Cross-sensor spectral fusion for coastal water physicochemical assessment using optimized ensemble machine learning approaches.

Rising coastal pollution has increased use of remote sensing-machine learning monitoring frameworks. However, most studies rely on limited spectral features and conventional models, limiting their ability to capture the complex optical variability of coastal waters. Addressing these gaps, this study enhances random forest (RF) modeling of China's coastal water quality index (WQI) by incorporating teaching-learning-based optimization (TLBO) and student-psychology-based optimization (SPBO) using various spectral features. WQI was calculated using chemical oxygen demand, dissolved inorganic nitrogen, dissolved inorganic phosphorus, dissolved oxygen, total petroleum hydrocarbons, and potential hydrogen, while Landsat 8 and Sentinel 2 data served as inputs. Over 44% of WQI observations were rated poor, very poor, or severely polluted, highlighting severe pollution in Shanghai, the Yangtze River, Guangzhou, Hong Kong, and the Pearl River. While spectral indices correlated more with WQI than individual bands, the model was more sensitive to a satellite's full feature set than to a single spectral type, indicating that multisource satellite information is crucial for improving accuracy. RF-TLBO slightly outperformed RF-SPBO, achieving lower errors (mean absolute error: 0.141 vs. 0.143; mean squared error: 0.034 vs. 0.035), higher Willmott index (0.786 vs. 0.785) and coefficient of determination (0.462 vs. 0.459), while both models showed close to zero mean bias error. It also demonstrated lower uncertainty, with a 95% prediction interval coverage probability of 0.950 and a prediction interval width of 0.193. The proposed framework enhances coastal water quality monitoring and high-risk area detection. Future research could apply multi-target regression to better capture interdependencies among water quality parameters.