MULTISCALE DEEP-LEARNING MODELING METHODS FOR PREDICTING STREAM TEMPERATURE AT LOCAL SPATIAL RESOLUTIONS IN WELL-OBSERVED AND DATA-SPARSE BASINS

Stream temperature is governed by processes that operate on multiple scales, resulting in temperatures that are spatially and temporally heterogeneous. National or regional models accurately capture broad patterns driven by climate, topography, and seasonality, but lack the resolution to simulate local-scale variation in groundwater discharge, riparian shading, or tributary inflows, and therefore to predict the thermal refugia that are critical to habitat management. In contrast, local models may simulate finer-resolution processes, but are limited in spatial extent and may miss drivers of broad patterns. Computational approaches exist for model downscaling, including deep learning-based super resolution, transferring models from coarse to fine resolutions, and predicting the residuals between coarse- and fine-resolution models. However, these methods were not designed for capturing complex spatial and temporal processes governing stream temperature, and thus the relative accuracy and efficiency of these approaches when applied to stream temperature models are uncertain.

Our objectives are 1) to produce fine-resolution daily stream temperature predictions by leveraging models trained at different scales and 2) to assess both standard and innovative approaches to multiscale modeling. The coarse-resolution model is a deep learning model that simulates daily stream temperature on the National Geospatial Fabric (median reach length of 10.5 km ; 55,100 km total length in the northeastern US). The fine-resolution model simulates daily stream temperature on the National Hydrography Dataset (median reach length of 1.1 km; 292,000 km total in the same area). The architecture of the fine-resolution model varies across scaling approaches. We are analyzing multiple ecologically-relevant metrics of predictive accuracy (e.g. July root-mean-square error [RMSE], days over a threshold) and feature importance across reaches of differing thermal regimes. We expect this analysis will improve our multiscale modeling tools that can guide management of thermal-refugia habitat.