Predicting Water Temperature Dynamics of Unmonitored Lake Systems with Meta Transfer Learning

Abstract:

Though sensor-based monitoring and machine learning (ML) applications have recently seen rapid growth in the environmental sciences, the majority of freshwater lakes remain unmonitored and thus have been inaccessible for ML models to predict water temperature. We demonstrate a Meta Transfer Learning (MTL) framework that uses ML to allow the systematic selection of data-driven models from well-observed lakes to predict unmonitored lakes. To do this selection, we construct a meta-model that predicts how well a model will transfer to another lake based on lake attributes. We show that both calibrated process-based (PB) models and the recently developed process-guided deep learning (PGDL) models can be transferred to make predictions of water temperature dynamics of these systems, outperforming alternative methods. MTL performance was evaluated using 145 PB models (PB-MTL) and 145 PGDL models (PGDL-MTL) built from well-observed lakes and transferred to 305 lakes treated as unobserved in Minnesota and Wisconsin. We show a significant increase in performance (decrease in RMSE), compared to the uncalibrated process-based General Lake Model, with PB-MTL, PGDL-MTL, and an ensemble approach combining several models from PGDL-MTL. We further evaluate PGDL-MTL against a leading PGDL approach for sparsely observed lake systems and find that PGDL-MTL can often outperform these models trained on data from the target lakes themselves, with a threshold of roughly 30-45 days of observations required before a locally trained PGDL outperforms the ensembled PGDL-MTL. Lastly, we show that this approach can scale to thousands of lakes in the Midwestern United States. This work demonstrates that integrating scientific knowledge into both the candidate source models for well-observed systems (the PGDL approach) as well as the meta learning model deciding which models to transfer (MTL) shows promise for predicting many different kinds of unmonitored systems.