Prior-Informed Spatially Adaptive Network (PISA-Net) for multi-parameter water quality retrieval in Poyang Lake using Sentinel-2 imagery

Abstract

Accurate retrieval of water-quality parameters from remote-sensing imagery is essential for monitoring large inland lakes, yet pronounced spatial heterogeneity and limited transferability across varying environmental conditions remain major challenges. This study proposes a Prior-Informed Spatially Adaptive Network (PISA-Net) for multi-parameter water-quality retrieval from Sentinel-2 imagery in Poyang Lake. PISA-Net integrates a CNN branch for nonlinear spectral–spatial feature learning with a knowledge-guided branch based on empirical optical descriptors and coordinate-derived geographic priors. A geographic soft-assignment mechanism constructs the spatial prior vector, which conditions CNN features through Feature-wise Linear Modulation (FiLM), while a stop gradient consistency loss (SG-CL) promotes agreement between the two prediction pathways with controlled gradient propagation. Chl-a, TP, and CODMn were modeled independently and evaluated using 5-fold spatial-block cross-validation against six representative empirical, statistical, machine-learning, and deep-learning baselines. PISA-Net achieved R 2 values of 0.6113, 0.5112, and 0.5054 for Chl-a, TP, and CODMn, respectively, consistently outperforming the strongest baseline, 2D-CNN, while reducing RMSE by approximately 29.6%, 21.4%, and 19.0%. An additional April 2018 cross-season experiment revealed substantial degradation under direct zero-shot transfer, whereas few-shot adaptation markedly reduced prediction errors, indicating useful cross-season adaptability after limited local calibration. Overall, PISA-Net provides a prior-informed spectral–spatial framework for water-quality retrieval under spatially heterogeneous conditions, while broader multi-season validation remains necessary to establish comprehensive seasonal generalizability.