“Amw‘uanwm a

v

3 . .J
.dwwkmrufi

. gunman: ...
11,,
wank”?

‘
1
3..

1 .4. I».
kn. u gmwMVA“: w
Ekmnuflmr

.1.

.v

9. ”.me

“WW.

,

iv...

.v r
fa-

{I

 

.
an... luau] u.
. V-ll _ ‘1
‘1‘ If V]
:5.
«4.4.9.9 unn. .

U 'O.
(n?!

than? «if!

y‘mnh~..\.u a.

. V. rm}? ..

:2»... :6“

nhgmt...fi

THESIS

Q
city‘l
This is to certify that the
dissertation entitled
IMPROVING REGIONAL CLIMATE MODELING IN EAST
AFRICA USING REMOTE SENSING PRODUCTS
.93
£6
E 35 E
< E g presented by
I ‘5 >
e 2’2
_J o D J G
"" ian'un e
:5 ’

 

has been accepted towards fulfillment
of the requirements for the

PhD. degree in Department of Geography

 

 

"‘ 0/3;

 

' Professor’s Signature
f/B A2997
/

Date

 

MSU is an affirmative-action, equal-opportunity employer

.-_t-.--.----—-—.—-—--o—.-.-.-------‘-n-u---.-u—n—-.. 4

PLACE IN RETURN BOX to remove this checkout from your record.
To AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DAIEDUE

DAIEDUE

DAJEDUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/07 p:/C|RC/DateDueindd-p.1

IMPROVING REGIONAL CLIMATE MODELING IN EAST AFRICA
USING REMOTE SENSING PRODUCTS

By

Jianjun Ge

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of requirements
for the degree of

DOCTOR OF PHILOSOPHY
Department of Geography

2007

ABSTRACT

IMPROVING REGIONAL CLIMATE MODELING IN EAST AFRICA
USING REMOTE SENSING PRODUCTS

By
Jianjun Ge

Accurate representation of the land surface in regional climate models is
becoming unprecedentedly crucial as numerous studies are now focused to
simulate the influences of human modification of the Earth’s surface on regional
and global climate. The unique advantages of remote sensing technique in
monitoring the land surface have been recognized for decades. The climate
modeling community, however, has yet to fully utilize this technique, especially
the recently developed remote sensing products which have been proven to be
more suitable for global change studies. The objectives of this study are to
improve the representation of the land surface and to investigate the impacts of

land cover classification accuracy on regional climate modeling in East Africa.

Several land cover datasets from different sources now exist in almost
any region of the world. A new statistical measure Q is developed to evaluate the
land cover classification specifically for the purpose of climate modeling. In terms
of this Q measure, Global Land Cover 2000 (GLC2000) is ranked the best

among four land cover datasets.

To better represent the land surface newly developed MODIS Leaf Area

Index (LAI) and Vegetation Fractional Cover (VFC) imageries are incorporated

directly in the Regional Atmospheric Modeling System (RAMS). The default land
cover dataset is updated by GLC2000 as well. The impact is examined by
comparing the model simulated land surface temperature (LST) and precipitation
with MODIS LST and precipitation from the Tropical Rainfall Measuring Mission
(TRMM) satellite. This study finds that the incorporation of MODIS LAl and VFC
greatly improves the spatial and temporal characteristics of LST. The

precipitation, however, is less sensitive to the improved land surface conditions.

The uncertainty originating from the land surface and its propagation need
to be examined to truly improve the representation of the land surface in climate
models. This study focuses on the land cover classification accuracy, which is
the first such investigation. This study finds that classification accuracy under
80% has significant impact on simulated precipitation, especially when the land
surface has a greater control of the overlying atmosphere. As the accuracy
worsens, the effect becomes much stronger. In remote sensing community, an
85% overall accuracy has been brought up as a guideline of classification quality
control. This study shows that this accuracy target can indeed satisfy the
requirement of climate modeling in the East Africa region. In reality, however, the
classification accuracy can be much lower as historically reconstructed and
future projected land cover datasets are extensively used in many climate

modeling studies.

To my family

iv

ACKNOWLEDGEMENTS

Thanks to my advisor, Jiaguo Qi, for his good ideas and for helping me
through bumps in the road. Thanks to members of my committee, David Lusch,
Jeffrey Andresen, Brent Lofgren, and Lijian Yang. My thanks to the co-authors of
the papers coming out of this dissertation: Nathan Moore, Nathan Torbick, and
Jennifer Olson. Thanks to all the students and project members I have worked

with over the years. I appreciate your help and input.

Funding for this study was provided by the National Science Foundation
(NSF) Biocomplexity of Coupled Human and Natural Systems Program award
BCS 0308420 entitled “An Integrated Analysis of Regional Land-Climate
Interactions”, and by the NASA's Land Cover and Land Use program through a

grant to Michigan State University (NNG05GD49G).

TABLE OF CONTENTS

LIST OF TABLES ............................................................................................... viii
LIST OF FIGURES ............................................................................................... ix
Chapter 1 Introduction .......................................................................................... 1
1.1 Climate modeling and the importance of land surface ................................. 1
1.2. The role of remote sensing ......................................................................... 4
1.3. Land cover accuracy and uncertainty assessment ..................................... 9
1.4. Research objectives ................................................................................. 12
Chapter 2 Research Design ............................................................................... 15
2.1. Regional Atmospheric Modeling System (RAMS) overview ..................... 15
2.2. Research design ....................................................................................... 16
Chapter 3 Biophysical Evaluation of Four Land Cover Products for Regional
Climate Modeling ................................................................................................ 19
3.1 Introduction ................................................................................................ 19
3.1.1 Study area .......................................................................................... 19
3.1.2 Land cover classification evaluation ................................................... 20
3.2 Methodology .............................................................................................. 22
3.2.1 0 development ................................................................................... 23
3.2.2 Illustration ........................................................................................... 25
3.2.3 Evaluation design ............................................................................... 25
3.3 Data description ......................................................................................... 26
3.3.1 Land cover products ........................................................................... 26
3.3.2 MODIS LAI product ............................................................................ 28
3.4. Results ...................................................................................................... 29
3.4.1 Q values ............................................................................................. 29
3.4.2 Significance test ................................................................................. 31
3.4.3 Single class investigation ................................................................... 33
3.5. Conclusion and discussion ....................................................................... 36

Chapter 4 Improving the Land Surface Representation in RAMS and Impacts

Analysis ............................................................................................................... 38
4.1 Introduction ................................................................................................ 38
4.2 Incorporating MODIS LAI & VFC in RAMS ................................................ 44
4.3. LAI and VFC comparison .......................................................................... 48

4.3.1 Temporal comparison ........................................................................ 48
4.3.2 Spatial comparison ............................................................................. 51
4.4 RAMS and configuration ............................................................................ 55
4.5 TRMM precipitation and MODIS land surface temperature ....................... 59
4.5.1 Weather station data .......................................................................... 59

vi

4.5.2 TRMM precipitation ............................................................................ 60

4.5.3 MODIS land surface temperature ....................................................... 61
4.5.4 LST data quality ................................................................................. 64
4.6. Results ...................................................................................................... 67
4.6.1 LST .................................................................................................... 67
4.6.1.1 Temporal comparison ..................................................................... 67
4.6.1.2 Spatial comparison .......................................................................... 72
4.6.1.3 Diurnal LST ..................................................................................... 75
4.6.2 Precipitation ....................................................................................... 78
4.7 Conclusions and discussions ..................................................................... 81

Chapter 5 Impacts of Land Cover Classification Accuracy on Regional Climate

Simulations ......................................................................................................... 84
5.1 Introduction ................................................................................................ 84
5.2 Methodology .............................................................................................. 86
5.3 Land cover dataset .................................................................................... 87
5.4 Land cover accuracy ................................................................................. 88
5.5 Results ....................................................................................................... 91

5.5.1 Basic experiment ................................................................................ 91
5.5.2 Follow-on experiments ..................................................................... 100
5.6 Discussion ............................................................................................... 103
5.7 Conclusions ............................................................................................. 106

Chapter 6 Conclusions and Future Research .................................................. 109

APPENDIX Acronym List ................................................................................. 114

BIBLIOGRAPY .................................................................................................. 115

vii

LIST OF TABLES

Table 4-1. Biophysical parameters for some important land cover types ............ 41
Table 4-2. TRMM 3B-42 characteristics .............................................................. 61
Table 4-3. Data quality of MODIS LST in 2003 ................................................... 66

viii

LIST OF FIGURES

Figure 2-1. Flowchart of research design in this study ........................................ 16

Figure 3-1. Study area illustrated by GLC2000 for Africa. Area 1, 2, and 3 refer to
Lake Victoria, Rwanda, and Burundi, respectively. ............................................. 21

Figure 3-2. A simulated 4X4 pixel area. Quadrate a) has 2 classes and quadrate
b) has 3 classes. Numbers shown are LAI values ............................................... 25

Figure 3-3. Q maps for four land covers at the quadrate size of 30X30 km:
GLC2000 (a), MODIS land cover (b), OGE (c), and LEAF (d). ........................... 30

Figure 3-4. Mean Q for all land covers at three scales. .......................................... 31

Figure 3-5. Multiple comparison (Tukey’s method) results for mean Q values for a
quadrate size of 30X30 km (a) and 50X50 km (b). MOD means MODIS land
cover. Confidence intervals were built on significance level of 0.05. ................... 33

Figure 36 Q was applied to a single class, croplands (>50%), in GLC2000. (a) is
Q map at 30X30km quadrate size; (b) croplands in the hotspot in (a), pointed by
two lines; (c) Africover corresponding to (b). ....................................................... 35

Figure 4-1. Schematic representation of‘heat and moisture transfer between
components of LEAF-2 for two patches within a model grid cell (Walko et al.
2000). See text for detailed description. .............................................................. 40
Figure 4-2. Built-in seasonal cycles of LAI (top) and VFC (bottom), illustrated by
Crop/Mixed Farming land cover type. “N5” and “$5” refer to 5 degree north and 5
degree south respectively. ........................................................................................... 43

Figure 4-3. Illustration of domain decomposition with four nodes: I, II, III, and IV.
............................................................................................................................ 48

Figure 4-4. Seasonal variations of leaf area index (top) and vegetation fractional
cover (bottom) for wooded grassland. The black lines are LEAF2 built-in
variations, and the red lines are from MODIS products ....................................... 49

Figure 4-5. Same as in Figure 4-4 but for crop/mixed farming. ........................... 50

Figure 46 Spatial comparison of the built-in, improved and original 1 km LAI for
March, June and September, 2003. .................................................................... 53

Figure 4-7. Same as in Figure 4-6 but for VFC. .................................................. 54

ix

Figure 4-8. RAMS domain with Ax = 50 km. ....................................................... 57
Figure 49 Distribution of weather stations in the study area. ............................ 59
Figure 4-10. Cloud contamination in nighttime MYD LST for November, 2003...65

Figure 4-11. Temporal comparison of maximum daily LST from three RAMS
runs: OGE, GLC, and GLC+LAI+VFC, and monthly MYD observations for the
whole study domain. ........................................................................................... 71

Figure 4-12. Same as in Figure 4-11 but for minimum daily LST and for eastern
domain only. ........................................................................................................ 71

Figure 4-13. Spatial comparison of LST from MODIS observations (Top). RAMS
with default land cover and biophysical parameters, DEF run (Middle), and
RAMS with GLC2000 and MODIS LAINFC, GLC+LAI+VFC run (Bottom). ........ 73

Figure 4-14. Monthly averaged diurnal cycles from MODIS and three RAMS runs:
OGE, GLC, and GLC+LAI+VFC. ......................................................................... 76

Figure 4-15. Domain averaged three-hour precipitation rate from TRMM for 2003
and three RAMS runs: DEF, GLC, and GLC+LAI+VFC. ..................................... 79

Figure 4-16. Accumulated precipitation in 2003 from TRMM and three RAMS
runs: DEF, GLC, and GLC+LAI+VFC .................................................................. 80

Figure 5-1. Cross referenced GLC2000 with 1km resolution (a) and 50km
resolution (b), and simulated land cover classification errors: 10% (c), 30% (d),
and 50% (e). Land cover types in (b), (c), (d), and (e) only represent the biggest
patches in grid cells. See texts for more details. ................................................. 89

Figure 5-2. Spatial comparison of simulated accumulated precipitation (mm) in
RAMS and that from TRMM for March — May, 2003. .......................................... 93

Figure 5-3. Temporal comparison of simulated cumulative precipitation in RAMS
and from TRMM. ................................................................................................. 93

Figure 5-4. Differences of accumulated precipitation (mm) between simulation
without land cover error and simulations with 10% error (R10-R00, a), 30% error
(R30-R00, b), and 50% error (R50-R00, c). ........................................................ 95

Figure 5-5. Maximum absolute differences (a), mean absolute differences (b) and
standard deviations (c) from the basic experiment (Kain-Fritsch), follow-on 1
(Kuo), follow-on 2 (Kain-Fritsch with interior nudging), and follow-on 3 (Kuo with
interior nudging). ................................................................................................. 99

Figure 5-6. Simulated accumulated precipitation (mm) with different model
parameterizations. Results from runs without land cover error are presented. .102

xi

Chapter 1

Introduction
1.1 Climate modeling and the importance of land surface

Modern climate change is dominated by human influences, which are now
likely large enough to exceed the bounds of natural variability (IPCC, 2001 ). It is
believed that climate changes resulting from human impacts are slow to develop
and, therefore, may not become apparent until their effects have become
irreversibly advanced. At global scale, anthropogenic emissions of greenhouse
gases, such as C02 that results from the combustion of fossil fuels, change the
atmospheric composition and influence the natural flows of radiant energy. At
regional scale, human activities are transforming the land surface at an
accelerated pace, which is often more pronounced where people live, work, and
grow food. Deforestation in Amazonia (Hahmann and Dickinson, 1997) and
desertification in the Sahel (Xue, 1997) are two instances where evidence

suggests there is likely to be significant human influence on regional climate.

Over the past four decades, climate models have been employed as
primary tools to enhance understanding of past climate changes and to aid
prediction of future climates. In climate models, the behavior of the atmosphere is
described using a set of differential equations that describe external forcings to
the system and the response of the atmosphere to the forcing. Historically,

modeling of the atmospheric component of the climate system has received the

most attention. As the availability of computers has become greater and
computer storage capacity and speed of computation have increased, climate
models have increased in complexity. More efforts have been concentrated on
including four major components — atmosphere, oceans, sea ice and land
surface — and the interactions and feedbacks among them (McGuffie and

Henderson-Sellers, 2001; Washington and Parkingson, 2005).

The land surface is the lower boundary of the atmosphere and thus a key
component of climate models. It controls the partitioning of available energy at
the surface between sensible and latent heat, and it also controls the partitioning
of available water between evaporation and runoff. Lewis F. Richardson was the
father of today’s climate models. In the first description of a method for
constructing a weather forecasts numerical calculations Richardson recognized

the importance of the land surface (Richardson, 1922). He noted:

“The atmosphere and the upper layers of the soil or sea form together a
united system.

Leaves, when present, exert a paramount influence on the interchange of
moisture and heat. They absorb the sunshine and screen the soil
beneath. Being very freely exposed to the air they very rapidly
communicate the absorbed energy to the air, either by raising its
temperature or by evaporating water into it. A portion of rain, and the
greater part of dew, is caught on foliage and evaporated there without
ever reaching the soil. Leaves and stems exert a retarding friction on the
air.”

These is increasing evidence from modeling studies that the influence of

the land surface on climate is significant and that changes in the land surface can

influence regional to global scale climate on time scales from days to millennia.
Studies have demonstrated the sensitivity of climate to land surface albedo
(Charney et al., 1977; Cunnington and Rowntree, 1986; Laval, 1986; Sud and
Fennessy, 1982; Lofgren, 1995). Recently, Betts (2000) shows that the positive
radiative forcing induced by decreases in albedo can offset the negative forcing
that is expected from carbon sequestration by afforestation. Studies have found a
change in global temperature and precipitation following a prescribed change in
leaf area index (LAI) (Chase et al., 1996). Precipitation patterns in the tropics are
altered substantially in a Global Circulation Model (GCM), where two 10-year
simulations were performed: one with the current global seasonally varying LAI
and one with the potential seasonally varying leaf area index (Nemani et al.,
1996). A new area of recognition in climate models is the role of plant roots. Zeng
et al. (1998) found, for example, that the root distribution influences the latent
heat flux over tropical land. Kleidon and Heimann (2000) concluded that in order
to realistically represent the tropical climate system deep-rooted vegetation must

be adequately represented.

Instead of focusing on a single parameter change, many studies examined
the climate change caused by land use/cover change (LUCC) (Chase et al.,
1996, 2000; Pielke et al., 1999; Betts, 2000; Zhao et al., 2001a,b; DeFries et al.,
2002; Taylor et al., 2002; Feddema et al., 2005), deforestation (Bonan et al.,

1992; Henderson-Sellers et al., 1993; Lean and Rowntree, 1997), and

desertification (Xue, 1997; Nicholson et al., 1998). All these studies have found

significant continent-scale changes in temperature, rainfall and other variables.

1.2. The role of remote sensing

Ever since Deardorff’s (1978) pioneering work on parameterizations for
both soil and vegetation, land surface submodels within climate models have
evolved from quite simple treatments of the surface energy, moisture and
momentum exchanges to increasingly complex descriptions (Dickinson, 1991;
Sellers et al., 1997; Pitman, 2003; Yang, 2003). Also, our ability to sense
characteristics of the land surface remotely has improved dramatically, enabling

much better data to be input to the more sophisticated parameterizations.

The land surface in climate models is usually represented by a discrete
set of land cover types, each characterized by a suite of biophysical parameters
(eg. LAI, vegetation fractional cover (VFC), albedo, root depth and roughness
length). Until the last decade, the land cover products used in most climate
models were initially compiled from maps, ground surveys, and various national
sources, which have inherent limitations (Mathews 1983, Olson et al. 1983,
Cihlar 2000). In the mid-19903, global-scale land cover products generated from
remotely sensed images became available. The Global Land Cover
Characteristics Database (Loveland et al. 2000), generated from one year 1 km

Advanced Very High-Resolution Radiometer (AVHRR) data, have been widely

implemented in various major soil-vegetation-atmosphere transfer (SVAT)
schemes, such as the Biosphere-Atmosphere Transfer Scheme (BATS;
Dickinson et al. 1986), the Simple Biosphere Model (SB and SiBZ; Sellers et al.
1986, Sellers et al. 1996a, b), the Land Ecosystem-Atmosphere Feedback Model
(LEAF; Lee 1992, Walko et al. 2000), and the National Center for Atmospheric
Research Land Surface Model (NCAR LSM; Bonan 1996, Oleson and Bonan,

2000)

For more accurate specification of landscape patterns, many climate
modeling studies have also attempted to use biophysical variables derived from
AVHRR imagery, particularly spatial and seasonal vegetation structure
information. For example, AVHRR normalized difference vegetation index (NDVI)
was used to determine the temporal variation of LAI and fraction of absorbed
photosynthetically active radiation (fPAR) in the revised Simple Biosphere Model
(SIBZ) (Sellers et al. 1996a,b). Oleson and Bonan (2000) studied the effect of
plant functional type and LAI derived from AVHRR imagery on the simulation of
surface fluxes for boreal forest using the NCAR Land Surface Model and found a
substantial impact of spatial heterogeneity. LAI from AVHRR NDVI was
incorporated into the Regional Atmospheric Modeling System (RAMS) to
investigate the sensitivity of regional climate simulations to changes in vegetation

(Lu 2002).

