Predicting Water Temperature Dynamics of Unmonitored Lakes with Meta Transfer Learning: 5 Model predictions

Multiple modeling frameworks were used to predict daily temperatures at 0.5m depth intervals for a set of diverse lakes in the U.S. states of South Dakota, North Dakota, Minnesota, Wisconsin, and Michigan. Process-Based (PB) models were configured and calibrated with training data to reduce root-mean squared error. Uncalibrated models used default configurations (PB0; see Winslow et al. 2016 for details) and no parameters were adjusted according to model fit with observations. Process-Guided Deep Learning (PGDL) models were deep learning models with an added physical constraint for energy conservation as a loss term. These models were pre-trained with uncalibrated Process-Based model outputs (PB0) before training on actual temperature observations. PBall and PGDL models were transfered to other lakes with a meta transfer learning model (PB-MTL; PGDL-MTL) and for PGDL transfers, an additional ensemble estimate was created based on the top nine source lakes (PGDL-MTL9).