Predicting Water Temperature Dynamics of Unmonitored Lakes With Meta?Transfer Learning

Most environmental data come from a minority of well-monitored sites. An ongoing challenge in the sciences is transferring knowledge monitored sites to unmonitored Here, we demonstrate novel transfer-learning framework that accurately predicts depth-specific temperature lakes (targets) by borrowing models (sources). This method, meta-transfer learning (MTL), builds meta-learning model predict transfer performance candidate source targets using lake attributes and candidates' past performance. We constructed at 145 calibrated process-based (PB) modeling recently developed approach called process-guided deep (PGDL). applied MTL either PB or PGDL (PB-MTL PGDL-MTL, respectively) temperatures 305 target treated as Upper Midwestern United States. show significantly improved relative uncalibrated General Lake Model, where median root mean squared error (RMSE) for 2.52°C. PB-MTL yielded RMSE 2.43°C; PGDL-MTL 2.16°C; ensemble nine sources per 1.88°C. For sparsely lakes, often outperformed trained on themselves. Differences maximum depth between were consistently most important predictors. Our readily scales thousands States, demonstrating with meaningful predictor variables high-quality promising many kinds systems variables.