Predicting Lake Surface Water Temperature With Transfer‐Based Physics‐Informed Deep Learning

Ongoing climate change has intensified lake surface warming. Enhanced accuracy in lake temperature modeling can better assess the risk of ecosystems being harmed by thermal tipping points, informing more sustainable water management. Recent progress in physics‐informed deep learning (PIDL) has opened new avenues for improving such modeling. Yet, the site‐specific focus of PIDL training continues to pose challenges to its broader applicability. Here, we demonstrate Transfer‐PIDL, a transfer learning framework that enhances PIDL generalizability for lake surface temperature prediction. This approach employs a three‐stage training strategy, consisting of pre‐training based on large‐scale satellite observations, re‐training using process‐based (PB) model simulations, and fine‐tuning with local measurements. Our first experiment revealed that, given sufficient pre‐training (e.g., on 40 source lakes), Transfer‐PIDL outperformed local PIDL by 20%–39% in validation root‐mean‐square‐error (RMSE). Additionally, to achieve comparable performance, Transfer‐PIDL showed a reduced fine‐tuning data requirement compared with both local PIDL and purely data‐driven deep learning (DL) across three initial cases. Transfer‐PIDL further demonstrated consistent accuracy across 43 additional lakes with in situ temperature observations (mean validation RMSE of 1.2°C), surpassing local PIDL (1.6°C), DL (1.8°C), and PB (1.9°C) models. In a global‐scale experiment involving 869 lakes, Transfer‐PIDL exhibited cross‐thermal‐system transferability, with the poorest‐performing validation scenario still achieving a RMSE of 1.5 ± 0.36°C, mean‐absolute‐error of 1.1 ± 0.25°C, and R2 of 0.85 ± 0.12 (mean ± SD). This study demonstrates the synergy between transfer learning and PIDL, offering a promising approach for large‐scale lake temperature modeling.