For many years, terrestrial ecosystem monitoring at moderate spatial
resolutions suitable for climate studies, relied almost exclusively on AVHRR
imagery. However, reliance on AVHRR imagery with its associated spectral and
geometric constraints has limited the ability of the land research community to
develop the range of products needed for global change research (Cihlar 1997).
Now a new sensor, the Moderate Resolution Imaging Spectroradiometer
(MODIS), is providing a series of products of unparalleled quality and
sophistication for biophysical observation of the terrestrial environment. MODIS
is designed to satisfy the requirements of three different disciplines: atmosphere,
ocean and land, with spectral bands and spatial resolution selected to meet

different observational needs (Salomonson et al. 1989).

Standard products produced from MODIS imagery include LAI, fraction of
photosynthetically active absorbed radiation (fPAR), enhanced vegetation index
(EVI), land surface temperature (LST), net primary productivity (NPP), land
cover, albedo, etc (Justice et al. 2002). Initial results of validation have suggested
that there is good agreement between MODIS LAI product values and field
measurements as well as those scaled-up from very high spatial resolution
satellite data (Myneni et al. 2002). Through the BigFoot project (Running et al.
1999), the MODIS LAI algorithm has been assessed for distinct biomes over the
world, such as savannah and shrub in southern Africa (Tian et al. 2002),

broadleaf forest in eastern United States (Shabanov et al. 2003), an agricultural

and broadleaf forest site in North America (Cohen et al. 2003), and the tundra of

Alaska (Verbyla 2005).

The enhanced vegetation index (EVI) was developed to optimize the
vegetation signal to improve vegetation monitoring by considering the canopy
background signal and a reduction in atmospheric influences. The MODIS EVI
has been demonstrated to have improved capability to capture multi-temporal
vegetation variations, land cover variations, and biophysical parameter variations

(Huete et al. 2002).

Along with the new MODIS land cover products (Friedl et al. 2002), the
Global Land Cover 2000 dataset (GLC2000) has recently become available (Fritz
et al. 2003). Sponsored by the Joint Research Center (JRC) of the European
Commission (EC), GLC2000 was developed from SPOT VEGETATION 1 km
data with enhanced spectral, spatial, radiometric, and geometric quality.
Compared to the AVHRR-based land cover datasets, these two products are
more suitable for monitoring land surface properties at regional to global scales
and have great potential to be employed in climate modeling systems in the

future (Latifovic et al. 2004, Giri et al. 2005).

Remote sensing data can not only be used to provide better initial and
boundary conditions, but also to evaluate the model performance and the

accuracy of the forecasts. Currently, validation of global and regional climate

models is based on the comparison between model outputs of standard
meteorological fields and meteorological observations. The main problem with
traditional meteorological observations when used to validate models is their
poor representation of the grid-point average simulated by a model. Well
documented station observations of climate variables are mainly located in
populated and industrialized regions. Observations are very sparse both spatially
and temporally over many rural regions of the world, which brings inevitable

challenges for climate modeling studies.

Several climatic variables can now be measured from space. Rainfall and
land surface temperature (LST) are of the two most important ones. Since late
1997, the Tropical Rainfall Measuring Mission (TRMM; Kummerow et al. 2000)
has been successfully collecting passive microwave and radar-derived tropical
precipitation data over both the land and ocean. The latest rainfall data product
(3842 version 6) has a quarter-degree spatial resolution and 3-hour temporal
resolution covering'from 50 degree south to 50 degree north globally. Although
the primary objective of TRMM was to improve climate models and to aid them in
climate prediction (Kummerow et al. 2000), very few studies have fully utilized

this dataset for model validation and rainfall data assimilation.

AVHRR derived LST was compared with the output from the NCAR
Community Climate Model version 2 (CCM2) coupled with BATS to illustrate the

differences between air temperature and skin temperature at the global scale (Jin

et al. 1997). Since then, few other efforts have attempted to apply satellite LST in
climate modeling. This is partly because the algorithm of AVHRR LST was a
simple extension of the sea surface temperature (SST) method, which
sometimes produced to unacceptable errors (Price 1984, Becker, 1987). AVHRR
data are not suitable for accurate cloud detection over land which is required for
an operational LST algorithm. With multiple thermal-infrared bands specifically
designed for LST retrievals. MODIS is providing a much more accurate LST

product for both regional and global climate modeling (Wan et al. 2002, 2004).

1.3. Land cover accuracy and uncertainty assessment

Modeling of the climate system which has a wide variety of components is
a formidable task, and as a result climate models have uncertainties (Mahlman
1997, Palmer 2000, Vidale et al. 2003). Simulated climatic variables differ
considerably between models. The uncertainties in the climate models originate
from several components, of which the SVAT scheme is believed to be very
important. The primary aim of SVAT is to take into account the essential land
surface processes and provide estimates of latent and sensible heat fluxes
between the land surface and atmosphere. The uncertainties in SVAT schemes
arise from three elements: the model parameterization, meteorological forcing

data, and vegetation/soil inputs.

In SVAT schemes, land surface features are differentiated by land cover
types, and then biophysical parameter sets are assigned to each land cover type.
While numerous studies have been conducted to investigate the sensitivity of the
climate model to some particular biophysical parameters (e.g., Franks et al.
1997, Bastidas et al. 1999, Lynch et al. 2001), some attention has been given to
the uncertainty stemming from the land cover dataset. By examining the impact
of different land cover datasets (digitalized vs. satellite derived) on the predicted
variables, Molders et al. (1997) found that the distribution of daily averages of
temperature and humidity changed less than 0.2 °C (1%) and 0.2 g/kg (1.5%) if
the coverage of the various land-use and soil types differed by only about 5%
between the datasets. Pauwels and Wood (2000) compared the effect of
differences in spatial resolution of land cover data to land-atmosphere model
results relative to the effect of differences in land cover sensors and classification
schemes. They found that the uncertainty in model results arises mainly from the
land cover classification and that the lack of spatial resolution had a lower effect.
Overall, an uncertainty of approximately 15% in modeled energy and water fluxes

and states has to be assigned in evaluating the model simulation.

Until all important biophysical parameters in any SVAT scheme can be
mapped directly by satellites, land cover datasets will continue to play a key role
in representing the land surface conditions. No land cover product is 100%
accurate even those produced from the most advanced remote sensing imagery.

The uncertainties caused by the classification inaccuracy thus need to be

10

investigated. Because several land cover datasets are sometimes available for a
study, the effect of land cover classification accuracy on model results can help
climate modelers to decide which one to select. For land cover dataset
producers, the effect of classification accuracy can be used as a guideline for the
accuracy assessment and quality control. If the effect is not significant, no extra

efforts are needed to pursue higher mapping accuracies.

In the remote sensing community, some target accuracy thresholds have
recently been recommended in an attempt to provide guidelines to the
classification quality. Thomlinson et al. (1999), for example, set as a target an
overall accuracy of 85% with no class less than 70% accurate. However,
classification accuracy is usually interpreted differently from the viewpoint of
various users. The effect of land cover accuracy for a particular application, such
as climate modeling in this study, remains an unanswered question. The
accuracy targets commonly specified have largely not been tested from the
perspective of the operational use of land cover data. As anthropogenic impacts
on the land surface and climate at various scales are attracting more and more
attention from both scientists and policy makers, an effective way to evaluate
classification accuracy needs to be developed specifically for climate modeling
studies and the impact of classification uncertainty on climate simulations needs

to be fully investigated.

11

1.4. Research objectives

The overall research objectives in this study are to improve the
representation of the land surface in a regional climate model using newly
developed remote sensing products and to address the uncertainty of land cover
classification and its impact on model simulations. The specific research

objectives are:

1. To develop a new method to evaluate various land cover products for
regional climate modeling

2. To improve the representation of the land surface in a regional climate
model using the highest-quality land cover product from Objective 1 and
newly developed MODIS biophysical products (LAI and VFC).

3. To investigate the land cover classification uncertainty and its impact on

regional climate simulations.

The new classification accuracy evaluation method is designed specifically
for regional climate modeling, addressing the shortcomings of traditional methods
used in the remote sensing community. Impacts of improved land surface
representation will be examined by comparing model outputs with remote
sensing observations (TRMM and MODIS LST). This will be the first study to
directly incorporate MODIS biophysical products into a regional model and to use
TRMM and MODIS LST to evaluate the model performance in a region where

station observations are extremely scarce. Also, a classification accuracy

12

threshold will be identified by the third objective. This will provide an additional
guideline for land cover accuracy assessment in the remote sensing community.
When the classification accuracy is low, climate simulations must be interpreted
with caution, especially when historically-developed and future-projected land

cover products are used.

The dissertation is outlined as follows:

Chapter 2 describes the methodology of this study. A more detailed
discussion of the various remote sensing products is provided and where and
how they are used is illustrated. In addition, the regional climate model used in

this study is briefly introduced.

Chapter 3 (Objective 1) focuses on the theoretical development and
application of a new statistical method to evaluate land cover classification
accuracy. The most appropriate land cover product is identified from several

existing ones.

Chapter 4 (Objective 2) describes the implementation of the new land
cover product identified by Objective 1 and new MODIS LAI and VFC products.
Spatial and temporal characteristics of the improved land surface map are

compared with the default representation in the regional model. The impact of the

-13

improved land surface characterization is examined by comparing model-

simulated LST and precipitation with MODIS LST and TRMM data respectively.

Chapter 5 (Objective 3) investigates the impacts of land cover

classification accuracy on simulated precipitation. A threshold of accuracy is

idenfified.

Chapter 6 is the conclusion. Future studies in this area are discussed.

14

Chapter 2

Research Design
2.1. Regional Atmospheric Modeling System (RAMS) overview

The regional climate model used for the numerical simulations in this
study is the Regional Atmospheric Modeling System (RAMS) Version 4.4 (Pielke
et al. 1992, Cotton et al. 2003). RAMS is a three-dimensional, nonhydrostatic,
general purpose atmospheric simulation modeling system, which solves
equations of motion, heat, moisture, and mass continuity in a terrain-following
coordinate system. It is an atmospheric model which is capable of both numerical

weather prediction and regional climate simulation.

The SVAT scheme employed in RAMS is the Land Ecosystem-Atmosphere
Feedback model, version 2 (LEAF-2) (Lee 1992, Walko et al. 2000). LEAF-2
represents the storage and vertical exchange of water and energy in multiple soil
layers, temporary surface water or snow cover, and vegetation and canopy air.
The special feature of LEAF-2 is its ability to represent fine-scale surface
variations by dividing surface grid cells into sub-grid patches, which are assigned
based on the land cover types in a model grid cell. Each patch has one land
cover type and influences the overlying atmosphere in its own unique way
according to its fractional area of coverage. The biophysical characteristics, such
as albedo, leaf area index, fractional vegetation cover, etc., are prescribed for the

land cover type in each patch. However, as in many other SVAT schemes

15

seasonal variation of vegetation is characterized by simple mathematic

equafions.

2.2. Research design

The following flowchart summarizes the research design in this study.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

LC #1
LC #2
LC #3 1
Assessment
Selected RAMS _ MODIS LAI

LC I V MODIS VFC
{.5 '-

Uncertainty !

<:> MODIS LST
1TUHAJ

 

 

 

 

 

 

Figure 2-1. Flowchart of research design in this study

A new land cover product and one-year monthly MODIS LAI & VFC
datasets will be directly ingested into RAMS to improve the land surface
conditions both spatially and temporally, and the effects of these new land cover
and biophysical products will be examined separately. RAMS will be run three
times: As a control, the first run uses the default land cover dataset and
prescribed LAI & VFC. The second run uses a new land cover product, but
prescribed LAI & VFC. For the third run, both new datasets will be used. Model-
simulated precipitation and LST will be compared with TRMM precipitation and
MODIS LST. RAMS does not have LST as an output variable. An algorithm thus
will be used to calculate LST based on the temperature of vegetation and soil in

a grid cell.

A new statistical evaluation method is developed and applied to several
existing land cover products to select the most appropriate one for climate
modeling in the study region. Therefore, various land cover products will be

evaluated and compared before running the RAMS model.

To study the uncertainty effect of land cover classification inaccuracy, a
range of misclassifications (5% - 50% at 5% interval) will be simulated on the
baseline land cover dataset. During this process, spatial locations (excluding
water) will be randomly selected and original land cover types will be converted

to random predominant types. RAMS will then run multiple times, each with a

17

different amount of classification error. Model outputs will be compared and the
threshold of classification accuracy will be identified. Also, the influence of model

configuration will be examined.

Images in this dissertation are presented in color.

18

Chapter 3
Biophysical Evaluation of Four Land Cover Products for

Regional Climate Modeling

3.1 Introduction

3.1.1 Study area

This study focuses on the East Africa region (Figure 3-1), which covers
Kenya, Uganda, Rwanda and Burundi, as well as parts of Congo and Tanzania
and a small section of the Indian Ocean. The geographic coordinates for the
upper-left and lower-right corners of this study area are (3.71°, 23.80°) and (-
934°, 41 .80°), and the size is about 2000km by 1500km. This region contains
some of the most varied topography in the world, including large lakes (e.g., Lake
Victoria), rift valleys, and snow-capped mountains (e.g., Mount Kilimanjaro) on
the equator. This region also has a wide range of land use/cover types: from
savannas to forest to intense agriculture. Rainfall in East Africa mostly occurs
during the boreal spring (long rains, March-May) and autumn (short-rains,
September/October—December) seasons as the intertropical convergence zone

(ITCZ) migrates through the equator from south to north, and vice versa.

The land surface in this region has been substantially changed by human

activities during the past decades due to increased population and other factors,

which may likely trigger climate change. At the same time, people in this region

19

are highly vulnerable to climate variability and change. Much of the region is arid,
semi-arid or sub-humid and yet highly populated by pastoralists and agro-
pastoralists dependent on rainfed pasture and crops. This study is in the context
of the Climate-Land Interaction Project (CLIP), which integrates diverse methods
to quantify the two-way interactions between land use and regional climate

systems.

3.1.2 Land cover classification evaluation

To represent the land surface conditions realistically, RAMS needs an
accurate land cover dataset as input. Traditional land cover classification
accuracy assessment is primarily based on ground-based surveys or
interpretation of high spatial resolution aerial photos and satellite images. By
comparing the classified land cover with ground-truth data, error metrics can be
developed to report the commission and omission errors (Congalton 1991 ).
Shortcomings of this traditional method are apparent. Firstly, it is not cost
effective. Secondly, reported accuracy is highly susceptible to sampling error
when the sample taken is far sufficient to represent the study area, especially
when classifications have regional and global coverage. Thirdly, the ground-truth
data are usually obtained at one particular time, which can’t take into account the
phenological consistency within each class. Finally, classification accuracy is
often interpreted differently from different users (Lark 1995, Brown et al. 1999).

The traditional method is not designed for the purpose of climate modeling.

20

 

 

 

    
 

Tanzania
Ocean

 

 

 

- Background [:3 Open gassland with sparse shrubs
- Closed evergreen lowland forest [4] Open gascland

- Submontane fared (900-1500m) L j Sparse gassland

- Montane forest (>1500n) E] Swamp bushland and gassland
E] Mosaic Forest/Croplands [j Croplands (>50%)

E: Mosaic Forest/Savanna L Oroplands with open woody vegdation
- Closed deciduous forest E Sandy desert and dunes

- Deciduous woodand :3 Salt hartbans

m Deciduous shrubland with sparse tress - Waterbodies

Open deciduous shrubland Cities

[—1 Closed gassland

 

Figure 3—1. Study area illustrated by GLC2000 for Africa. Area 1, 2, and 3 refer to
Lake Victoria, Rwanda, and Burundi, respectively.

21

 

Researchers have proposed to use biophysical products with observable,
unambiguous, and continuous structural variables to make an optimized
classification for specific needs (Running et al. 1995, DeFries et al. 1997, Cihlar

2000). This idea has not yet been used in the phase of classification assessment.

In this study, a new statistical method, called Q, is developed to
biophysically evaluate the seasonal consistency of classification. This method

eliminates sampling error and is designed specifically for climate modeling.

3.2 Methodology

In the RAMS model, land cover types in a grid cell are treated as subgrid
‘patches’, each with its own vegetation, soil, canopy air, and etc. (Walko et al.
2000). The surface characteristics of a patch are then represented by a group of

biophysical parameters: LAI, VFC, albedo, etc.

Biophysical parameters are now available from remote sensing products.
For each patch, biophysical parameters can be taken directly from corresponding
remote sensing products. As a result, variation of these parameters in a patch
can be used to indicate if this patch (land cover type) has been appropriately
assigned (classified). Aggregating all patch-level variations spatially can be used

to evaluate land cover products from the perspective of climate modeling. The

22

statistical measure Q presented in this paper is developed on LAI which is one of

the most important biophysical parameters.

3.2.1 Q development

For an area of interest (a RAMS grid cell) with k classes, LAI variation, V,
within one class ican be calculated using statistical error sum of squares
(Stapleton 1995).

p- _
Vi = :ILA’in -LA/i)2 (3-1)

n=1

where p; is the total number of pixels for class i, n refers to any particular pixel,

and L—A—Ii is the mean LAI value. Also, seasonal changes of LAI must be taken
into account. Including the temporal dynamics of vegetation is expected to
improve the evaluation. By adding this temporal LAI information, equation (3-1)

can be rewritten as:

T
Wi= Zvit ’7' (M)
i=1

where trepresents time. In this paper two years of monthly LAI data (January
2002 to December 2003) are used with tranging from 1 to 24, and Ttherefore
equals the maximum number of time periods used (For some areas not all 24
time periods can be used because of quality control for LAI pixels. This is

discussed later.) Summing W,- for all classes will give a good indication of within-

23

class LAI variation. However, there is still another very important issue left out.
For a given geographic area, different land cover products may classify an
unequal number of classes to it because of different classification schemes.
Therefore, summation of WI must be normalized by (N — k) to yield the final
statistical measure Q. It is expressed as:
1 k
Q = WEN (3-3)

where N is the total number of pixels for the area of interest (e.g., a grid cell in
RAMS) and is the same for different land covers, and k is the total number of
classes. It is assumed here that N is larger than one. In most regional climate
modeling studies, the spatial resolution is usually coarser than 10km because of
the limitation on computing resources. By combining equation (3-1) and (3-2),

equation (3-3) can be written as:

Q:(___N-1-k)z ZZILA/rm- [MY ’7 <3-4)

— t=1=n1

In summary, Q aggregates LAI variation in each land cover type for each
time period for an area. A smaller Q value indicates more consistent biophysical
characteristics over a two-year time period and therefore, a more appropriate

classification for climate modeling.

24

3.2.2 Illustration

As an illustration, Figure 3-2 represents a simulated area with 4X4 pixels.
The LAI value in each pixel is also presented at one particular time. For the
purpose of clarity, in this study this type of rectangular area will be referred as
‘quadrate’ to differentiate from ‘pixel’ in image processing and ‘grid 09/7 in climate
modeling. One quadrate is composed of smaller image pixels and corresponds to
a grid cell. In this illustration two quadrates in Figure 3-2 represent two different
classifications for one area. Quadrate a) has two classes and its Q value is 0.02.
In comparison, quadrate b) has three classes and its Q value is 0.47. These two

Q values suggest quadrate a) is more properly classified.

 

 

 

 

 

 

 

 

 

 

 

Figure 3-2. A simulated 4X4 pixel area. Quadrate a) has 2 classes and quadrate
b) has 3 classes. Numbers shown are LAI values.

3.2.3 Evaluation design

Non-overlapping equal size quadrates are tiled completely for the study
area. Q values were calculated quadrate by quadrate for each land cover

products. The mean Q value was utilized as the final evaluator. By doing so, the

25

effect of spatial resolution of the climate model can be studied by adjusting the
quadrate size. As quadrate size becomes smaller, the Q value may decrease. In
the case where the size of the quadrates is one pixel, there would be just one LAI

value and thus no variation at all.

