Stream temperature is an important parameter of water quality and developing models capable of reliable predictions are critically important in stream management. In addition to the structure of these models (e.g., predictive variables), there are other factors that may influence model performance such as the selection of data granularity (i.e., level of temporal aggregation) and seasonal coverage of data collection. Data granularity and seasonal extent of data collection vary widely in the literature and have often been arbitrarily selected in stream temperature modeling studies in the past, but the consequences of these choices have not been explored. I applied different data granularity and time period selections to regression models, which were developed by Andrews (2019) to predict temperature gradient (i.e., stream temperature change) in Michigan’s streams. Applying higher data granularity increased overall model performances and changed model selection results, however applying different time periods did not have a substantial effect on model performances. Using higher data granularity also changed model parameter estimates by increasing the multicollinearity in best-fitting models. In addition to temporal data granularity, data may be pooled spatially across streams within a thermal class to reduce the costs of data collection. I examined the impact of stream classification on model performance by applying data pooling within stream classes. Stream-Specific Models had better performance compared to Class-Based Models. Additional analyses suggested that classifying streams based on temperature gradient instead of stream temperature may result in better Class-Based Model performance.
Read
