Machine learning water quality inversion via optimal band selection from highly correlated subsets

Rapid urbanization has triggered severe water contamination in urban fluvial ecosystems, making efficient water quality monitoring and precise retrieval an urgent research priority. Taking Shanghai's Lianqi River as the study area, this paper targets COD and ammonia nitrogen with distinct concentration ranges. On the basis of UAV multispectral remote sensing observations, this work explores water quality retrieval via diverse feature screening strategies and machine learning approaches. Exhaustive search is adopted to generate abundant feature combinations, and six algorithms including CatBoost, XGBoost, LightGBM, KNN, RF, and MBLERM are employed to build massive optimized model clusters. A multi-dimensional feature evaluation system is constructed to assess feature properties from correlation, distance, and spectral morphology perspectives. Results show that linear correlation cannot determine model precision; model accuracy declines beyond seven features due to dimensional catastrophe; pollutant spatial distribution presents obvious heterogeneity closely associated with land-use patterns.