Estimación de parámetros para la evaluación del índice de calidad de agua en el lago George mediante el uso de imágenes sentinel-2 y machine learning

Abstract

To implement an alternative methodology for water quality monitoring in Lake George, linear regression models were developed to estimate water quality parameters using field measurements collected between September 2022 and April 2024 as dependent variables and a combination of Sentinel 2 spectral bands, indices, and band ratios as predictors. Of the datasets analyzed, ten achieved satisfactory performance, with an adjusted R2 greater than 70%. However, due to the strong temporal autocorrelation detected in the residuals, generalized least squares with autoregressive errors (GLSAR) was employed instead of ordinary least squares, resulting in statistically valid models. The estimated parameters were subsequently used to calculate the Canadian Council of Ministers of the Environment Water Quality Index. A multivariate adaptive regression splines (MARS) model was then fitted to relate the calculated index to Sentinel-2 spectral bands from four representative images corresponding to the different seasons of the year. The July image provided the best predictive performance, with an RMSE of 2.76 and a mean absolute error (MAE) of 1.99. Nevertheless, the best model explained only 0.9% of the variability in the water quality index because most WQI values were concentrated between 80 and 85, indicating limited variability in the response variable and, consequently, poor predictive performance of the MARS model