Advancing reservoirs water quality parameters estimation using Sentinel-2 and Landsat-8 satellite data with machine learning approaches

Abstract

Reservoir eutrophication, caused by human activities and climate change, has emerged as a critical environmental concern that has attracted both governmental and public attention. However, accurate measurement of water quality parameters, such as chlorophyll a (CHL-a), water clarity (Secchi depth; SD), and total suspended solids (TSS), in inland waters is challenging due to the optical complexity of individual water bodies, which impedes the optimization of conventional bio-optical algorithms. The aim of this study was to demonstrate the viability of harmonizing Sentinel-2 Multi-Spectral Imager (MSI) and Landsat-8 Operational Land Imager (OLI) satellite imagery surface reflectance (SR) products to facilitate the monitoring of inland reservoir CHL-a, SD, and TSS using the Google Earth Engine (GEE) platform and machine learning algorithms. Machine learning models were trained using Landsat-8 OLI and Sentinel-2 MSI surface reflectance products to identify bands and combinations predicting CHL-a, SD, and TSS. Among the machine learning algorithms tested, random forest (RF) (S-2 MSI: R2 = 0.61, mean absolute error [MAE] = 6.56%, root-mean-square error [RMSE] = 12.51 μg/L, L-8 OLI: R2 = 0.56, MAE = 8.44%, RMSE = 16.01 μg/L) yielded the best results in the test set for CHL-a prediction from the Sentinel-2 MSI and Landsat-8 OLI, outperforming the k-nearest neighbor (KNN), AdaBoost, and artificial neural network (ANN) models. It also showed superior performance for SD and TSS prediction. The feature importance analysis revealed that specific band ratios, such as (red/red edge1)*red edge2 for Sentinel-2 MSI and red/blue for Landsat-8 OLI, were significant predictors of CHL-a and TSS. The red/blue band ratio and the green band were highly predictive for SD in Sentinel-2 MSI and Landsat-8 OLI, respectively. The results of fall CHL-a predictions revealed varying trophic levels in the reservoirs. Landsat-8 OLI indicated that 2% of the reservoirs were oligotrophic, while 46%, 43%, and 9% were mesotrophic, eutrophic, and hypertrophic, respectively. Meanwhile, the Sentinel-2 MSI results showed that 51% of the reservoirs were mesotrophic, while 6%, 35%, and 8% were oligotrophic, eutrophic, and hypertrophic, respectively. Overall, this study demonstrates the effectiveness of Landsat-8 OLI and Sentinel-2 MSI surface reflectance products in estimating and monitoring reservoir water quality parameters using machine learning algorithms. This approach has the potential to yield valuable insights aiding the assessment and management of water quality at the regional, national, and global levels.