Assessment of lake water quality status based on Sentinel-2 images, location information, and machine learning algorithms.

Assessing lake water quality is essential for ecological protection, public health, resource management, and policymaking. Traditional approaches assess water quality at specific sampling points or localized areas, failing to represent the lake's overall condition. To improve this, the study combines measured water quality parameters with the entropy weight method to create the entropy-weighted water quality index (EWQI). It also integrates Sentinel-2 imagery and accounts for spatial variability in water quality assessment. The contribution of each feature to EWQI is analyzed using an interpretable model (SHAP). The study compares the effectiveness of EWQI and the comprehensive trophic level index (TLI) in evaluating lake water quality. Results show that (1) machine learning models, particularly XGBoost, outperform multiple linear regression (MLR), with a training R2 of 0.98, RMSE of 10.30, MAE of 7.93, and RPD of 7.27. The testing results include an R2 of 0.90, an RMSE of 23.53, an MAE of 18.10, and an RPD of 3.17. (2) Adding location data significantly improves model accuracy, increasing R2 by up to 0.1, reducing RMSE by 14.55 and MAE by 10.83, and increasing RPD by 4.32. (3) Location information is a key factor in estimating EWQI. (4) For Hongjiannao Lake, the EWQI mostly ranges from 110 to 150. Only 0.25% of water bodies are classified as good, 62.18% as medium, 19.64% as poor, and 17.93% as extremely poor. Water quality tends to decline from the lake edge toward the center, with poorer conditions in the middle and small areas of good quality in the southeast and southwest. Compared with TLI, the EWQI more accurately reflects the water quality of Hongjiannao Lake.