Disentangling Spectral and Environmental Controls on Inland River–Lake Water Quality Using Satellite Earth Observation-Driven Optimized Machine Learning

Abstract

Accurate monitoring of surface water quality remains challenging due to pronounced spatial heterogeneity and limited ground observations. Satellite remote sensing offers scalable solutions, yet the extent to which environmental drivers enhance predictive performance, particularly in complex river–lake systems, remains insufficiently understood. This study develops a parallel comparative river–lake modeling framework to estimate permanganate index (CODmn), total phosphorus (TP), and total nitrogen (TN) by integrating satellite spectral data with climatic and land-use variables. Four model configurations were evaluated: spectral predictors alone (M1), spectral predictors combined with climate variables (M2), spectral predictors combined with land-use information (M3), and full integration of all predictors (M4). LightGBM models were optimized using Bayesian hyperparameter tuning (Optuna) and trained over rivers (January 2021–July 2025) and lakes (January 2021–December 2024) datasets in the Dongliao Basin, China. Spectral predictors alone (M1) provided robust performance for CODmn (R2 = 0.78), with marginal improvement when land-use variables were included (M2) in rivers (R2 = 0.80). In contrast, nutrient predictions showed stronger dependence on environmental covariates. TN predictions improved substantially with land-use inputs (M3) (R2 = 0.75 in rivers and 0.63 in lakes with M2), with further gains with full integration (M4) in lakes (R2 = 0.66). TP predictions exhibited marked improvements with land-use variables in rivers (R2 = 0.76) and with full integration in lakes (R2 = 0.72). Model interpretability analysis using SHAP revealed that spectral features dominate CODmn estimation, while climatic and watershed characteristics exert greater influence on TN variability. Seasonal analysis indicated that hydrological drivers dominate during wet seasons, while land-use effects and internal biogeochemical processes become more important in dry seasons. The proposed framework advances predictive accuracy and process understanding, supporting more effective monitoring and management of water quality in complex river–lake systems.