For the climate modeling in this study, the grid spacing of RAMS is 50 km.
This resolution was designed based on the topographical characteristics in the
study area and computational requirement. The evaluation was extended to
quadrate size of 30X30km and 100><100km to study how resolution affects the
application of Q. In the following sections, Q values for all four land covers for
quadrate size of 30X30km, 50XS0km and 100><100km will be calculated and

analyzed.

3.3 Data description

3.3.1 Land cover products

Q was applied to the following four land cover products, which exist for

climate modeling in East Africa:

1) GLC2000 for Africa (See Figure 3-1)

This product was developed by the Joint Research Centre’s Global

Vegetation Unit based primarily on SPOT VEGETATION daily 1km data, which

26

were acquired from 1 November 1999 to 31 December 2000 (Mayaux et al.
2004). Other data sources such as NDVI, radar and Defense Meteorological
Satellite Program (DMSP) were also used. GLC2000 uses the Land Cover
Classification System (LCCS) developed by the Food and Agriculture
Organization of the United Nations (FAO) and the United Nations Environment
Programme (UNEP), which contains 27 land cover classes (Di Gregorio et al.
2000). The GLC2000 for Africa was downloaded at http://www-gvm.jrc.it/gI02000.
It was in geographic coordinates and then reprojected to Lambert Azimuthal

Equal-area projection.

2) MODIS land cover

This product (MOD1ZQ1) was developed by MODIS Land Cover group at
Boston University. It was prepared using MODIS Terra daily data acquired from
15 October 2000 to 15 October 2001, except for June 2001, which is missing due
to instrument down time. This product was based on the International
Geosphere-Biosphere Programme (IGBP) global vegetation classification
scheme which has 17 classes (Friedl et al. 2002). MOD1201 for Africa was
downloaded from http://duckwater.bu.edu/lc/mod12q1.html and was already in

Lambert Azimuthal Equal-area projection with 1km resolution.

3) OGE

27

This is the Global Land Cover Characterization database (version 2.0)
using the Olson Global Ecosystem legend (Olson 1994a, b). It was developed by
the US. Geological Survey (USGS), the University of Nebraska-Lincoln (UNL),
and the European Commission’s Joint Research Centre (JRC), based on 1km
AVHRR data spanning April 1992 through March 1993. It is archived at the
EROS Data Center http://edcdaac.usgs.gov/glcc/glcc.html. This dataset is called

OGE for short in this paper.

4) LEAF

This is the default land cover used in LEAF-2 within the RAMS model,
termed LEAF herein. LEAF aggregates the OGE dataset into 31 classes to use

the BATS land-surface parameters (Walko et al. 2000).

3.3.2 MODIS LAI product

The purpose of this evaluation is to select an appropriate land cover for
climate simulations at a time period of interest (2002-2003) rather than to
examine whether these land covers were accurately mapped when they were
produced. 1km monthly MODIS LAI data from January 2002 to December 2003
were used in this evaluation. The newly reprocessed MOD15_BU LAI C4.1 data
were downloaded at ftpzllprimavera.bu.edu/pub/datasets/MODISI. They were
reprojected to Lambert Azimuthal Equal-area projection with 1km resolution. For

LAI pixels with normal values (0-7), Quality Assessment (QA) flags were used to

28

detect cloud contamination in each pixel (Myneni et al. 2002). Only high quality
LAI pixels having QA values from one to four were selected for Q assessment.
These pixels were produced by the main algorithm, without saturation under
clear sky conditions. For LAI pixels with fill values (200, pixels outside projection;
253, Barren, desert, or very sparsely vegetated; 254, water; 255, non-computed
pixels or missing pixels), only 253 and 254 were used by replacing them with
zero. It needs to be noted that there is persistent cloud cover over part of the
Congo forest in the study area, especially during the wet season. But, this is not

a problem since Q can be calculated with only one month of data.

3.4. Results

3.4.1 Q values

Q values were calculated quadrate by quadrate for all four land cover
products at three quadrate sizes (30X30, 50X50 and 100><100km). Only Q values
at quadrate size of 30X30km are graphically presented here (Figure 3-3). It is
noticeable that Q has a similar spatial pattern for all land covers. It has larger
values for areas with complex and heterogeneous landscapes, such as ecotone
boundaries, and has smaller values for more homogeneous areas, such as
forests and deserts. Large water areas like Lake Victoria have Q of zero because

of no LAI variation.

29

 
 

.o-o.1-o.1-o.5 mos-1[j1.1.5:]1.5.2[:j2-2.5u2.5.3“345-154-445

 

Figure 3-3. Q maps for four land covers at the quadrate size of 30X30 km:
GLCZOOO (a), MODIS land cover (b), OGE (c), and LEAF (d).

Having Q values at quadrate level, the mean Q value was calculated and
used as the indicator of evaluation. The mean Q values for each land cover
product at three different quadrate sizes are plotted in Figure 3-4. The mean Q
value for GLCZOOO is the smallest, while the Q value for LEAF is greater than the
other three land covers for every quadrate size. The rank for mean Q values for
all four landcovers is QGLc < QMoms < Qoee < QLEAF. This suggests that GLCZOOO
has the best land cover classification and LEAF has the least appropriate
classification in terms of LAI variation. Also, MODIS land cover is better than

OGE. As expected, for a given land cover, Q values increase as the size of

30

quadrates increases. For example, the mean Q value for GLCZOOO increases

from 0.69 for 3060 to 0.80 for 100x100_

 

   

1.0' *‘***’ *zi””””’

0.7
0.6

Mean Q value

0.5 *

 

 

0.4

Quadrate size

 

 

 

Figure 3-4. Mean Q for all land covers at three scales

3.4.2 Significance test

A null hypothesis that mean Q values from the four land cover products
are all equal was tested using the One Way Analysis of Variance (ANOVA)
(Scheffe 1959). One advantage of using ANOVA, rather than multiple t-tests, is
that it gives one p value for a large number of groups. It would have required six
pairs for the t-test to evaluate the four land cover classifications in this study. Like
other statistical tests, ANOVA assumes that within each sample the values are
independent and normally distributed. This may not be satisfied for Q values

since they may be spatially correlated, which is common for most spatially

31

distributed data. In this study, I assumed there was no spatial correlation for Q
values. A significance level of 0.05 was chosen to conduct the test for each
quadrate size. Sample sizes (number of quadrates) for four land cover
classifications are equal on all three scales: 3484 for 30X30km, 1240 for
50XS0km and 300 for 100><100km. The resulting p-values for each test are
0.0002, 0.0054, and 0.0739, respectively. This analysis suggests that mean Q
values at the quadrate size of 30km and 50km are significantly different (0.0002,

0.0054 << 0.05), while mean Q values at 100km are not significantly different.

In order to identify specific differences between pairs of groups, Tukey’s
method of the multiple comparison procedure (MCP) was used following the
ANOVA (Zhuang et al. 1995, Stapleton 1995). ANOVA only tells whether there is
a difference among the four land cover products, but not which ones are
significantly different. MCP provides which land cover classification is different if

significant difference has been found by ANOVA.

Figure 3-5 presents part of the results from the MCP for quadrate sizes
30X30km (Figure 3-5 a) and 50XS0km (Figure 3-5 b). In this figure, the positions
of the dots represent the differences between sample means for Q. Parenthesis
and dashed lines indicate the extent of the confidence intervals for differences
between population means. If a confidence interval does not contain zero, the
difference for that pair is significant. As Figure 3-5 illustrates, GLC2000 has a

significantly smaller Q value than LEAF at both the 30km and 50km quadrate

32

sizes, and at 30km quadrate size GLC2000 is also significantly different from
OGE. All other paired differences are not significant. Since ANOVA determined
that there is no significant difference among the four land cover classifications at
the 100><100km quadrate size, it was not necessary to conduct MCP at this

scale.

 

glc-leaf
gIc-mod
gIc-oge
leaf-mod I—fi————e- ————— .L
leaf-oge I——————e- ————— .I.

mod-oge I —————— .._..._.__.).

 

 

 

 

I I I I l I I
-0.1 O -0.06 -0.02 0.0 0.02 0.04 0.06 0.08

 

glc-Ieaf I —————— -e— ————— .I.

gIc-mod
gIc-oge ( —————— + —————— I'
leaf-mod
Ieaf—oge
mod-oge

 

 

 

 

-0.16 -0.1 2 -0.08 -0.04 0.0 0.04 0.08 0.1 2

Figure 3-5. Multiple comparison (Tukey’s method) results for mean Q values for a
quadrate size of 30X30 km (a) and 50X50 km (b). MOD means MODIS land
cover. Confidence intervals were built on significance level of 0.05.

3.4.3 Single class investigation

For illustrative purposes, Q was also applied to a single class in GLC2000

to investigate the LAI variation within that class. The croplands (>50%) class was

33

selected for this study because it is one of the most prevalent land cover types,
occupying about 19% of the study area (Figure 3-1). It is defined as regions of

intensive cultivation and/or sown pasture.

In Figure 3-6a, Q values of this single class calculated at the quadrate size
of 30X30km are presented. Map cells in white represent those quadrates not
containing any croplands. The mean Q value for this class is 1.017, which is
higher than that for all other classes of GLC2000. This is because agricultural
fields in Africa are usually small and mixed with savanna and fallow patches,
which preclude a reliable mapping at 1km spatial resolution. The spatial pattern
is very similar to the map in Figure 3-3a. High Q values (yellow and red) tend to

occur in areas with complex landscapes.

A hotspot with high Q values was identified visually in Figure 3-6a.
Geographically, it is in the Mount Elgon area that straddles the border between
Kenya and Uganda (Figure 3-1). In the Q map it consists of four contiguous
quadrates (total size of 60X60km). The mean Q value for this area is 3.265,
which is much higher than the total mean value (1.017). The spatial distribution of
the croplands (>50%) class in this hotspot area is presented in Figure 3-6b. It
occupies 1069 1km pixels (about 29.7%) in the original GLC2000 dataset. To
further investigate the cause of the high Q, Africover land cover at 100m
resolution was examined for this patch classified as croplands by GLC2000

(Figure 3-6c). Africover was produced from a visual interpretation of TM data

34

from the year 2000 to 2001 (Jansen et al. 2003). It is considered as ground truth
here because of its much higher resolution (30 x 30m). Africover shows the
GLCZOOO cropland patch in this hotspot area is actually composed of up to five
diversified land cover types: forest, open shrubland, closed shrubland, savanna
and crop. Integrating such a complex landscape to one single type will certainly

produce high within-class LAI variation.

 

 

 

7 8, background *
- cm I
- open shrub L ,,
.forest
I closed shrub

- savanna

 

 

   

 

 

 

 

 

 

 

Figure 3-6. Q was applied to a single class, croplands (>50%), in GLC2000. (a) is
Q map at 30X30km quadrate size; (b) croplands in the hotspot in (a), pointed by
two lines; (c) Africover corresponding to (b).

35

3.5. Conclusion and discussion

Based on the needs of climate modeling, a statistical measure, Q, was
developed to biophysically evaluate four land cover products: GLC2000, MODIS
land cover, OGE and LEAF, using the monthly MODIS LAI product. Evaluations
were conducted at three spatial scales (quadrate sizes): 30X30, 50X50 and
100><100km. This evaluation found that in terms of Q, GLC2000 ranks the best
and LEAF ranks the lowest at every scale. MODIS land cover is better than OGE.
As quadrate size increases, the differences between the land cover products
tended to decrease. For the quadrate sizes of 30km and 50km, GLC2000 is
significantly better (i.e., the smallest mean Q value) than LEAF, and for the
quadrate size of 30km GLCZOOO is also significantly better than OGE. This
suggests that the LEAF dataset (built into the RAMS model) needs to be updated
by GLC2000 in order for the model to better capture the surface conditions in

East Africa.

There are several comments to be made on the proposed approach. First,
the time period of LAI data should be fairly long. Based on the quality control
flags, only high quality LAI pixels should be picked. Thus, not every LAI pixel is
valid for Q calculation. If for example only a couple of months of LAI are used,
there might not be enough high quality LAI pixels to calculate Q in some
quadrates. This is especially true for tropical areas, where persistent cloud cover

could exist for months.

36

The second comment is related to the overall quality of the MODIS LAI
data. The accuracy of using Q to evaluate alternative land cover classifications
depends on the accuracy of LAI product used. Two measures were taken to
address this issue. One was using monthly LAI data, composited from 8-day data
by selecting high quality pixels over a month period. The accuracy of the 8-day
LAI product is about 0.5 LAI (Wang et al. 2004). The other was further filtering
the LAI pixels finally used by Q according to the quality assessment flags. Only
pixels produced by the main algorithm under clear sky conditions were selected.
Nonetheless, the overall LAI quality in this region is still not well known. Some
publications have already shown that the LAI product has limitations (Wang et al.

2004).

Finally, LAI is only one of the biophysical variables representing surface
properties of land cover in climate modeling. Other biophysical variables (e.g.,
fractional vegetation cover or albedo) are equally important. Q may also be
applied to these variables. More complete conclusions may be drawn by
evaluating alternative land cover classifications using a suite of biophysical

variables.

37

Chapter 4
Improving the Land Surface Representation in RAMS

and Impacts Analysis
4.1 Introduction

The land surface plays a prominent role in climate modeling, since it
exchanges momentum, energy, water, and other important chemical constituents
with the atmosphere. The land surface is characterized by pronounced spatial
heterogeneity that spans a wide range of scales. In the last decade or so, the
importance of representing land surface heterogeneity representation, especially
at the subgrid scale, has been increasingly recognized in a large number of
climate modeling studies. Ideally, surface heterogeneity effects could be
accounted for by running a coupled atmosphere-surface model at a very high
spatial resolution, so that the heterogeneity is explicitly captured. However, this
approach is too computationally demanding even for most regional modeling.
The land surface heterogeneity, therefore, needs to be parameterized within the

framework of complex land surface process schemes.

The Land Ecosystem-Atmosphere Feedback model version 2 (LEAF-2),
which is the SVAT scheme in RAMS, represents subgrid heterogeneity through a
discrete number of homogeneous subregions, also referred to as “patches” or
“tiles” (Avissar and Pielke 1989). Patches are selected on the basis of the land

cover types (evergreen broadleaf forest, savanna, water, etc), and they directly

38

exchange fluxes with the atmosphere independently of each other. Each patch
occupies a fractional horizontal area of a grid cell in RAMS. Net momentum,
moisture, sensible heat, longwave, and shortwave fluxes are integrated over all
patches, weighted by the corresponding patch fractional areas. This approach
does not keep track of the actual geographical location of the patch within the
model grid cell; all the subareas belonging to a given land cover class are treated

equally.

The LEAF-2 components and the flux pathways between them are
illustrated in Figure 4-1. This example includes two patches beneath an
atmospheric column (A) within a single RAMS grid cell. Both patch 1 and patch 2
have partial vegetation cover (V) and patch 2 alone has snowcover (S). Two soil
layers (G) and canopy air (C) are also shown in this figure. Fluxed are denoted
by smaller letters. The prefix (w, h, or r) indicates whether the flux involves the
transfer water, heat or longwave radiation, and the two suffix letters denote the
source and the receptor (9 for ground, 3 for snow, v for vegetation, c for canopy,
and a for reference height air). The one exception is wgvc, which denotes a flux
of water from the soil to the canopy air through the vegetation by the means of

transpiration.

39

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

C
w'ee
4
rga wgc hgc rgv was
WWW (22 wsvc2-- Gz
I l I
hgg wgg wgg "99
W9V°1----‘ GI wgvc‘I-ua G1
PATCH 1 PATCH 2

Figure 4-1. Schematic representation of heat and moisture transfer between
components of LEAF-2 for two patches within a model grid cell (Walko et al.
2000). See text for detailed description.

In LEAF-2, the standard land cover map that generates the subgrid
patches during the initialization of RAMS is the Global Ecosystem dataset (OGE,

Olson 1994a, b). Currently, the 94 ecosystem classes in OGE have been cross
referenced to the 18 BATS classes (Dickinson et al. 1986) plus some additional
classes, which were referred to as LEAF-2 cover in Chapter 1. The RAMS user

decides the number of patches, and LEAF-2 fills those patches with the most

predominant classes.

40

These various land cover types (patches) are characterized by a suite of
biophysical parameters: leaf area index (LAI), vegetation fractional cover (VFC),
displacement height, roughness height, albedo, and emissivity. Some of these
biophysical variables are specified from observations in field experiments, while

others are educated guesses.

Table 4-1. Biophysical parameters for some important land cover types

 

 

 

 

 

 

 

 

 

 

Evergreen .
Crop/Mixed Open
LC types Eroadleaf Farming Shrublan d Grassland Woodland
orest
Albedo 0.06 0.20 0.12 0.11 0.08
Emissivity 0.95 0.95 0.97 0.96 0.96
LAI 6.00 6.00 6.00 2.60 5.70
D LAI 1.00 5.50 5.40 2.00 2.30
VFC 0.80 0.85 0.22 0.73 0.80
D VFC 0.10 0.60 0.12 0.11 0.17
R°Ughness 2 21 o 06 o 08 o 04 o 83
length ' ' ' ' '
Displacement
height 20.7 0.7 0.2 0.2 7.4
Root depth 1 .20 1.00 0.60 0.70 1.00

 

 

 

 

 

 

 

 

Table 4-1 presents the default biophysical parameters in LEAF-2 for some
important land cover types in the study area. LAI and VFC are maximum leaf
area index and vegetation fractional cover; D LAI and D VFC are the maximum

decrease in leaf area index and vegetation fractional cover. See

41

http://www.atmet.com/html/docs/rams/RT1-leaf2-3.pdf for more detailed

biophysical characteristics of all land cover types defined in LEAF-2.

Of these biophysical variables, LAI and VFC are assumed to have a
simple seasonal dependence (Figure 4-2), which is the function of a cosine
distribution, latitude and time in a year. Vegetation is assumed to peak in late
July (Julian day = 200) in the northern hemisphere and the reverse in the
southern hemisphere. For locations close to the equator, such as a large part of
East Africa in this study, LEAF-2 assumes that seasonal variation is reduced to
zero. According to Figure 4-2, the built-in spatial and temporal vegetation

variations are extremely unrealistic for near-equatorial regions.

42

 

Crop/Mixed Farming - LAI

 

6.1 I 7‘7

 

 

1 61 121 181 241 301 361
Julian Day

Crop/Mixed Farming - VFC

0.9 , , ._,,,.,,, _,,,,,,,, "new Lt,

 

0.88

E

 

(00
0|

 

 

 

0.8

1 61 121 181 241 301 361
Julian Day

 

 

 

Figure 4-2. Built-in seasonal cycles of LAI (top) and VFC (bottom), illustrated by
Crop/Mixed Farming land cover type. “N5" and “S5" refer to 5 degree north and 5
degree south respectively.

In chapter 3, the default land cover dataset in LEAF-2 was found to have
significantly lower quality than the newly developed land cover products, such as
GLCZOOO. As previously discussed, there are major shortcomings of the built-in
vegetation characteristics in RAMS. This component of study is, therefore, to
improve the land surface representation in RAMS by updating the default land

cover dataset and the built-in spatial and temporal vegetation dynamics with

43

GLC2000 and MODIS LAI and VFC products, respectively. Impacts of these

improvements will also be investigated.

4.2 Incorporating MODIS LAI 8. VFC in RAMS

The monthly, 1km MODIS LAI and EVI products for 2003 were
downloaded from MODIS group at Boston University
ftp://primavera.bu.edu/pub/datasets/MODIS/ (see Chapter 3 for more details).
These images were transformed to a Polar Stereographic projection centered at
(33°E, 2°S), which corresponds to the projection of the RAMS domain. The LAI

data were converted to binary files so that RAMS can read them directly.

