N-PCLN: A physics-informed deep learning framework for short-term chlorophyll-a prediction with sparse satellite observation

Chlorophyll-a (Chla) is a key indicator of lake eutrophication. Its short-term prediction is essential for water quality monitoring and bloom early warning. However, traditional deep learning models often rely heavily on dense observations and provide limited process interpretability. To address these limitations, this study proposes a physics-informed deep learning framework, named N-PCLN, for short-term Chla prediction in Erhai Lake. The framework integrates Sentinel-3 OLCI imagery and meteorological datasets from 2016 to 2024. It uses Chla, land surface temperature, wind speed, and precipitation as inputs. Total phosphorus is incorporated through a nutrient–phytoplankton model to construct a physical loss. This design enables the model to jointly learn spatiotemporal patterns and nutrient–phytoplankton process constraints. The N-PCLN outperforms ConvLSTM, Transformer, MDN, CNN, and LSTM models, achieving an R 2 of 0.767, an MAE of 1.483 mg/m 3 , and an RMSE of 2.665 mg/m 3 . Sparse-data experiments show that N-PCLN maintains more stable performance than the selected baseline models under reduced data availability. The ablation experiments indicate that LST is the dominant meteorological factor driver of short-term Chla dynamics, while wind speed and precipitation mainly affect hydrodynamic redistribution and runoff-related nutrient transport. Physical-loss gradient analysis further reveals that the N-P constraint acts differently across seasons and Chla concentration. These results demonstrate that N-PCLN improves both prediction performance and process interpretability. It provides a useful framework for satellite-based lake water quality prediction.