Mapping nutrient pollution in inland water bodies using multi-platform hyperspectral imagery and deep regression network.

Inland waters face multiple threats from human activities and natural factors, leading to frequent water quality issues, particularly the significant challenge of eutrophication. Hyperspectral remote sensing provides rich spectral information, enabling timely and accurate assessment of water quality status and trends. To address the challenge of inaccurate water quality mapping, we propose a novel deep learning framework for multi-parameter estimation from hyperspectral imagery. A deep convolutional spatial-spectral joint learning method incorporating high-dimensional attention-weighted differences is proposed to optimize the deep features. The model was used to accurately estimate the distribution of three key eutrophication-related water quality parameters: total nitrogen, total phosphorus, and ammonia nitrogen. Through scale analysis, ablation experiments, and model comparisons, the results demonstrate stable regression performance with the proposed model. Specifically, the coefficient of determination (R2) values are 0.8315, 0.8137, and 0.8245, the mean absolute error (MAE) values are 0.2035, 0.0056 and 0.0134, and the mean squared error (MSE) values are 0.0733, 0.00008 and 0.0003 for the three parameters in the test set, respectively. Compared to the traditional feature analysis and regression methods, the R² values are improved by approximately 30 %, while the MAE and MSE values are reduced by approximately 60 % and 80 %, respectively. The model was applied to airborne hyperspectral imagery for nutrient pollution mapping. To assess the model's generalizability, we applied the trained model to multi-temporal satellite hyperspectral imagery and validated against in situ monitoring data, where the proposed model demonstrated promising cross-platform and temporal transferability.