It needs to be mentioned that MODIS LAI has a meaning different than
that in RAMS. As is standard in the remote sensing community, MODIS LAI is
defined as “the area of green leaves per unit area of ground” (Curran 1983, Price
1992), which is sometimes referred to as “effective” LAI. In contrast, LAl in LEAF-
2 (Table 4-1) and other land surface models is defined as “the number of leaf
layers over the vegetated part of the pixel" (Dickinson, personal communication),
which is also referred to as “clump” LAI (Choudhury et al. 1994). Therefore,
MODIS LAI was divided by MODIS VFC so that the ingested LAI has the same

meaning as defined in RAMS.

44

VFC data were developed from the 1km monthly MODIS Enhanced
Vegetation lndex (EVI) product (see Chapter 1 for more description). Compared
to the traditional Normalized Difference Vegetation Index (NDVI), EVI takes full
advantage of MODIS’ state-of-the-art measurement capabilities and has much
improved quality (Huete et al. 2002). One noticeable advantage of EVI is that it
does not become saturated as easily as NDVI when viewing tropical rainforests

and other areas of the Earth with large amounts of green biomass.

Calculating fractional cover from vegetation indices is based on the theory
of “Mosaic Pixel”, which assumes that a remote sensing pixel has a patchy
(mosaic) structure (Kerr et al. 1992, Valor and Caselles 1996). A quantity

measured by satellite (¢, e.g. vegetation index) fora pixel can be interpreted as

the sum of linear contributions from the vegetated area (fv) and bare soil (1 - fv ):

¢=¢va +¢s(1—fv) (4-1)
where the subscripts v and 3 denote values over fully vegetated and bare soil

areas. From this equation, a simple formulation for fractional cover fv can be

derived as:

f = ¢_¢S
V ¢v-¢s <4-2>

While some studies have suggested a non-linear relationship between vegetation

 

fractional cover fv and vegetation indices ¢ (e.g., Myneni and Williams 1994,

Carlson and Ripley 1997), others found the linear relationship is sufficient

45

(Wittich 1997, Gutman and lgnatov 1998). For this study, the linear relationship is

assumed, and Equation (4-2) was used to calculate VFC from MODIS EVI.

EVI for fully vegetated (¢v ) and bare soil (¢S ) areas needs to be derived
in order to use Equation (4-2). Some investigators have suggested that ¢(, and
(b3 are dependent on vegetation and soil types, season and geographic regions

(e.g., Price 1992, Huete et al. 1994, Zeng, et al. 2000). This study is focused on
the East Africa region rather than the whole globe, so that the problem is
minimized. In addition, Gutman and lgnatov (1998) found that there are many
land cover types for which it may be difficult to find these two constants during
the course of a whole year. For example, only one of these two constants can be
derived for deserts and evergreen forests. Following the study by Gutman and
lgnatov (1998), «A, and ¢S in this study are prescribed as seasonally and
geographically invariant constants, which correspond to the yearly maximum EVI

of the Congo Forest (0.86) and minimum EVI of deserts (0.05) in northern Kenya

(Figure 1-1).

Monthly MODIS LAI and VFC were then interpolated linearly to determine
daily values so that RAMS can update LAI and VFC values at each time step. It
is assumed that monthly MODIS LAI and VFC are observed in the middle of each

month.

46

To ingest the MODIS LAI and VFC data properly into RAMS, the parallel
processing architecture must be considered. RAMS is structured in a standard
master-node configuration, where the master process handles initialization and
output, while the nodes are the main workers. In this configuration, model domain
is horizontally decomposed into rectangular subdomains, which cover different
sets of grid points and a surrounding boundary region. The nodes work on
subdomains and exchange information at the subdomain boundaries. As lower
boundary information, LAI and VFC are not assigned during the initialization of
RAMS, but are updated dynamically during the simulation. Therefore, the MODIS
LAI and VFC data must be assigned to the proper nodes for smooth and correct

data incorporation in RAMS.

Figure 4-3 is a schematic illustration of domain decomposition for four-
node (I, II, III, and IV) computing in RAMS. The whole domain is 40 (x) by 30 (y)
grid cells, which is the domain configuration in this study (This will be further
discussed in Section 4.5). Dark areas are boundaries and overlaps between
nodes. They exchange information and are coordinated by the master process

during each model time step.

47

 

Figure 4-3. Illustration of domain decomposition with four nodes: I, II, III, and IV.

4.3. LAI and VFC comparison

In this section, the built-in and new LAI and VFC are compared both
temporally and spatially to illustrate the improved representation of the land

surface in the RAMS model.

4.3.1 Temporal comparison

Seasonal LAI and VFC were compared for individual land cover types.

Wooded grassland (Figure 4—4) and crop/mixed farming (Figure 4-5) are

48

illustrated here. The wooded grassland is located at (8.9°S, 23.8°E) (latitude and

longitude), while the crop/mixed farming is at (6.2°S, 32.8°E).

 

 

 

 

 

 

 

5 a
g 4 .. —— LEAF2
.. —— MODIS
‘ Wooded Grassland
g 3 4 (lat-8.9. lon:23.8)
2 'l \ .‘l \‘v‘. \
i / \\ /’/ \ \
\x’ , / . \
1 -L. _ I‘\¥\/ \'A __1
o I V I fi I I I I I
O 50 100 160 200 350 300 350
Tine (day)
1.0
l
0.8 -
o '\
E 0.6 _ \z,.——~\\ / /\
U- 044 \\ [/fd/
3% —— LEAF2 \ /
— MODIS \\\ //
E 02 . Wooded Grasshnd \
(lat-8.9, lon123.8)
0.0 I f 1 f I

 

 

501w'1éo'250'250'350350
Tima(day)

Figure 4-4. Seasonal variations of leaf area index (top) and vegetation fractional
cover (bottom) for wooded grassland. The black lines are LEAF2 built-in
variations, and the red lines are from MODIS products.

49

 

 

 

 

 

 

 

 

 

6
l
5 a.
'3 4 - Crop/Mbted Farming
_ 1 /, (lat-6.2. lon:32.8)
\ -— LEAF2
g 3‘ / \v"’/\\\ .-— moons
"i w \
3
2 - /,,’-#
J V
1.
o ' I ' I 1 I ' I ‘ I ' I ' I
O 50 100 150 200 250 300 350
This (day)
1.0
0.8 -————m——~f‘_’_—‘l
g .
g Crop/Mixed Farmhg
g 0-6 . (lat-6.2, lon:32.8)
\ /fi\ — LEAF2
l, ‘N/ ~ _ — moons
E 0.4 a ~
5 4 \“\ I//
§ \ /
> 0.2 -‘ “MW—d
0.0 . , . , T T . j E, , r , I
0 50 100 150 200 250 300 350
Time (day)

Figure 4-5. Same as in Figure 4-4 but for crop/mixed farming.

50

According to these two figures, RAMS built-in LAI and VFC displayed little
annual variation and had much higher values compared to the MODIS
observations. As discussed previously, the seasonal variation of vegetation in
RAMS is assumed to have a cosine distribution, which varies with latitude (see
Figure 4-2). The focus area in this study straddles the equator and has low
latitude. As a result, the built-in seasonal vegetation cycles in RAMS have

unrealistically small temporal variation.

4.3.2 Spatial comparison

Figure 4-6 and Figure 4-7 present the spatial comparison of LAI and VFC
respectively for different time periods in 2003: March, June and September. In
addition to the built-in and improved LAI and VFC, the original 1 km MODIS
imagery is also included in the comparison. The resolution for the built-in and
improved LAI and VFC maps is 50 km, which corresponds to the RAMS grid
spacing (see section 4.5 for model configuration). Colors in the 50 km maps

indicate LAI and VFC values for the biggest patches in the grid cells.

Overall, the built-in LAI and VFC are extremely homogeneous spatially.
Except for deserts and lakes, vegetation has little variation over the domain. The
built-in data show the Congo forest to have similar biophysical characteristics as
the semi-arid areas in the east. As shown in the previous section, the built-in LAI
and VFC also present unrealistic temporal variation. For example, the observed

VFC in a large part of southeast portion of the study area decreases significantly

51

in the second half of the year, which is likely related to the pattern of ITCZ.

However, this is completely missed in the default VFC.

52

meow gonEofiom ucm 0:2. .552 .8 .5 Ex r 3598 new 3505:: 53:3 05 Co sawtquoo 6:me .o-v 2:9“.

noon
new

moon
scan

noon
Eras.

  

.5: 200: .8355. $.53

53

no.0

.95 .2 Sn o4. 059“. 5 mm mEmw NA. 9:9”.

.5: 28.2 3.65....

 

5-25m

8cm
new

moon
«can

mean
:93:

54

In the following sections, RAMS will be run with the updated land cover
dataset (GLC2000) and improved LAI and VFC for the whole year of 2003. The
impacts will then be examined on the basis of simulated precipitation and land

surface temperature.

4.4 RAMS and configuration

RAMS 4.4 is an atmospheric model which is capable of both numerical
weather prediction and regional climate simulation. In a philosophical sense,
numerical weather prediction depends on the initial values of the state variables
of the atmosphere. On the other hand, climate simulation is run for longer periods
of time, so that it is insensitive to the initial conditions, but dependent on
boundary conditions such as ocean temperature, land use, and greenhouse gas
concentrations (Giorgi and Mearns 1999). This simulation includes some parts of
the climate system such as a full treatment of atmospheric dynamics,
thermodynamics and moisture processes, along with a Soil-Vegetation-
Atmosphere Transfer (SVAT) scheme. However, unlike some climate models,
RAMS does not include a fully interactive ocean. Rather, it treats ocean surface

temperature as a prescribed boundary condition.
The soil model in LEAF-2 consists of 11 vertical layers spanning a depth of

2.1m, and the soil temperature profile in the initial condition was determined by a

deviation from the initial air temperature in the lowest atmospheric level. The soil

55

moisture content for the top layer was initialized as 35% of the saturation value,
which was horizontally homogeneous over the domain. This percentage was
increased with depth to a maximum of 55% at 48 cm and below. Moisture flux
between soil layers was parameterized in LEAF-2 based on the multilayer soil
model described by Tremback and Kessler (1985). Both energy and moisture
fluxes between LEAF-2 components (i.e., vegetation, canopy air, and each soil
and snow cover layer) are illustrated in detail in Walko, et al. (2000, also refer to

section 4.1).

Soil moisture can play an important role in surface-atmosphere interactions
particularly through moisture “memory” in semi-arid regions like in Kenya and
Tanzania (Fig 3-1). The presence of soil moisture influences the partitioning of
latent and sensible heat, thereby affecting the development of shallow
convection. However, the soil types in East Africa are poorly mapped, and
available soil moisture values for the region are speculative due to data scarcity.
It needs to be emphasized that the role of surface parameters, including soil
moisture, can strongly affect the model solution (Ducharne and Laval 2000). In
the absence of reliable data, and to avoid introducing more complex uncertainties

into this experiment, this homogeneous approach was chosen.

56

 

Figure 4-8. RAMS domain with Ax = 50 km

In Figure 4-8, a single model grid with 50 km horizontal grid spacing was
used to cover the study area (see Fig 3-1). The horizontal grid spacing was set at
50 km in consideration of the domain size and the computational requirements.
For the land surface, the standard RAMS 30-arc sec topography dataset was
used. The grid extended over 32 vertical levels, with a layer thickness of 80m
near the surface and stretching to 1900m at the top of the domain. The model
was driven by 6-hourly lateral boundary conditions derived from the National
Centers for Environmental Prediction (NCEP) atmospheric reanalysis product

(Kalnay et al. 1996). The model time step was 90 seconds with the output period

57

set to every six hours. At each time step, the reanalysis data were nudged over

five outer grid points.

The radiative transfer scheme of Chen and Cotton (1983) was used to
parameterize the vertical flux of shortwave and longwave radiation. Horizontal
diffusion coefficients were computed based on the modified Smagorinsky
formulation (Smagorinsky 1963), and the vertical diffusion was parameterized
according to the scheme of Mellor and Yamada (1982). The bulk microphysics
parameterization was activated, which allows the model to consider the effect of
moisture in all phases. The sea surface temperature was specified using the 1°

monthly climatological dataset from NCEP (Reynolds and Smith 1994).

Three experiments were carried out to evaluate the impacts of the
improved land surface parameterization in RAMS. In the first run, here called the
“default” (DEF) run, the default land cover dataset (OGE) and the built-in LAI and
VFC were used. In the second run, here called the GLC run, the built-in LAI and
VFC were used but with OGE land cover replaced by GLCZOOO. Both GLCZOOO
and MODIS derived LAI and VFC were used for the third run, which is called

GLC+LAI+VFC in this study.

58

4.5 TRMM precipitation and MODIS land surface temperature
4.5.1 Weather station data

Observed climatic variables, e.g., rainfall and temperature, are usually
used for assessing model performance and outputs. The main problem with
traditional meteorological observations is their poor representation of the grid-
point average simulated by a model. Well documented station observations of
climate variables are mainly located in populated and industrialized regions.
Observations are very sparse both spatially and temporally over many regions of

the world, which brings inevitable challenges for climate modeling studies.

 

BY 0 ‘5
- o
' a o . 0'. h
' a o
. do
3 D 0 l o
o o.- _r
K o 3 ....’I‘8 o
3. o. o . .' ' 'r ‘3’
. 3" ‘ .9.“

 

 

 

I‘\ {I .7 ‘ . ..
...»r L. a
I f5 O o >.
“ / . . ‘ s7
'. , . f/
/
k . t\
o o o.
\\ f"
_'\\ ’ ° 7'
- — ‘\ . K:

 

Figure 4-9. Distribution of weather stations in the study area

Figure 4-9 shows the weather stations in the study area. Spatially, there
are very few stations in Congo, Tanzania, and east Kenya. More importantly,

many of the stations shown here do not have a continuous time series of

59

observations. To solve this problem, climatic variables observed by satellites are
used in this study: rainfall data from the TRMM satellite and land surface

temperature from MODIS.

4.5.2 TRMM precipitation

Tropical Rainfall Measuring Mission (TRMM) is a joint US — Japanese
satellite that was launched in November 1997 (Simpson et al. 1998, Kummerow
et al. 2000). The primary mission of TRMM is to measure precipitation in the
Tropics where conventional surface observations are scarce. TRMM provides the
most accurate global tropical rain estimates to date by using a unique
combination of instruments (the TRMM Microwave lmager and the Precipitation
Radar) designed specially for rain observation and by using a low-inclination
(35°) orbit to provide excellent coverage of the Tropics (Kummerow et al. 2000).
The TRMM products have recently been evaluated over East Africa using station
observations (Dinku et al. 2007), and they perform reasonably well. Although the
primary objective of TRMM was to improve climate models and to aid them in
climate prediction (Kummerow et al. 2000), very few studies have fully utilized

this dataset for model validation and rainfall data assimilation.
For this study, the latest TRMM product (3842 version 6) is used. Its

primary characteristics are listed in Table 4-2

(http://daac._qsfc.nasa.gov/precipitation/TRMM README ).

60

Table 4-2. TRMM 38-42 characteristics

 

 

 

 

 

 

 

Temporal Coverage Start Date: Jan 1, 1998; Stop Date: -

Spatial Coverage Latitude: 50°S — 50°N; Longitude: 180°W — 180°E
Temporal Resolution 3-Hourly

Spatial Resolution 0.25° x 0.25°

Average File Size Original: ~ 4.5MB

File Type HDF

 

 

 

3B42 V6 data with global coverage for 2003 were downloaded from
http://Iake.nascom.nasagov/Giovanni/tovas . These data were subset and
compared with the 3-hour precipitation file produced by RAMS. An online
interface for visualization and analysis of TRMM data is also available from this

website.

4.5.3 MODIS land surface temperature

Land surface temperature (LST, also referred to as radiometric
temperature or skin temperature) is inferred from the thermal emission of the
land surface and is more directly related to the surface properties than the
surface air temperature. For bare soil, LST is the soil surface temperature. For
vegetated areas, LST is generally some average of the temperature of various

canopy and soil surfaces (Qin and Karnieli 1999, Jin et al. 1997). It is one of the

61

key parameters in the physics of land surface processes on regional to global
scales. For example, sensible heat flux is proportional to the difference between
instantaneous LST and the temperature of the overlying air. Many researchers
have used LST to estimate the sensible heat flux as well as soil moisture (Sun

and Mahrt 1995, Wang 1992).

Before satellites, LST could only be obtained by ground-based or airborne
radiometers, which have very limited capabilities to provide convincing regional
to global views. In the late 19705, global measurements of LST derived from
AVHRR became available. The split-window algorithm (Price 1984, Becker
1987), which was originally devised for sea surface temperature (SST), is the
most extensively used methodology to drive LST from the two thermal bands of
AVHRR: band 4 (10.5 - 11.5 um) and band 5 (11.5 —— 12.5 um) (Qin and Karnieli
1999). A major problem in using split-window methods is that the surface
emissivities are required with accuracy better than 0.01, which are almost
impossible to estimate for many land cover types such as semi-arid and arid

areas (Wan and Li, 1997).

With multiple thermal-infrared bands designed for LST retrievals, MODIS
is providing a much more accurate LST product for both regional and global
climate modeling. A physical-based day/night algorithm (Wan and Li 1997) was
developed to retrieve LST at 5 km resolution from a pair of daytime and nighttime

MODIS data in seven thermal bands (bands 20, 22, 23, 29, and 31 - 33).

62

Compared with in situ measurements, the MODIS LST accuracy is better than 1K

in the range from 263K to 300K (Wan et al. 2004).

More importantly, MODIS is capable of providing diurnal LST
observations. Two MODIS instruments (Salomonson et al. 1989) have been
launched. The first was launched on the Earth Observing System (EOS) morning
platform Terra (called MOD in this study), and the second was launched on the
EOS afternoon platform Aqua (called MYD in this study). The Terra overpass
times are around 10:30 (local solar time) and 22:30, while the Aqua overpass
time is around 13:30 and 1:30. For this study area, 10:30, 22:30, 13:30, and 1:30
in local solar time correspond to approximately 08:00, 20:00, 11:00, and 23:00 in
Coordinated Universal Time (UTC). MOD products have been made available to
the public for about five years, while MYD became available only recently.
Despite four observations per day, MOD and MYD combined are now providing
unprecedented diurnal LST information. In addition, MYD measurements are
obtained at 11:00 UTC which are more close to the time of maximum
temperature of the land surface. It is therefore more suitable for regional and

global change studies.

AVHRR LST has been compared with that simulated in a global climate

model (Jin et al. 1997). In this study, the newly developed MOD and MYD are

used for the first time to evaluate the LST calculated in a regional climate model

63

(RAMS) and to study the impact of changing the land surface representation.

Diurnal LST produced by RAMS is also examined.

RAMS does not output LST directly. In this study, canopy temperature and
top-layer soil temperature are combined to calculate LST in each grid cell on the

basis of vegetation fractional cover (VFC), using the following equation:
4 _ 4 4

where 0V is VFC, TV is the canopy temperature, and T9 is the top-layer soil

temperature (Jin et al. 1997). From equation (4-3), vegetation density (leaf area
index) and vegetation fraction can have a significant impact on simulated LST.
By incorporating the more spatially and temporally realistic MODIS LAI and VFC,

RAMS is expected to produce LST with improved characteristics.

4.5.4 LST data quality

Monthly MOD and MYD LST data (version 4) with 0.05° spatial resolution
was downloaded from the EOS Data Gateway (http://edcimswww.cr.usgsgov ).
They were subset and resampled to 50 km resolution to match with RAMS
output. It needs to be pointed out that MODIS LST is valid only under clear-sky
conditions. For this low-latitude study area (Fig 3-1), the daily product is usually
severely contaminated by clouds. Monthly MODIS LST is thus used in this study.
It is composited and averaged on the basis of clear-sky observations. However,

monthly LSTs in several months of 2003 were still adversely affected by the

64

persistent cloud cover, especially over the western part of the study area (Congo
forest) during nighttime. Figure 4-10 is an illustration of cloud contamination in
the original (005° spatial resolution) nighttime MYD LST for November, 2003.
Grey pixels have valid LST values, while yellow pixels indicate invalid LST

observations, which include the ocean and a portion of Congo forest.

 

Figure 4-10. Cloud contamination in nighttime MYD LST for November, 2003.

The data quality of MOD and MYD LST in 2003 is summarized in Table 4-
3, where «1 indicates no significant cloud cover and X indicates significant cloud
contamination. It is noticeable that most cloud contamination occurs during night

time and in the raining seasons.

65

Table 4-3. Data quality of MODIS LST in 2003

 

 

 

 

 

 

 

 

 

 

 

 

 

 

MYD MOD
Month Day Night Day Night
1 v x v v
2 v v v v
3 v v v x
4 v v v v
5 x! ~/ \I ~/
6 \l \l \J ‘J
7 \l J \l \l
8 ‘J \l ‘1 \l
9 ‘1 X \I ~/
10 ‘1 X \l X
1 1 \l X ‘1 X
12 J x v x

 

 

 

 

 

 

 

In this study, daytime MYD (1 1 :00 UTC) is assumed as the maximum daily
LST and is compared with RAMS LST at (12:00 UTC) for the whole domain.
Nighttime MYD (23:00 UTC) is assumed as the minimum daily LST, recognizing
that minimum temperature may actually occur a couple of hours later in the early
morning. From 0:00 UTC, RAMS outputs after every three hours. Thus, there is a
one hour difference between MOD and MYD observing times and the simulation
times in RAMS. These time differences are assumed to have a negligible
influence on the comparison in this study. Since nighttime LST is usually
contaminated with clouds in the western part of the domain (Fig 4-10), the

minimum LST comparison is restricted to the eastern part of the domain. In

66

addition, diurnal LST is compared from May to August, due to minimum cloud

contamination in this time period.

4.6. Results

In this section, LST and precipitation in 2003 simulated by the three RAMS
runs (DEF, GLC, and GLC+LAI+VFC) are compared spatially and temporally to
examine the effects of the improved land surface representation. They are also

evaluated against MODIS and TRMM products.

4.6.1 LST

4.6.1.1 Temporal comparison

Figure 4-11 shows the RAMS simulated maximum daily LSTs and the
MYD monthly composited observations, which are averaged for the whole study
area excluding water bodies. The three runs are differentiated in colors and bold
lines are simply 30-day averages of daily results. The impacts of the improved
land surface conditions are significant, especially when MODIS LAI and VFC
products are used. In contrast with the RAMS default land surface, LST seasonal
variation has been dramatically improved in GLC+LAI+VFC (green line in Figure
4-11). According to MYD observations the lowest LST in 2003 occurs in May.
This is correctly simulated by GLC+LAI+VFC, but is incorrectly simulated to be in

July by DEF and GLC runs. LSTs peak at about the same time periods (March

67

and September) in all three runs. However, RAMS strongly underestimates LST
in the second half of the year, probably as a result of its unrealistic LAI and VFC
information. Overall, GLCZOOO increases the maximum daily LST by about one
degree, but fails to change the seasonal dynamics. LST flattens out later in the

year, both with and without GLCZOOO.

As discussed previously, the land surface is primarily represented by both
land cover types and their related biophysical parameters (see Table 4-1).
Introducing a new land cover dataset, GLC2000, provides a better description of
the spatial distribution of land cover types across the study domain. This should
help to produce better spatial characteristics of LST (discussed later). However,
GLC2000 is not able to modify the seasonal dynamics which are still prescribed
as simple mathematical equations in LEAF-2 (Figure 4-2). In Equation (4-3), the
temperatures of vegetation and bare soil as well as their relative contributions in
a grid cell, vary not just by location, but also over time. As a result, using

GLC2000 alone does not improve the LST seasonality.

By incorporating temporally explicit MODIS LAI and VFC (Figs. 4-6 and 4-
7) RAMS improves its performance of simulating the seasonal dynamics of LST
in two ways. lnfonned by the MODIS LAI data, a more realistic vegetation density
of the vegetated area in a grid cell may help to improve the interactions of plants
and overlying atmosphere. The other one, which is more important, is the relative

contributions of vegetation and bare soil as determined by VFC. Significantly

68

increased areal extent of bare soil in the eastern domain during the second half
of the year (Figure 4-7) may help to lift the flat LST curves after June (Figure 4-

11).

It is noticeable that the MODIS-observed LST is much higher than the
RAMS simulation for a large part of the year. Remember that monthly MODIS
LST is composited on the basis of clear-sky observations. By screening out cloud
contamination, the MODIS observations are biased toward the highest values in

a month, since clouds diminish solar insulation and thus cause much lower LSTs.

Minimum daily LST is also compared in Figure 4-12. Because a significant
amount of invalid LST pixels exist in the nighttime MYD in the western domain
from September to January (see table 4-3), this comparison was conducted only
for the eastern domain, which is approximately the eastern half of the study area.
Water bodies were excluded in this comparison as well. Generally, RAMS
produces much better minimum LST dynamics in all three runs. This is possibly
due to less moisture in this arid/semi-arid area. Unlike the maximum daily LST,
the lowest minimum LST occurs in July. Seasonal variation is as large as about
four degrees. Similarly as in Figure 4-11, RAMS produces lower LSTs at the
beginning of the year, but higher LSTs after June using the GLC+LAI+VFC run.
In addition, GLC2000 does not have much impact on the LST seasonality. Also,

MODIS observations are lower than RAMS which is similarly due to MODIS’

69

clear-sky observations. The surface is usually cooler at night with less cloud

COVGI’.

70

 

     
 

 

 

38
3 33
2 v v
E (I V I V M
E J DI!‘ ':\ llll I It} I?!“ IVA. v
,_ 32 l ”6'55 III] If I“: l l v
E III!" I "ill. “II; '5. , “NM!“
0:) 30 I’ ”'1" i I “'1‘ [MIIL'M J“, .’ "'II'II "I‘l‘.mh'!l ’4 III
1: - ~tl| II, I! "'1 Nil!”
: It ”I‘ll I I'II I
3 28 l I I I “I; ‘ ~ I!
[ll -—OGE ,
’ —GLC+LAI+VFC
2‘ V MYDLST
6'53f160'1éo'2éo'25fi06307fi950
Days in 2003

Figure 4-11. Temporal comparison of maximum daily LST from three RAMS
runs: OGE, GLC, and GLC+LAI+VFC, and monthly MYD observations for the
whole study domain.

 

 

 

 

E
E
E
a
E
I-
8
E
3
to
'U
C
1'.

— OGE

14 — GLC

—GLC+LAI+VFC
A MYD LST
12 I ' I ' I ' I ' I ' I ' I '7 I
o 50 100 150 200 250 300 350
Days in 2003

Figure 4-12. Same as in Figure 4-11 but for minimum daily LST and for eastern
domain only

71

4.6.1.2 Spatial comparison

In Figure 4-13, the maximum daily LSTs from MODIS (MYD daytime) and
RAMS are spatially compared month by month. Images on the top are from the
MODIS observations, resampled to 50 km resolution. Only LSTs from the DEF
run (middle) and the GLC+LAI+VFC run (bottom) are included. Because the
comparison is in terms of land surface temperature all water bodies (lakes and
oceans) were masked out (white color). The magnitude of LST is indicated by

color: yellow to red show high LSTs while green to blue show relatively low LSTs.

MODIS LSTs peak from February to March and again from August to
September, and reach their minimum around May. This is consistent with the
bimodal temporal pattern shown in Figure 4-11. This feature is obviously
captured by the GLC+LAI+VFC run, but is completely missed by the DEF run.
LSTs do not vary much after June for the DEF run. Spatially, the western domain
which is covered by more vegetation and has more moisture and rainfall
(discussed later) has lower LSTs compared to the eastern part. Both the DEF
and the GLC+LAI+VFC runs seem to capture this feature, but the latter captures
this contrast much better. More importantly, MODIS-observed LSTs in the east
shows a strong ITCZ related pattern, in which high LSTs migrate from north to
south with time. This feature is fully captured by the GLC+LAI+VFC run. In Figure
4-7, the MODIS VFC shows a similar pattern. This confirms the importance of

VFC in calculating LST as is manifested in Equation (4—3).

72

€289 :2 83.555 .825 98.2 95 883m 5;, 92¢ Em 322% .5. ”an @2398

@5235 .25 .98 95. .333 £3 225.

.v.v 3 an

 

30.: 223238 9005. E0: .54 v0 comtmano 5:25 .94 9:9”.

0
on am 3 8 R om
‘IIII'V

 

73

.3

TV

8:538 .m _.-¢ 059“.

mm. mm mm

mm

mm

0N

 

74

4.6.1.3 Diurnal LST

The impacts of the land surface representation in RAMS are further
examined by looking at diurnal LSTs, taking advantage of the multiple daily
observations by MODIS Terra (MOD) and MODIS Aqua (MYD). Due to cloud
contamination (see table 4-3), only four months, May to August, are analyzed. In
Figure 4-14, the RAMS simulated LSTs, which are output eight times per
simulated day (every three hours), are shown by the color lines, while the four
MODIS observations are shown in red squares. Time is in UTC, which is about

three hours different from local time in the study area.

This figure shows that RAMS captures the diurnal cycles quite well. LST
reaches its daily maximum at about 12:00 UTC and the daily minimum at about
03:00 UTC. As shown in Figure 4-12, RAMS produces higher LST at night (0:00
and 21:00 UTC). During the day (09:00 and 12:00 UTC), the simulated LST is
close to the MODIS values in May and June. However, in July and August the
differences between simulated and observed daytime LSTs magnify. This is also
shown in Figure 4-11. During the day in July and August, only the LSTs

simulated by the GLC+LAI+VFC run are close to the MODIS observations.

75

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

35
May, 2003
A m‘
93
2
5.!
8 25-
if
.—
3 204 ,
g - ' I MODIS
‘3 —°—OGE
c 15- —‘—GLC
3 —v—GLC+LAI+VFC
10 T l f I f fi l f j ' r m l
o 3 6 9 12 15 1s 21
Time (UTC)
35
June, 2003
30
A 1
93
g d
E 25-
E?
l- .
8 20‘
g \, - MODIS .
(D , I —°—0GE
E 15- “—GLC
3 —- . — GLC+LAI+VFC
10 j I I f l ' l ' I ' I ' I
o 3 6 9 12 15 1s 21
Time (UTC)

Figure 4-14. Monthly averaged diurnal cycles from MODIS and three RAMS runs:
OGE, GLC, and GLC+LAI+VFC

76

Land Surface Temperature (°c)

Land Surface Temperature (’0)

 

   

 

    

 

 

 

 

 

 

 

   

 

   

 

 

 

 

 

 

July, 2003 .
30 d \
1 , .
25 .
20 .-
‘ z"; I MODIS .
. — . — OGE
15 - — . — GLC
-— ' — GLC+LAI+VFC
10 I ' I ' I ' I I T ' I ' I ' I
0 3 6 9 12 15 18 21
Tums (UTC)
35
August, 2003 '
30 _I I./\\
25 ..
20 4
I K I MODIS I
I ‘— ' — OGE
15 _ — 1 — GLC
— . — GLC+LAI+VFC
10 I . I . I . I . I . I . I I I
0 3 6 9 12 15 18 21
Time (UTC)

Figure 4-14. continued

77

4.6.2 Precipitation

Figures 4-15 and 4-16 show the TRMM-observed 3-hour precipitation
rates and accumulated precipitation in 2003, respectively. These TRMM
observations are compared with the three RAMS simulations: DEF, GLC, and
GLC+LAI+VFC. RAMS is able to produce the general temporal and spatial
dynamics of rainfall in this area considering RAMS has a much lower spatial
resolution than the TRMM observations (50km vs. about 27km). The two wet
seasons separated by a dry season in-between (around June) are clearly
captured in Figure 4-15. Spatially, the major features are captured by RAMS,
such as high rainfall areas over the Congo forest and Lake Victoria and relatively
dry areas in eastern Kenya and Tanzania. However, RAMS produces little rainfall
at the left and right boundaries, which maybe due to the boundary nudging
employed in the model. However, as shown in Figure 4-15, RAMS
underestimates the major wet season (March - May), but overestimates the
second wet season (about October to November) in 2003. In addition, RAMS
tends to generate a larger contiguous area of high rainfall over the Congo forest

compared to the TRMM-observed, three insular areas of high rainfall in this area.

In contrast to LST, the RAMS simulated precipitation seems to be less
sensitive to different land surface representations. All three experiments
produced similar spatial and temporal rainfall patterns. In Figure 4-16, however,
the GLC+LAI+VFC run produced a somewhat more realistic compact high rainfall

area over the Congo forest compared to the other two experiments.

78

Om>+_<4+0._0
new .040 due ”2:: m_2<m p.25 ocm moom L2 22w: :8: 99 8:32:83 52-8.5 pom—29m EmEoD .34. 93mm

88 88
302 “mm .3... 232. E 2. >02 mum .5_. >52 as). 5...

P

~ _

   

(me/tum) d!°°-'d

95.7.5540

 

 

 

 

 

. d

_ I.

No a

7 m.

d

)

to W

no no on.“

0
id no

two A ESE fi

3

 

 

79

9.3.3540 Em .o._o .“mE ace wszi 92.: can .225 so: 88 E 853605 863582 9+ 9:9“.

00m

00¢

cow

QCDF

oom_

00¢—

oom—

QQON

comm

 

80

4.7 Conclusions and discussions

In RAMS, the default land surface is represented by an outdated land
cover dataset and a set of prescribed parameters. Vegetation phenology is
unrealistically simplified by mathematical equations, which primarily are functions
of temperature and latitude. In comparison to the corresponding MODIS LAI and
VFC products, the spatial and seasonal vegetation dynamics of East Africa are
poorly represented by the default approach in RAMS. Overall, the default LAI and
VFC in RAMS are too homogeneous among different vegetation types across the
domain and are almost invariant over the year, which are very uncharacteristic of

the real world situation.

In this study, the GLCZOOO land cover product, which was evaluated in
Chapter 3 to have the best quality, and newly developed spatially and temporally
explicit MODIS LAI and VFC products are directly ingested into RAMS to improve
the representation of the land surface. Their impacts on regional climate
simulations were examined by three different RAMS experiments: DEF (default
OGE land cover and the built-in LAI and VFC), GLC (OGE is updated by
GL02000 but with the built-in LAI and VFC), and GLC+LAI+VFC (GLCZOOO plus
the MODIS LAI & VFC are used). Results show that the temporal and spatial LST
dynamics from the GLC+LAI+VFC run are greatly improved over the other two
experiments. Temporally, incorporating MODIS LAI and VFC enables RAMS to

capture the bimodal LST characteristics, especially the peak from August to

81

October in 2003. Spatially, RAMS’ capability to simulate ITCZ-related LST
migrations in the eastern domain is greatly improved. GLC2000 alone, however,
was not sufficient to improve the LST seasonal variation, but did change the

overall magnitude by about one degree.

In temporal terms, RAMS produces precipitation fairly well compared with
precipitation observations from the TRMM satellite. Two wet seasons and the
relatively dry period in 2003 were reasonably captured. Spatially, the general
major high rain features in the western domain and the dry areas in the eastern
domain are very well produced. Unlike LST, precipitation in RAMS seems to be
less sensitive to different land surface conditions. Rainfall outputs from the three
RAMS experiments: DEF, GLC, and GLC+LAI+VFC have closely similar
temporal and spatial characteristics as examined by the precipitation rate over

time and the accumulated rainfall in 2003.

The improvement of the daytime LST can be attributed to more realistic
description of the surface evapotranspiration process. The MODIS biophysical
products, especially VFC, capture the ITCZ related less-vegetated zone in the
eastern domain during the later part of the year (Figure 4-7). There is thus less
moisture evapotranspiration activity. The surface temperature in GLC+LAI+VFC
therefore increases compared to DEF and GLC with unrealistically covered

vegetation. Precipitation in RAMS is governed by large-scale fields and model

82

parameterizations. The improvement of the land surface representation is not

sufficient to change the simulated precipitation significantly.

It is worth mentioning that climatic variables from satellite observations
can help to improve the climate modeling studies in areas with sparse spatial and
temporal station measurements, as demonstrated in this study. TRMM
precipitation retrievals with three-hour temporal resolution is an excellent
validation variable for regional climate modeling. Though it was designed for
climate modeling studies, very few published works use these data except the
regular comparison with traditional data conducted outside of the climate
modeling community. Furthermore, the land surface temperature measured by
satellites has been shown to be very useful in GCM studies. For the first time,
this study compared the model-simulated LSTs with the newly developed MODIS
Terra and MODIS Aqua LSTs in a regional model (RAMS). Spatially and
temporally explicit, as well as diurnal, LST information is fully utilized in the

RAMS model in this study.

83

Chapter 5
Impacts of Land Cover Classification Accuracy on

Regional Climate Simulations
5.1 Introduction

Human activities are transforming the surface of the Earth at an
accelerated pace. Such disturbance of the land can affect the local, regional, and
global climate by changing the energy balance on the Earth’s surface and the
chemical composition of the atmosphere (Chase et al. 1998, Houghton et al.
1999, Pielke 2001). Over the past few decades, land use/cover has been widely
recognized as a critical factor mediating socioeconomic, political and cultural
behavior and global climate change (IGBP 1990, Lambin et al. 1999, Watson et
al. 2000). Numerous attempts have been made to understand past climate
changes and to project potential future climate changes by incorporating
reconstructed historical land cover changes and projected possible future land
cover changes into numerical simulations (Xue 1997, Pielke et al. 1999, Chase et
al. 2000, DeFries et al. 2002, Taylor et al. 2002). Recent studies have suggested
that land use/cover change has a first-order climate effect at the global scale

(Feddema et al. 2005).

No land cover dataset is one hundred percent accurate, even if developed

from the most advanced satellite images. Other factors, such as the classification

84

method, the sample size of evaluation data, and the inherent subjective
characteristics of classification, can increase the uncertainties contained in land
cover datasets. Such limitations have been recognized in the remote sensing
community, and therefore quantitative accuracy assessment has been
emphasized in most recent land cover classification research (Foody 2002).
Some target accuracy thresholds have recently been recommended in an
attempt to provide guidelines to the classification quality. Thomlinson et al.
(1999), for example, set as a target an overall accuracy of 85% with no class less
than 70% accurate. However, classification accuracy is usually interpreted
differently from the viewpoint of various users. The effect of land cover accuracy
for a particular application, such as climate modeling in this study, remains an
unanswered question. The accuracy targets commonly specified have largely not

been tested from the perspective of the operational use of land cover data.

The objective of this chapter is to examine how the classification accuracy
of a land cover dataset employed in a land surface scheme affects simulated
cumulative precipitation in a regional climate model. (Here, “regional climate
model” means a limited area model with high resolution, generally with grid
spacing less than 100 km, run for a simulation time of more than approximately
two weeks’ length, so that the initial atmospheric conditions have been forgotten
(Jacob and Podzun 1997).) The hypothesis of this study is that degradation of
land cover classification accuracy will not result in a significant change in

simulated regional climate until it reaches a certain threshold. By identifying this

85

threshold, the requirement of classification accuracy and the uncertainty
originating from the land cover classification in the regional climate simulation

analysis can be determined.

In addition, three follow-on experiments were conducted to investigate
how certain model parameterizations influence this effect. The parameterizations
examined in this study are the convection schemes and interior nudging, which
have been shown to influence the atmospheric response to surface boundary
forcing (Weaver et al. 2002, Castro et al. 2005). These follow-on experiments
help to illustrate how land classification error can propagate to factors that govern

precipitation in the climate model.

5.2 Methodology

The RAMS model is utilized to simulate the main wet season in East
Africa from March to May for the year 2003. Unlike in Chapter 4, only three
months are simulated in this chapter because the ensemble run in this study
requires significantly more computing resources. The study area is shown in the
Figure 3-1. To better represent the land surface characteristics, the default land
cover in the model (OGE) is replaced by the improved GLC2000, which ranked
the highest in the biophysical evaluation in Chapter 3. Based on GL02000,

classification error with increasing magnitude is then simulated. Cumulative

86

precipitation from simulations with different classification accuracies is then

examined.

The RAMS model and its configuration in this study have been described
in detail in Section 4.4 of Chapter 4. In the basic experiment, the Kain-Fritsch
(KF) convection scheme (Kain and Fritsch 1993) was used with no interior
nudging. In the three follow-on experiments, the effects of a different convection

scheme by Kuo (1974) and interior nudging are explored.

5.3 Land cover dataset

For an updated representation of the land surface cover, the GLC2000
dataset was used in these experiments, replacing the default OGE file. In order
to be able to use the biophysical parameters adopted from the Biosphere-
Atmosphere Transfer Scheme (BATS) (Dickinson et al. 1986), the GLC2000
classes were cross-referenced based on the results of multiple assessments
(Torbick et al. 2006). The predominant five non-water land cover types after
cross-referencing are presented in Table 4-1, with the most important biophysical
parameters listed. Combined, the five predominant types comprise 62.1% of the
total area, while ocean and inland water combined comprise 12.3%. The largest
inland water body in this area is Lake Victoria in the center of the model domain
(Figure 3-1). In the default LEAF-2 methodology,the original 1 km land cover

data were sampled to reduce the demand on computing resources used to

87

initialize the model. Only one pixel’s value, for example, is taken from a 5X5 pixel
block for a configuration of a 50 km horizontal spacing. As a result, details of the
input land cover are lost. In this study, detailed land cover input is needed, and

therefore the sampling strategy was modified to take every 1 km land cover pixel

in a grid cell.

5.4 Land cover accuracy

Land cover accuracy is commonly defined as the degree to which the
derived classification agrees with reality (Foody 2002). Here, classification error
at the1 km level was simulated as a random difference from GLCZOOO (Figure 5-
1a), the initial baseline land cover which was assUmed to be 100% accurate.
Specifically, random locations in the 1 km GLCZOOO were selected, and the
original land cover type at each of these selected locations was replaced by a
type randomly chosen from the five predominant types (Table 4-1). Only the land
cover types could be randomly altered since in practice it is less likely that water
bodies are misclassified. The five predominant land classes were chosen,
because it is reasonable to assume that they have more chance to be

misclassified than less abundant classes.

88

I I I
I
(C)
I I
I
I
' I
I I I
I
I I
- I
I II
I
(d) I I n
I l I I
I I
I
_ I
I
I
‘3'. - I - I
I I I
I I I. - I.
I I I
I I . J
I I r I
Land Cover Types (3) ' ' .
I I I
- Ocean _ Evergreen Broadbaf Forest I I :
- Inland Water - Deciduous Broadlear Forest I I
ET] Evergreen BroadleafTree -WoodIend . I I I
[:1 Semi-desert 7 Wooded Grassland . " I
Deciduous Shrub [j Open Shrubland ' I I
. . ,__ " I - I
- Crop/Muted Fanning . Grassland I .
-IrrIgaIedCrop -BareGround I‘Iti. -, I}. III-
' I
[Jaw/Marsh -uman . _ 5 I II-_

 

 

 

 

 

 

 

 

 

Figure 5-1. Cross referenced GL02000 with 1km resolution (a) and 50km
resolution (b), and simulated land cover classification errors: 10% (c), 30% (d),
and 50% (e). Land cover types in (b), (c), (d), and (e) only represent the biggest

patches in grid cells. See texts for more details.

89

Classification errors with magnitudes ranging from 5% to 50% at 5%
intervals were generated. The magnitude of error is determined by the proportion
of converted pixels in the 1 km GLC2000. Fifty percent error was the maximum
level tested as it is assumed that most land cover products could reach 50%
accuracy levels. These 1 km land covers with degraded classification accuracies
were used to initialize the land surface in sequential RAMS simulations, and the

behavior of the simulated results was examined.

In Figure 5-1, 10% (Figure 5-1c), 30% (Figure 5-1d), and 50% (Figure 5-
1e) classification errors are presented. For the sake of clarity, only the most
predominant patches in each 50 km RAMS grid cell are illustrated because
simulated errors and their gradual increase would be hard to see at a 1 km
resolution (Figure 5-1a). Figure 5-1b presents the land cover at a 50 km
resolution, assumed to be 100% accurate, with each grid cell showing only its
most predominant cover type. Figures 5-1c, d, and e show those model grid cells
with the predominant land cover changed following the introduction of random

classification errors.

Despite random selection at the 1 km resolution, the errors do not appear
to be distributed randomly over the domain when viewed at 50 km level. Instead,
they tend to occur at the transition zones between major types (Figure 5-1 b),
where it is likely that two land cover types are approximately equal in frequency

within the grid cell. Converting a few pixels may alter which land cover type is the

90

predominant patch. For example, most changes in evergreen broadleaf forests in
Figures 5-1c, d, and e occur at the edge of the Congo forest. For grid cells with
strongly dominant types, such as the Congo forest, random errors are less likely
to change the dominance of the biggest patch. Transitions to woodland appear to
have a higher frequency than do the other four types (see especially in Figure 5-
1a). This is due to the woodland appearing in a fragmented arrangement (Figure
5-1a). Similarly, transitions to water as the largest patch show up at the edges of
lakes and the ocean, as seen in Figures 5-1c, d, and e, although water was not

considered in the process of randomization (Table 4-1).

In each experiment, RAMS was run 11 times, each with different amounts
of classification error ranging from zero to 50 percent. The effect of classification
accuracy on simulated regional climate is then examined by comparing the
behavior of simulated precipitation within this range of accuracies to determine
patterns. Then, patterns of behavior are compared across experiments to

investigate the impacts of model parameterizations.

5.5 Results

5.5.1 Basic experiment

In the basic experiment, RAMS was run with the KF convection scheme

and without interior nudging. As in Chapter 4, the performance of RAMS over

91

March — May 2003 is first assessed by comparing the RAMS-simulated

accumulated precipitation to the TRMM observations.

Figure 5-2 shows both the simulated accumulated rainfall from RAMS and
the observed accumulated rainfall from TRMM. As discussed in Section 4.6.2 of
Chapter 4 for the whole year of 2003, RAMS underestimated precipitation in
some areas, especially near the left and right boundaries, which may be due to
the effect of boundary nudging. But it captured some major features, such as
over the Congo forest. The spatial distribution of simulated precipitation is fairly
similar to the TRMM observations, especially considering that no attempt was
made to “tune” model parameters and that our configuration of RAMS has a
lower spatial resolution (50 km vs. about 27 km). In Figure 5-3, precipitation is
compared over time. Domain-averaged daily precipitation is normalized to 1 for
the sake of comparison for the study area. The correlation coefficient is 0.336 for
the whole time period, and 0.438 when the spin-up time of the first 20 days is
omitted. Fidelity to observation improved over time, and the cessation of the

“long rains” (day 77) is well-replicated.

92

TRMM

RAMS

, n—
100 200 300 400 500 600 700 BOD 900

   

Figure 5-2. Spatial comparison of simulated accumulated precipitation (mm) in
RAMS and that from TRMM for March — May, 2003

 

0.035 I

I —--—RAMS .l'

_ 0.025
I! - .
'5 0020 '0 .
g . \-
8 ' I.
5 0.015 \- .
To _ /~. \
E O. _ I
'5 0.010
2 _ . \X / W .
I \ I. o
0.005 _ _ .i‘\

0.000

 

 

 

 

30 4o 50 60 70 80 90
Simulation Day

Figure 5-3. Temporal comparison of simulated cumulative precipitation in RAMS
and from TRMM

The differences in precipitation between the simulation without land cover
errors and simulations with errors (5%, 10% 45%, 50%) were examined. For
the convenience of discussion, let R00, R05, R10 R45, R50 denote these 11
runs and R05-R00, R10-R00 R45-R00, R50-R00 denote the differences
between runs. R10-R00, R30-R00, and R50-R00 are presented in Figure 5-4. If
classification accuracy does not have any impact on simulated precipitation, then
these differences are expected to be close to zero. However, as illustrated in
Figure 5-4, precipitation differences are not minute. The impact on precipitation
increases as classification accuracy worsens. It is also noticeable that most of
the largest differences occur in the Lake Victoria area, even though errors are

scattered across the whole domain (Figure 5-1).

94

 

 

 

I
(a) «a:
. .
I
I
- d a
I B
a
I Cs .1
0 V B

 

 

I I
I Jo
I I
B
--I Q
R. d

 

 

 

 

 

 

 

—40 -30 -20 -1O 1O 20 30 4C 50

Figure 5-4. Differences of accumulated precipitation (mm) between simulation
without land cover error and simulations with 10% error (R10-R00, a), 30% error
(R30-R00, b), and 50% error (R50-R00, c).

95

Although a full investigation of this is beyond the scope of this dissertation,
it is likely that the general spatial pattern of changes in precipitation shown in
Figure 5—4 is due to the mechanism stated by Charney et al. (1977) and followed
by other researchers (Lofgren 1995, Xue 1997, Wang et al. 2004). In this
mechanism, a change in land surface parameters (e.g. albedo, vegetation
fractional cover) alters the energy budget of the coupled surface-atmosphere
system. Particularly at low latitudes, reduced heating of the atmosphere,
resulting from increased surface albedo, leads to a relative sinking motion and
reduced precipitation, while decreased surface albedo and increased

atmospheric heating have the opposite effect.

In the domain considered here, the unperturbed GLC2000 land cover has
nearly solid evergreen broadleaf forest in the western part of the domain, the
class with the lowest surface albedo (Table 4-1). Thus, the insertion of random
errors into this land cover class will necessarily increase the surface albedo,
resulting in reduced precipitation in this area. Conversely, the region surrounding
Lake Victoria is initially dominated by the crop/mixed farming class (in the
GLC2000 dataset), which has the highest surface albedo of any of the classes
used in the random replacement process, so the imposition of random errors
reduces the surface albedo in this region. Combined with the ready access to
water evaporating from Lake Victoria itself, this can lead to an increase in
precipitation. As in Lofgren (1995), the heating of the atmosphere near Lake

Victoria due to reduced surface albedo and the changes of other parameters is

96

likely compounded by the release of the latent heat of condensation associated

with the increased precipitation.

Three measures were utilized to depict the precipitation differences
between runs (Figure 5-4). The first measure is the maximum absolute difference
(both positive and negative), which highlights only one hot spot. It represents the
largest possible difference caused by land cover errors, but it does not give
information on the overall differences. The other two measures used are the
mean absolute difference and the standard deviation calculated over the whole
domain. They characterize the overall magnitude and variation of the difference.
The maximum absolute differences for R10-R00, R30-R00, and R50-R00 were
30.6mm, 56.7mm, and 84.4mm. The mean absolute differences were 4.6mm,
6.7mm, and 10.5mm while the standard deviations were 6.7mm, 9.5mm, and
14.7mm. The three measures all indicate an increase in precipitation difference

as land cover accuracy decreases.

These three measures can evaluate precipitation differences against a
range of classification errors (5% to 50%). For the basic experiment, the black
lines in Figure 5-5 show the behavior of precipitation difference for this range of
classification errors by illustrating the maximum and mean absolute differences
and the standard deviation. In Figure 5—5a, the maximum absolute difference
from the basic experiment increases from 34.8mm for 5% error to 84.4mm for

50% error. In Figure 5-5b, the mean absolute difference increases from 5.5mm to

97

10.5mm. The standard deviation increases from 7.8mm to 14.7mm (Figure 5-50).
From these three plots, it is evident that precipitation differences increase with an
increase in land cover errors in RAMS. lmportantly, when the errors are less than
20%, the plots are relatively flat, and when errors are larger than 20%, the
differences increase sharply. This indicates that a classification error of less than
20% has little effect on the simulated precipitation in this particular experiment.
The accuracy target of 85% commonly specified in the land cover production
community can meet the requirements of regional climate modeling. If the land
cover accuracy is less than 80%, however, its effect on climate simulation and

propagation of uncertainty should be examined.

In the basic experiment results shown in Figure 5-5, the level area below
20% errors has non-zero differences. This is especially obvious for 5% error
level. Adding this small amount of classification error causes some notable
precipitation differences, which might be due to random noise. Above the 20%
threshold, the signal rises above the noise. It is noticeable that there is a slight
leveling off of the differences above the 40% level, which might be due to a

saturation effect of classification errors.

98

 

\

(a) \-

I

Monrnum Abwlute Difference (mm)
a
\ '. I] ‘- )1
b
0
\
//
'\
\.

 

 

 

 

 

 

 

 

 

 

 

‘ \\. v/' V
20 _ , \ ./ ,,.—/ V
10% 20% 30% 40% 50%
Lend Cover En'or Percentage
10- s\,/’
E (b)
v a _
8
O - A.-/" //°“‘Ao
a \/'\/ _/_A’/,//
4 p //0—\‘,/ ’/ V
i ./’. A v’—'V/V/
: ‘ W //
g 2 h v/'/
r/I/z . . . .
10% 20% 30% 50%
Land Cover Enor Percentage
w-
. (c) /\I/-
312- . .
g / — -— Basrc Experiment
. .l / _._Fd.ow-m1
, -\ /.\_/' /-~.-...-. _._ Follow-on 2
g 0 - ' A;,/"\‘-/°' ' — v— F ollow-on 3
z, - "i A I“ v/
. . , 2 v——-V‘/
A” v/"/
3 * ,//
. Y/T/A l A I 4 l A L
10% 20% 30% 40%
Lend Cover Error Percentage

Figure 5-5. Maximum absolute differences (a), mean absolute differences (b) and
standard deviations (c) from the basic experiment (Kain-Fritsch), follow-on 1
(Kuo), follow-on 2 (Kain-Fritsch with interior nudging), and follow-on 3 (Kuo with
interior nudging).

99

5.5.2 Follow-on experiments

To test the effect of model parameterizations on atmospheric response to
land cover accuracy, internal nudging and a different convection scheme are
investigated. A different convection scheme may dramatically change the surface
energy and moisture budget and, hence, surface feedback to the atmosphere.
The KF convection scheme, which was used in the basic experiment, is known to
produce more precipitation than the Kuo scheme, especially in areas of steep
terrain (Castro et al. 2002, 2005). Nudging is used to relax the model solution
towards the input reanalysis data continuously at each time step by adding
artificial tendency terms (based on the difference between the two states) to the
prognostic equations. With interior nudging, the surface boundary conditions tend
to have weaker control on the vertical motion and distribution of precipitation,
compared to no interior nudging (Weaver et al. 2002, Castro et al. 2005).
Therefore, both convection scheme and interior nudging may influence the effect
of land cover accuracy on simulated precipitation in RAMS. Other model aspects
can also modify the influence of surface forcing on simulated precipitation;
however, nudging and convection schemes are often used for such evaluations

(e.g., Weaver et al. 2002, Castro et al. 2005).

In the basic experiment described previously, the KF convection scheme
is used with no interior nudging, which allows the model to have a stronger
response to surface boundary forcing. In follow-on experiment 1, the Kuo

scheme was used with no interior nudging. ln follow-on experiment 2, the KF

100

scheme was used with interior nudging applied and in the follow-on experiment
3, the Kuo scheme was used together with interior nudging. When interior
nudging was used, the time scale was set to one day, which is larger than that
specified in the RAMS User Guide (Castro et al. 2005). In each of these three
follow-on experiments, RAMS was run 11 times, each with a different amount of

classification error ranging from zero to 50%, similar to the basic experiment.

In Figure 5-6, accumulated precipitation is presented for the basic
experiment and the three follow-on experiments, all with no classification error.
As expected, the Kuo scheme produces much less precipitation over the whole
domain. The major peak over the Congo forest, which is seen in experiments
with the KF scheme and in satellite observations (Figure 5-2), is not shown
clearly in experiments with the Kuo scheme. There is not much difference
between these two schemes over dry areas, where both schemes tend to
underestimate the precipitation. The interior nudging seems to have little effect

on the accumulated precipitation.

In each follow-0n experiment, precipitation differences between
simulations with and without classification errors were investigated by examining
the maximum and mean absolute differences and the standard deviation as in
the basic experiment. In Figure 5-5, the behaviors of these three measures
against a range of classification errors are presented for both the basic

experiment and the three follow-on experiments. Precipitation in the follow-on

101

experiments is much less sensitive to classification errors, especially when the
Kuo scheme was used. With both convection schemes, interior nudging tends to
reduce this sensitivity. Interestingly, standard deviation plots for interior nudging
are very close to straight lines. This may be due to the effect of interior nudging
reducing the strength of small-scale variability, which has also been reported in

other studies (e.g., Weaver et al. 2002).

Kain-Fritsch Convection Scheme Kuo Convection Scheme
(Basic Experiment) (Follow-on 1)

    

' l

 

a...
I 000
700
000
Kain-Fritsch Convection Scheme eon Kuo Convection Scheme
Interior Nudging I Interior Nudging
(Follow-on 2) I“)0 (Follow-on 3)
300
i...
'100

   

Figure 5-6. Simulated accumulated precipitation (mm) with different model
parameterizations. Results from runs without land cover error are presented.

5.6 Discussion

In RAMS, each land cover type is represented by a suite of biophysical
variables: albedo, leaf area index (LAI), fractional vegetation cover, etc. These
biophysical variables determine the energy and moisture exchanges between the
land surface and overlying atmosphere. Thus, the effect of land cover
classification accuracy on simulated precipitation is ultimately controlled by the
changes in the biophysical variables. Therefore, the effect of classification
accuracy relies on how the surface scheme (LEAF-2 in this study) defines these
biophysical variables for each type. As the biophysical parameters of different
land cover types become more differentiated, the effect observed in previous
sections will be more pronounced. In the hypothetical case where all land covers
have exactly the same biophysical characteristics, classification accuracy will not

have any effect on simulated precipitation.

In Figures 4-6 and 4-7, the default RAMS LAl and VFC are compared to
the MODIS LAI and VFC respectively. It is evident that the LAI and VFC in RAMS
are unrealistically uniform over most of the domain, with several regions poorly
represented. Other biophysical parameters in RAMS version 4.4, such as albedo,
may also have this characteristic of being overly homogeneous since they are
defined by simple mathematical functions. Therefore, it is reasonable to expect
that the impact of classification accuracy on simulated precipitation might be

even greater than described in this study.

103

As shown in the previous sections, a land cover accuracy lower than 80%
can substantially affect simulated precipitation, especially when the surface has a
greater control of the atmosphere. This effect becomes stronger as the accuracy
decreases. Although an 85% accuracy target has already been recommended for
land cover production, in reality this target is rarely obtained (Trodd 1995). For
example, the IGBP DlSCover Land Cover product, a land cover layer from GLCC
with global coverage, has an overall accuracy of 66.9%, which is considerably
lower than the specified target (Scepan 1999). The default land cover dataset in
RAMS (OGE) does not come with an accuracy estimate. Global accuracy for the
newly developed MODIS Land Cover product (V003) is stated to be
approximately 70 — 80%
(httpzllgeography.bu.edu/landcover/userguidelc/consistent.htm). When these
global land cover products are used for a specific region, such as East Africa, the
accuracy levels can be much lower than the global accuracy. Therefore, caution
is needed when using global land cover products at regional to local levels. It
should also be mentioned that global land cover products are usually developed
for land cover identification or other general use. In order to be used in SVAT
schemes in climate models, they usually need to be cross-referenced with

biophysical parameters, which can add additional uncertainties.

These global products, however, have their advantages. The importance

of classification accuracy is well recognized by producers. Quantitative

104

evaluation is therefore conducted to provide guidelines for users. Many land
cover datasets employed in climate modeling studies do not come with accuracy
information. This is especially true for historically reconstructed and future
projected land cover datasets that are often employed to examine the impact of
human activities on climate. Historical land covers are usually derived from
existing maps and other indirect evidence, while future projected land covers are
often developed from spatial models that simulate how changes in land use are
likely to affect land cover. There are simply not many options for accuracy
assessment of these types of land cover datasets. Historical and future land
cover datasets are usually used to simulate time periods that are decades or
even centuries long, much longer than the three months simulated in this study.
The impact of land cover accuracy may well increase over these longer time
frames. Uncertainties in those input land cover datasets may cause great

uncertainties in the output in climate models.

There are aspects in this work that can be further explored. One is the
strategy that was used to simulate classification errors. It is assumed that
classification errors occur randomly over space. In reality, they are more likely to
occur in areas with greater land surface heterogeneity and not in homogeneous
landscapes such as the Congo forest. Also, in this study the original land cover
types are replaced by random types from the five predominant types without

considering the biophysical similarities between types. Land cover types with

105

similar physical appearances or similar spectral features in satellite images are

more likely to be misclassified.

A second aspect relates to the configuration of RAMS. Factors such as
horizontal grid spacing and multiple nested grids may influence the effect
observed in this paper. The mosaic method of accounting for sub-grid variability
in land cover does not take into account certain factors. Notably, latent, sensible,
and radiative heat fluxes will be dependent on the land cover and on the
characteristics of the air in the planetary boundary layer. The boundary-layer
atmospheric characteristics are likely to be spatially correlated with the land
cover, but the mosaic approach does not account for this and thus will miss the
non-linear effects on fluxes. The situation is further complicated by sub-grid
mesoscale circulations that can be forced by land cover heterogeneity (e.g.,
Weaver and Avissar 2001). These factors can be sensitive to the scale over
which land cover is altered in addition to the model grid spacing. Further

investigations considering these factors are needed.

5.7 Conclusions

Human activities have substantially modified the Earth’s surface in the
past and will continue to do so in the future. The impact of human activities such

as land cover change on the regional and global climate can be studied using

106

climate modeling techniques. Land cover datasets, often derived from remote
sensing images, are widely used in land surface schemes in climate models to
describe the physical surface conditions. These datasets are not perfect, and
their value is a function of classification accuracy. In the land cover production
process, quantitative accuracy assessment has almost become a required
procedure. However, the uncertainty arising from the accuracy of land cover
classification and its impact on simulated climate have largely been ignored in

climate modeling research.

In this paper, the Regional Atmospheric Modeling System was utilized to
study the impact of land cover accuracy on simulated precipitation for the East
Africa region. Classification errors were simulated as random alterations to the
land cover dataset used in this study - GLC2000. The behavior of simulated
accumulated precipitation over a three month period was then examined over a
range of land cover errors (zero to 50%). It was found that, when the surface
boundary had greater control on the overlaying atmosphere, land cover accuracy
under 80% had a strong effect on simulated precipitation. As land cover accuracy
worsened, this effect became even stronger. This effect on simulated
precipitation was shown to be moderated by model parameterizations such as
convection schemes and interior nudging, which affect the strength of control that
the surface exerts on the atmosphere. When the Kuo convection scheme was
used, RAMS severely underestimated the precipitation over the entire domain,

and the land cover accuracy had little effect on simulated precipitation. With

107

interior nudging activated, the effect of land cover accuracy also decreased, even

though the overall magnitude of precipitation was affected only slightly.

Based on the results of this study, it can be concluded that land cover
datasets can meet general needs in climate modeling research if the
commonly recommended 85% accuracy target is obtained. In reality,
however, this is usually not the case. The reliability of land cover datasets
needs to be examined in climate modeling research, especially those using
historically reconstructed or future projected land covers for long-term

simulations.

108

Chapter 6

Conclusions and Future Research

The accurate description of the spatial heterogeneity and seasonal
variation of the landscape has been recognized as a key component in
meaningful regional and global climate modeling. Land cover datasets and
related biophysical variables are often used to represent the land surface in the
state-of—art climate models e.g., the Regional Atmospheric Modeling System
(RAMS) in this study. However, the land cover dataset in RAMS (OGE) was
developed more than a decade ago on the basis of satellite images with limited
capabilities, and the biophysical characteristics including leaf area index (LAI)
and vegetation fractional cover (VFC) are depicted by a lookup table and simple

mathematical equations.

This dissertation is devoted to improve the regional climate modeling in
East Africa by better representing the land surface using various newly
developed remote sensing products. Several new land cover products, together
with the default OGE, are evaluated by a new statistical measure designed
specifically for climate modeling purposes. The land cover with the best quality is
selected to replace OGE land cover. LAI and VFC imageries from MODIS are
incorporated directly in RAMS. The examination of the impacts of the improved
land surface becomes possible by utilizing the land surface temperature (LST)

and precipitation from remote sensing retrievals. For regions with very sparse

109

weather station observations such as East Africa, the focus of this study, model
evaluations are extremely challenging tasks. F urtherrnore, the uncertainty from
the land cover classification and its propagation in climate models, which have
not been fully addressed, are investigated in this study. The threshold of
classification accuracy is then identified, which can be used as a guideline for the
land cover mapping in the remote sensing community as well as for various
climate modeling studies, when historically reconstructed and future predicted

land cover datasets are employed.

Assuming better classification has more consistent biophysical
characteristics within each class, the new statistical measure Q developed for
land cover classification accuracy assessment in this study aggregates the
within-class LAl variation across the domain and over any time period. Unlike
traditional evaluation methods which are based on ground-truthing data, Q
utilizes newly developed MODIS LAI products and thus there is no sampling
error. Another major advantage of Q is that it considers the consistency of
classification over any specified time period when the seasonal variation of
vegetation is important. More importantly, Q is designed on the basis of the
mosaic approach of representing land surface heterogeneity in climate models. It
is therefore more suitable for evaluating land cover datasets for climate modeling

studies.

110

In terms of Q, GLC2000 ranks higher than MODIS land cover, OGE, and
LEAF at three different spatial resolutions: 30 km, 50 km, and 100 km.
Particularly, GLCZOOO is significantly better than LEAF at both 30 km and 50 km.
50 km is the RAMS grid spacing in this study. Although, the newly developed
MODIS land cover is better than OGE and LEAF at all three levels, the difference
is not statistically significant. This evaluation suggests that the default land cover
in RAMS needs to be replaced by GLC2000. In addition, this study finds as
quadrate size decreases, the difference between land cover products tend to

decrease.

The built-in LAI and VFC are compared spatially and temporally with
corresponding MODIS products. Spatially, the default LAI and VFC are too
homogeneous to differentiate distinct land surface types across the domain.
Temporally, simple cosine functions have been demonstrated to be extremely
unrealistic. For low-latitude regions, the prescribed LAI and VFC are almost
invariant as time evolves in a year. To better represent the land surface
conditions, spatially and temporally explicit MODIS LAI and VFC imageries are
directly ingested in RAMS. The OGE land cover is replaced by GLC2000 as well.
With insufficient weather station data available, the MODIS LST and TRMM
precipitation data are used to evaluate the model performance and examine the
impacts of improved land surface conditions. This study finds that the improved
land surface improves the LST characteristics spatially and temporally but has

little influence on simulated precipitation. Specifically, the bimodal feature of the

111

LST seasonal variation, which is completely missed in the default configurations,
is fully captured when MODIS LAI and VFC are incorporated. Regarding to the
spatial characteristics, the ITCZ-related seasonal migration of LST in the eastern
domain has been greatly enhanced. Both MODIS Terra and Aqua LST are used
for the first time to evaluate the surface impact on diurnal LST characteristics.
This study finds that diurnal LST cycles in the second half of the year are slightly
improved due to the new land surface representation. GLCZOOO alone, however,
is not sufficient to modify the seasonality of LST but is able to change the overall

magnitude by about one degree.

In order to truly improve the land surface conditions in climate modeling,
the uncertainty originating from the land surface and its propagation in climate
models need to be better understood. The impact of the land cover classification
accuracy on regional climate simulations remains a relatively unanswered
question. In this study, a range of random classification errors are simulated and
added to the baseline GLCZOOO. Differences in accumulated precipitation from
March to May 20003 are examined. This study finds that the overall classification
accuracy under 80% has a strong effect on simulated precipitation, especially
when the land surface has a greater control on the overlying atmosphere. This
effect becomes much stronger as the classification continues to worsen. In
addition, model parameterizations, e.g. convection schemes and interior nudging,
can influence this effect by mitigating the driving forcing of the land surface. This

study suggests that land cover datasets developed from remote sensing

112

imageries can meet the requirement of regional climate modeling if the overall
classification accuracy reaches 85%. In reality, however, this is usually not the
case especially when the land cover datasets in climate modeling studies are
historically reconstructed or future projected. In order to make convincing
simulations, the reliability of the land cover dataset and the uncertainty

propagation thus need to be fully investigated.

In the future, extended research is necessary to address the following
issues related to this study. Firstly, the property of the Q statistic needs to be
further studied. So far, its theoretical distribution and ranges under different
circumstances are still not quite clear. Without knowing these properties, it is
hard to tell whether a specific Q value is absolutely high or low. For land cover
classification, a specific range of Q as a quality control is necessary if Q is
applied in the land cover production community as an accuracy guideline.
Secondly, other newly developed biophysical products such as MODIS albedo,
have the potential to be incorporated into RAMS until the look-up table approach
is completely discarded. Before this is achieved, the land surface is still partially
improved. Of course, remote sensing techniques extracting other biophysical
variables such as rooting depth need to be further developed. Thirdly, for the
land cover classification accuracy, random classification errors are simulated in
this study. In reality, misclassification may more likely occur between closely
similar types such as shrubland and woodland not shrubland and forest. More

realistic misclassification simulation approach thus needs to be implemented.

113

ANOVA
AVH RR
BATS
CCM
CLIP
EVI
GCM
GLCZOOO
IGBP
ITCZ
LAI
LEAF
LST
LUCC
MCP
MODIS
NCEP
NDVI
OGE
RAMS
SST
SVAT
TRMM
UTC
VFC

APPENDIX

Acronym List

One Way Analysis of Variance

Advanced Very High-Resolution Radiometer
Biosphere-Atmosphere Transfer Scheme
Community Climate Model

Climate-Land Interaction Project

Enhanced Vegetation Index

Global Circulation Model

Global Land Cover 2000

International Geosphere-Biosphere Programme
lntertropical Convergence Zone

Leaf Area Index

Land Ecosystem-Atmosphere Feedback
Land Surface Temperature

Land Use/Cover Change

Multiple Comparison Procedure

Moderate Resolution Imaging Spectroradiometer
National Centers for Environmental Prediction
Normalized Difference Vegetation Index
Olson Global Ecosystem dataset

Regional Atmospheric Modeling System

Sea Surface Temperature
SoiI-Vegetation-Atmosphere Transfer
Tropical Rainfall Measuring Mission
Coordinated Universal Time

Vegetation Fractional Cover

114

BIBLIOGRAPHY

Avissar, R. and Pielke, R. A. (1989). A parameterization of heterogeneous land
surfaces for atmospheric numerical models and its impact on regional
meteorology. Monthlv Weather Review, 117, 2113-2136.

Bastidas, L. A., Gupta, H. V., Sorooshian, S., Shuttleworth, W. J., Yang, 2. L.
(1999). Sensitivity analysis of a land surface scheme using multicriteria methods.
Journal of Geophysical Research, 104, 19481-19490.

Becker, F. (1987). The impact of spectral emissivity on the measurement of land
surface temperature from a satellite. lnternaiional Jogrnal of Remote Sensing, 8,
1509-1522.

Betts, R. A. (2000). Offset of the potential carbon sink from boreal forestation by
decreases in surface albedo. Nature, 408, 187-190.

Bonan, G. B., Pollard, D., Thompson, S. L. (1992). Effects of boreal forest
vegetation on global climate. Nature, 359, 716-718.

 

Bonan, G. B. (1996). A land surfice mcfiel (LSM version 1.0) for ecolglical.
hygrolofial. and atmospheric stfiies: Techincal descrigtionJand LLSBT'S guide.
NCAR Tech. Note NCAR/TN-417 + STR, 150 pp.

Brown, J. F., Loveland, T. R., Ohlen, D. 0., Donald, 0., Zhu, Z. (1999). The
global land-cover characteristics database: the user’s perspective.
Photogrammetric Engineering and Remote Sensing, 65, 1069-1074.

Carlson, T. N. and Ripley, D. A. (1997). On the relationship between NDVI,
fractional vegetation cover, and leaf area index. Remote Sensing of Environment,
62, 241-252.

Castro, C. L., Cheng, W. Y. Y., Beltran, A. B., Marshall, C. H., Jr., Pielke, R. A.,
Cotton, W. R. (2002). The incorporation of the Kain-Fritsch cumulus
parameterization scheme in RAMS with a terrain-adjusted trigger function. file
Fifth RAMS Users and Related AgglicatkflWorkshog, Santorini, Greece.

Castro, C. L., Pielke, R. A., Leoncini, G. (2005). Dynamical downscaling:
Assessment of value retained and added using the Regional Atmospheric
Modelding System (RAMS). Jp‘urnal of GeophysicaflResearch, 110, D05108,
doi:10.1029/2004JD004721.

115

Charney, J., Quirk, W. J., Chow, S-H, Kornfield, J. (1977). A comparative study
of the effects of albedo change on drought in semi arid regions. Journal of the
Atmospheric Sciences. 34, 1366-1385.

Chase, T. N., Pielke, R. A., Kittel, T. G. F., Nemani, R., Running, S. W. (1996).
Sensitivity of a general circulation model to global changes in leaf area index.
Journ_a_| of Geophysical Research, 101, 7393-7408.

Chase, T. N., Pielke, R. A., Kittel, T. G. F., Baron, J. S., Stohlgren, T. J. (1998).
Potential impacts of historical land cover changes on global climate in northern
winter. Climate Dynamics, 16, 93-105.

Chase, T. N., Peilke, R. A., Kittel, T. G. F., Nemani, R., Running, S. W. (2000).
Simulated impacts of historical land cover changes on global climate in northern
winter. Climate Dynamics, 16, 93-105.

Chen, C. and Cotton, W. R. (1983). A one-dimensional simulation of the
stratocumulus-capped mixed layer. Boundary-Layer Meteorology, 15, 289-321.

Choudhury, B. J., Ahmed, N. U., ldso, S. B., Reginato, R. J., Daughtry, C. S. T.
(1994). Relations between evaporation coefficients and vegetation indices
studied by model simulations. Remote sensingoflinvironment, 50, 1-17.

Cihlar, J. (1997). GCOS/GTOS plan for terrestrial climate related observations.
Version 2. GCOS Report. Geneva: World Meteorological Organization.

Cihlar, J. (2000). Land cover mapping of large areas from satellites: status and
research priorities. lnternationail Journal of Remote Sensing, 21, 1093-1114.

Cohen, W. B., Maiersperger, T. K., Yang, 2., Gower, S. T., Turner, D. P., Ritts,
W. D., Berterretche, M., Running, S. W. (2003). Comparison of land cover and
LAI estimates derived from ETM+ and MODIS for four sites in North America: a
quality assessment of 2000/2001 provisional MODIS products. Remote Sensing
of Environment, 88, 233-255.

Congalton, R. G. (1991). A review of assessing the accuracy of classification of
remotely sensed data. Remote SensinLof Environment, 37, 35-46.

Cotton, W. R., et al. (2003). RAMS 2001: Current status and future directions.
Meteorology and Atmospheric Physics, 82, 5-29.

Cunnington, W. M. and Rowntree, P. R. (1986). Simulations of the Saharan
atmosphere-dependence on moisture and albedo. Quarterly Journal of the Royal
Meteorolggical Society, 112, 971-999.

116

Curran, P. J. (1983). Multispectral remote sensing for the estimation of green leaf
area index. Philosophical Transactions of the Royal Society (London). 309, 257-
270.

DeFries, R., Hansen, M., Steininger, M., Dubayah, R., Sohlberg, R., Townshend,
J. (1997). Subpixel forest cover in Central Africa from multisensor, multitemporal
data. Remote Sensing of Environment, 54, 209-222.

DeFries, R. S., Bounoua, L., Collatz, G. J. (2002). Human modification of the
landscape and surface climate in the next fifty years. Glogarl Change Biology, 8,
438-458.

Deardorff, J. W. (1978). Efficient prediction of ground surface temperature and
moisture, with inclusion of layer of vegetation. Journal of Geophysical Research.
83, 1889-1903.

Di Gregorio, A. and Jansen, L. J. M. (2000). Land cover classification system
(LCCS). Rome: F.A.O.

Dickinson, R. E., Henderson-Sellers, A., Kennedy, P. J., Wilson, M. F. (1986).
Biosphere-atmosphere transfer scheme for the NCAR community climate model.
NCAR Technical Note NCAR/TN;275+STR, Boulder, CO, USA.

Dickinson, R. E., Henderson-Sellers, A., Rosenzweig, C., Sellers, P. J. (1991).
Evapotranspiration models with canopy resistance for use in climate models, a
review. Agricultural and Forest Meteorology, 54, 373-388.

Dinku, T., P. Ceccato, E. Grover-Kopec, M. Lemma, S. J. Connor, C. F.
Ropelewski (2007). Validation of satellite rainfall products over East Africa’s
complex topography. International Journal of Remote Sensing, 28, 1503-1526.

Ducharne, A. and K. Laval (2000). Influence of the realistic description of soil
water-holding capacity on the global water cycle in a GCM. Journal of Climate,
13, 4393-4413.

Feddema, J. J., Oleson, K. W., Bonan, G. B., Mearns, L. 0., Buja, L. E., Meehl,
G. A., Washington, W. M. (2005). The importance of land-cover change in
simulating future climates. Science, 310, 1674-1678.

Foody, G. M. (2002). Status of land cover classification accuracy assessment.
Remote Sensipgof Environment, 80, 185-201.

Franks, S. W., Beyen, K. J., Quinn, P. F., Wright, I. R. (1997). On the sensitivity
of soil-vegetation-atmosphere transfer (SVAT) schemes: equifinality and the
problem of robust calibration. Agricultural and Forest Meteorolggy, 86, 63-75.

Friedl, M. A., Mclver, D. K., Hodges, J. C. F., Zhang, X., Muchoney, D., Strahler,
A. H., Woodcock, C. E., Gopal, S., Schneider, A., Cooper, A., Baccini, A., Gao,

117

F., Schaaf, C. (2002). Global land cover mapping from MODIS: algorithms and
early results. Remote Sensiryaof Environment, 83, 287-302.

Fritz, S., Bartholome, E., Belward, A., Hartley, A., Stibig, H.-J., Eva, H., et al.
(2003). Harmoniz_ation, mosaicking, and production of the Global Land Cover
2000 database. lspra, Italy: Joint Research Center (JRC).

Giorgi, F. and Mearns, L. O. (1999). Introduction to special section: regional
climate modeling revisited. Journal of Geophysical Research, 104, 6335-6352.

Giri, C., Zhu, Z. and Read, B. (2005). A comparative analysis of the Global Land
Cover 2000 and MODIS land cover data sets. Remote Sensing of Environment.
94, 123-132.

Gutman, G. and lgnatov, A. (1998). The derivation of the green vegetation
fraction from NCAA/AVHRR data for use in numerical weather prediction models.
International Joarnal of Remote Sensing, 19, 1533-1543.

Hahmann, A. N. and Dickinson, R. E. (1997). RCCM2-BATS model over tropical
south America: applications to tropical deforestation. Journal of Climate, 10,
1944-1964.

Henderson-Sellers, A., Dickinson, R. E., Durbidge, T. 8., Kennedy, P. J.,
McGuffe, K., Pitman, A. J. (1993). Tropical deforestation: modeling local- to
regional-scale climate change. Journal of Geophysical Resea_rc_h, 98, 7289-7315.

Houghton, R. A., Hackler, J. L., Lawrence, K. T. (1999). The US. carbon budget:
contributions from land-use change. Science, 285, 574-578.

Huete, A., Didan, K., Miura, T., Rodriguez, E. P., Gao, X., Ferreira, L. G. (2002).
Overview of the radiometric and biophysical performance of the MODIS
vegetation indices. Remote Sensingof Environment, 83, 195-213.

IGBP (1990). The international Geosphere-Biosphere Programme: a study of
global change — the initial core projects. I_G_BP Repom, Stockholm, Sweden.

IPCC (2001). Climate change_2001: the scientific basis. Cambridge University
Press.

Jacob, D. and Podzun, R. (1997). Sensitivity studies with the regional climate
model REMO. Meteorologyang Atmospheric Physics, 63, 119-129.

Jansen, L. J. M. and Gregorio, A. D. (2003). Land-use data collection using the

“land cover classification system”: results from a case study in Kenya. Land Use
Policy, 20, 131-148.

118

Jin, M., Dickinson, R. E., Vogelmann, A. M. (1997). A comparison of CCM2-
BATS skin temperature and surface-air temperature with satellite and surface
observations. Jouraai of Climag, 10, 1505-1524.

Justice, C. O., Townshend, J. R. G., Vermote, E. F., Masuoka, E., Wolfe, R. E.,
Saleous, M, Roy, D. P., Morisette, J. T. (2002). An overview of MODIS Land
data processing and product status. Remote Sensing of Environment, 83, 3-15.

Kain, J. S. and F ristch, J. M. (1993). Convective parameterization for mesoscale
models: the Kain-Fristch scheme, in The Representation of Cumulus Convection
in Niinericfial Models, Meteorological Monographs, 24, pp. 165-170, American
Meteorological Society, Boston, Mass.

Kerr, Y. H., Lagouarde, J. P., lmbemon, J. (1992). Accurate land surface
temperature retrieval from AVHRR data with use of an improved split-window
algorithm. Bamote sensing of Environment, 41, 197-209.

Kleidon, A. and Heimann, M. (2000). Assessing the role of deep rooted
vegetation in the climate system with model simulations: Mechanism, comparison
to observations and implications for Amazonian deforestation. Climate Dynamics,
16, 183-199.

Kummerow, C., Simpson, J., Thiele, O., Barnes, W., Chang, A. T. C., Stocker, E.,
Adler, R. F., Hou, A., Kakar, R., Wentz, F., Ashcroft, P., Kozu, T., Hong, Y.,
Okamoto, K., Iguchi, T., Kuroiwa, H., Im, E., Haddad, Z., Huffman, G., Ferrier, 8.,
Olson, W. S., Zipser, E., Smith, E. A., Wilheit, T. T., North, G., Krishnamurti, T.,
Nakamura, K. (2000). The status of the tropical rainfall measuring mission
(TRMM) after two yeas in orbit. Journal of Applied Meteorology, 39, 1965-1982.

Kuo, H. L. (1974). Further studies of the parameterization of the influence of
cumulus convection on large-scale flow. Mrnal of Atmospheric Sciences, 31,
1232-1240.

Lambin, E. F. et al. (1999). Land-use and land-cover change (LUCC):
implementation strategy. IGBP Report, 48, IGBP, Stockholm, Sweden.

Lark, R. M. (1995). Components of accuracy of maps with special reference to
discriminant analysis on remote sensing data. International Journal of Remote
Sensing, 16, 1461-1480.

Latifovic, R. and Olthof, I. (2004). Accuracy assessment using sub-pixel fractional
error matrices of global land cover products derived from satellite data. Remote
Sensing of Environment, 90, 153-165.

Laval, K. (1986). General circulation model experiments with surface albedo
changes. Climate Change, 9, 91-102.

119

Lean, J. and Rowntree, P. R. (1997). Understanding the sensitivity of a GCM
simulation of Amazonian deforestation to the specification of vegetation and soil
characteristics. Journal of Climate, 10, 1216-1235.

Lee, T. J. (1992). The impact of vegetation on the atmcgheric boundagy layer
app convective storms. Atmosphere Science Paper No. 509, Colorado State
University, CO, USA.

Lofgren, B. M. (1995). Surface albedo-climate feedback simulated using two-way
coupling. Mrnal of Climate, 8, 2543-2562.

Loveland, T. R., Reed, B. C., Brown, J. F., Ohlen, D. O., Zhu, Z., Yang, L.,
Meachant, J. W. (2000). Development of a global land cover characteristics
database and IGBP DlSCover from 1 km AVHRR data. International Journal of
Remote Sensing, 21 , 1303-1330.

Lu, L. and Shuttleworth, W. J. (2002). Incorporating NDVI-derived LAI into the
climate version of RAMS and its impact on regional climate. Journal of
flygrometeorology, 3, 347-362.

Lynch, A. H., Mcllwaine, S., Beringer, J., Bonan, G. B. (2001). An investigation of
the sensitivity of a land surface model to climate change using a reduced form
model. Climate Dynamics, 17, 643-652.

Mahlman, J. D. (1997). Uncertainties in projections of human-caused climate
warming. Science, 278, 1416-1417.

Mathews, E. (1983). Global vegetation and land use: new high resolution data
bases for climate studies. Journal of Climate and Applied Meteorology, 22, 474-
487.

Mayaux, P., Bartholome, E., Fritz, S., Belward, A. (2004). A new land-cover map
of Africa for the year 2000. Journal of Biogeography, 31 , 861-877.

McGuffie, M. and Henderson-Sellers, A. (2001). Forty yeas of numerical climate
modeling. lnternaticml JournaL of Climatolagy, 21, 1067-1109.

Mellor, G. L. and Yamada, T. (1982). Development of a turbulence closure model
for geophysical fluid problems. Review of Geophysics, 20, 851-875.

Molders, N., Strasser, U., Schneider, K., Mauser, W., Raabe, A. (1997). A
sensitivity study on the initialization of surface characteristics in meso-B / y
modeling using digitized vs. satellite derived land-use data. Contriputions to
Atmospheric Physics, 70, 173-187.

120

Myneni, R. B. and Williams, D. L. (1994). On the relationship between FPAR and
NDVI. Remote Sensinaof Environment, 49, 200-211.

Myneni, R. B., Hoffman, S., Knyazikhin, Y., Privette, J. L., Glassy, J., Tian, Y.,
Wang, Y., Song, X., Zhang, Y., Smith, G. R., Lotsch, A., Friedl, M., Morisette, J.
T., Votava, P., Nemani, R. R., Running, S. W. (2002). Global products of
vegetation leaf area and fraction absorbed PAR from year one of MODIS data.
Remote SensinlofaEnvironment, 83, 214-231.

Nemani, R. R., Running, S. W., Pielke, R. A., Chase, T. N. (1996). Global
vegetation cover changes from coarse resolution satellite data. Journal of
Geophysical Research, 101 , 7157-7162.

Nicholson, S. E., Tucker, C. J., Ba, M. B. (1998). Desertification, drought, and
surface vegetation: an example from the west African Sahel. Bulletin of the
American Meteorological Society, 79, 815-829.

Oleson, K. W. and Bonan, G. B. (2000). The effects of remotely sensed plant
functional type and leaf area index on simulations of boreal forest surface fluxes
by the NCAR land surface model. flrnal of Hydrometeorolgqy, 1, 431-446.

Olson, J. 8., Watts, J. A., Allison, L. J. (1983). Carbon in live vegetation of maior
world ecosystems. Oak Ridge, TN: Oak Ridge National Laboratory.

Olson, J. S. (1994a). Global ecosystem framework;_definitions. Internal report,
USGS EROS Data Center, Sioux Falls, SD.

 

Olson, J. S. (1994b). Gflml ecosystem framework-translation strategy. Internal
report, USGS EROS Data Center, Sioux Falls, SD.

Palmer, T. N. (2000). Predicting uncertainty in forecasts of weather and climate.
mods on Progress in Physics, 63, 71-116.

Pauwels, V. R. N. and Wood E. F. (2000). The importance of classification
differences and spatial resolution of land cover data in the uncertainty in model
results over boreal ecosystems. Journal of Hygrometeorology, 1, 255-266.

Pielke, R. A., et al. (1992). A comprehensive meteorological modeling system -
RAMS. Meteorology anal Atmosaheric Physics, 49, 69-91.

Pielke, R. A., Walko, R. L., Steyaert, L. T., Vidale, P. L., Liston, G. E., Lyons, W.
A., Chase, T. N. (1999). The influence of anthropogenic landscape changes on
weather in south Florida. Monthly Weather Review, 127, 1663-1673.

Pielke, R. A. (2001). Influence of the spatial distribution of vegetation and soils on
the prediction of cumulus convective rainfall. Review of Geophysics, 39, 151-177.

121

Pitman, A. J. (2003). The evolution of, and revolution in, land surface schemes
designed for climate models. International Journal of Climatology, 23, 479-510.

Price, J. C. (1984). Land surface temperature measurements from the split
window channels of the NOAA-7 AVHRR. flirnal of Geghysical Research, 79,
5039-5044.

Price, J. C. (1992). Estimating vegetation amount from visible and near-infrared
reflectances. Remote sensing of Environment, 41, 29-34.

Qin, Z. and Karnieli, A. (1999). Progress in the remote sensing of land surface
temperature and ground emissivity using NOAA-AVHRR data. lntemational
J_ogrnal of Remote Sensing, 20, 2367-2393.

Reynolds, R. W. and Smith, T. M. (1994). Improved global sea surface
temperature analyses using optimum interpolation. Journal of Climate, 7, 929-
948.

Richardson, L. F. (1922). Weather predictiorLby namerical procesa. Cambridge:
Cambridge University Press.

Running, S. W., Loveland, T. R., Pierce, L. L., Nemani, R. R., Hunt, E. R., Jr.
(1995). A remote sensing based vegetation classification logic for global land
cover analysis. Remote Sensing of Environment, 51, 39-48.

Running, 8., Baldocchi, D., Turner, D., Gower, S., Bakwin, P., Hibbard, K.
(1999). A global terrestrial monitoring network integrating tower fluxes, flask

sampling, ecosystem modeling and E08 data. Remote Sensing of Environment.
70, 108-127.

Salomonson, V. V., Barnes, W. L., Maymon, P. W., Montgomery, H. E., Ostrow,
H. (1989). MODIS: advanced facility instruments for studies of the earth as a
system. fiE Transactions on Geoscience aLd Remote Sensing, 27, 145-153.

Scepan, J. (1999). Thematic validation of high-resolution global land-cover data
sets. Photogmmetic Engineering and Remote Sensing, 65, 1051-1060.

Scheffe, H. (1959). The analysis of variance. New York: John Wiley & Sons.
Sellers, P. J., Mintz, Y., Sud, Y. C., Dalcher, A. (1986). A simple biosphere model

(SIB) for use within general circulation models. Journal of the Atmospheric
Sciences, 43, 505-531.

 

Sellers, P. J., Randall, D. A., Collatz, G. J., Berry, J. A., Field, C. B., Dazlich, D.
A., Zhang, C., Collelo, G. D., Bounoua, L. (1996a). A revised land surface

122

parameterization (SiBZ) for atmospheric GCMs. Part 1: model formulation.
Journal of Climate, 9, 676-705.

Sellers, P. J., Los, S. 0., Tucker, C. J., Justice, C. O., Dazlich, D. A., Collatz, G.
J., Randall, D. A. (1996b). A revised land surface parameterization (SiBZ) for
atmospheric GCMs. Part 2: the generation of global fields of terrestrial
biophysical parameters from satellite data. Journal of Climate, 9, 706-737.

Sellers, P. J., Dickinson, R. E., Randall, D. A., Betts, A. K., Hall, F. G., Berry, J.
A., Collatz, G. J., Denning, A. S., Mooney, H. A., Nobre, C. A., Sato, N., Field, C.
B., Henderson-Sellers, A. (1997). Modeling the exchanges of energy, water and
carbon between continents and atmosphere. Science, 275, 502-509.

Shabanov, N. V., Wang, Y., Buermann, W., Dong, J., Hoffman, S., Smith, G. R.,
Tian, Y., Hnyazikhin, Y., Myneni, R. B. (2003). Effect of foliage spatial
heterogeneity in the MODIS LAI and FPAR algorithm over broadleaf forests.
Remote Sensinflf Environment, 85, 410-423.

Simpson, J., Adler, R. F. and North, G. R. (1988). A proposed Tropical Rainfall
Measuring Mission (TRMM) satellite. Bulletin of the American Meteorological
Sociey, 69, 278-295.

Smagorinsky, J. (1963). General circulation experiments with the primitive
equations. Part I: the basic experiment. Monthly Weather Review, 91, 99-164.

Stapleton, J. H. (1995). Linear statistical models. New York: John Wiley & Sons.

Sud, Y. C. and Fennessy, M. J. (1982). A study of the influence of surface albedo
on July circulation in semi-arid regions using the GLAS GCM. Journal of
Climatology, 22, 105-125.

Sun, J. and Mahrt, L. (1995). Determination of surface fluxes from the surface
radiation temperature. Journal of Atmospheric Sciences, 52, 1096-1106.

Taylor, C., Lambin, E. F., Stephenne, N., Harding, R., Essery, R. (2002). The
influence of land-use change on climate in the Sahel. Journal of Climate, 15,
3615-3629.

 

Thomlinson, J. R., Bolstad, P. V., Cohen, W. B. (1999). Coordinating
methodologies for scaling landcover classifications from site-specific to global:
steps toward validating global map products. Remote Sensing of Environment.
70, 16-28.

Tian, Y., Woodcock, C. E., Wang, Y., Privette, J. L., Shabanow, N. V., Zhou, L.,

Zhang, Y., Buermann, W., Dong, J., Veikkanen, B., Hame, T., anersson, K.,
Ozdogan, M., Knyazikhin, Y., Myeni, R. B. (2002). Multiscale analysis and

123

validation of the MODIS LAI product I. Uncertainty assessment. Remote Sensing
of Environment, 83, 414-430.

Torbick, N., Lusch, D., Qi, J., Moore, M, Olson, J. and Ge, J. (2006). Developing
land use/land cover parameterization for climate-land modeling in East Africa.
International Jamal of Remote Sensing, 27, 4227-4244.

Tremback, C. J. and Kessler, R. (1985). A surface temperatgre and moistgya
parameterization for use in mesoscale models. Seventh Conference on
Numerical Weather Prediction, Montreal, Canada, American Meteorological
Society, 355-358.

Trodd, N. M. (1995)._l_Jncertainty in land cover mappi_ng for mgdalinfland cover
change. Proceedings of R8895 Remote Sensing in Action, Remote Sensing
Society, Nottingham, UK.

Valor, E. and Caselles, V. (1996). Mapping land surface emissivity from NDVI:
Application to European, African, and South American areas. Remote Sensing_of
Environment, 57, 167-184.

Verbyla, D. L. (2005). Assessment of the MODIS leaf area index product
(MOD15) in Alaska. lntemational Journal of Remote Sensing, 26, 1277-1284.

Vidale, P. L., Luthi, D., Frei, C., Seneviratne, S. I., Schar, C. (2003). Predictability
and uncertainty in a regional climate model. Journal of Geophysical Resea_rc_:g,
108, doi:10.1029/2002JD0028 10.

Walko, R. L., Band, L. E., Baron, J., Kittel, T. G. E., Lammers, R., Lee, T. J.,
Ojima, D., Pielke, R. A., Taylor, C., Tague, C., Tremback, C. J., Vidale, P. L.
(2000). Coupled atmosphere-biosphere-hydrology models for environmental
modeling. Mrnal of Applied Meteorology, 39, 931-944.

Wan, Z. and Li, Z.-L. (1997). A physics-based algorithm for retrieving land-
surface emissivity and temperature from EOS/MODIS data. IEEE Transactions
on Geoscience all Remote Sensiaq, 35, 980-996.

Wan, Z., Zhang, Y., Zhang, 0., Li, Z. (2002). Validation of the land-surface
temperature products retrieved from Terra Moderate Resolution Imaging
Spectroradiometer data. Remote Sensing ofEnvironment, 83, 163-180.

Wan, Z., Zhang, Y., Zhang, Q. (2004). Quality assessment and validation of the
MODIS global land surface temperature. International Journal of Remote
Sensing, 25, 261-274.

Wang, J. R. (1992). Active and passive microwave measurements of soil
moisture in FIFE. =lflrnal of Geophysical Researc_h, 97, 18979-18985.

124

Wang, G., Eltahir, E. A. B., Foley, J. A., Pollard, D., Levis, S. (2004). Decadal
variability of rainfall in the Sahel: results from the coupled GENESIS-IBIS
atmosphere-biosphere model. ClimfiDmamics, 22, 625-637.

Wang, Y., Woodcock, C. E., Buermann, W., Stenberg, P., Voipio, P., Smolander,
H., Hame, T., Tian, Y., Hu, J., Knyazikhin, Y., Myneni, R. B. (2004). Evaluation of
the MODIS LAI algorithm at a coniferous forest site in Finland. Remote Sensing
prnvironment, 91, 114-127.

Washington, W. M. and Parkinson, C. L. (2005). An introdfiuction to three-
gimensional climate modeling (second edition). University Science Books: Mill
Valley, Ca.

Watson, R. T., Noble, l. R., Bolin, B., Bavindranath, N. H., Verardo, D. J.,
Dokken, D. J. (2000). Land use, land use change, and forestry. IPCC Special
Report, Intemvernmental Panel on Climate Chapga, Genewa, Switzerland.

Weaver, C. P. and Avissar, R. (2001). Atmospheric disturbances caused by
human modification of the landscape. Bulletin of the American Meteorological
Socieg, 82, 269-281.

Weaver, C. P., Roy, S. B., Avissar, R. (2002). Sensitivity of simulated mesoscale
atmospheric circulations resulting from heterogeneity to aspects of model
configuration. JAJI‘I‘IaI of Geophysical Research, 107(D20), 8041,
doi:10.1029/2001JD000376.

Wittich, K.-P. (1997). Some simple relationships between land-surface emissivity,
greenness and the plant cover fraction for use in satellite remote sensing.
International Journal of Biometeorology, 41, 58-64.

Xue, Y. (1997). Biosphere feedback on regional climate in tropical North Africa.
_Ql_.@l‘t€l‘|V Joaurnal of the Royal Meteorolofial Society, 123, 1483-1515.

Yang, Z-L. (2003). Modeling land surface processes in short-term weather and
climate studies. Theogy and Modeling of Atmospheric Variapility, edited by Zu, X.,
Li, X., Cai, M., Zhou, S., Zhu, Y., Jin, F.-F., Zou, X., Zhang, M., World Scientific
Series on Meteorology of East Asia, World Scientific, New Jersey, 288-313.

Zeng, X., Dai, J., Dickinson, R. E., Shaikh, M. (1998). The role of root distribution
for climate simulation over land. Geophysical Research Letter, 25, 4533-4536.

Zeng, X., Dickinson, R. E., Walker, A., Shaikh, M., DeFries, R. S., Qi, J. (2000).
Derivation and evaluation of global 1-km fractional vegetation cover data for land

modeling. Journal of Applied Meteorology, 39, 826-839.

125

Zhao, M., Pitman, A. J., Chase, T. N. (2001a). Climatic effects of land cover
change at different carbon dioxide levels. Climate Research, 17, 1-18.

Zhao, M., Pitman, A. J., Chase, T. N. (2001b). The impact of land cover change
on the atmospheric circulation. Qimate Dynamics, 17, 467-477.

Zhuang, X., Engel, B. A., Xiong, X., Johannsen, C. J. (1995). Analysis of

classification results of remotely sensed data and evaluation of classification
algorithms. Photpqrammetric Engineeriryg and Remote Sensipg, 61, 427-433.

126

 

IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII

IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIILIIIIIIII