Estimation of dissolved oxygen in reservoirs from UAV-drone, Sentinel-2 and Landsat-9 multisensor data using stacked-ensemble machine learning: review and case study experimental results

Abstract

Accurate prediction of dissolved oxygen (DO) in reservoir systems is essential for assessing aquatic ecosystem health and informing water management decisions. This study develops and validates a stacked ensemble machine learning (SE-ML) modeling framework for estimating DO concentrations in a dam water reservoir by comparing multispectral reflectance data from UAV-drone, Sentinel-2 MSI, and Landsat-9 OLI-2 sensors. To overcome overfitting and weak generalizability in individual ML models, the study compared the prediction performance of Random Forest Regression (RFR), XGBoost, AdaBoost and CATBoost, and integrated the best performing MLs through a two-layer stacked ensemble architecture for optimal prediction of DO levels. The input parameters comprised of 84 spectral indices and water quality parameters (temperature, turbidity, total dissolved solids, pH and electric conductivity, and the validation comprised of train-test split, LOOCV, k-fold cross-validation and repeated k-fold cross-validation. For the SE-ML model, XGBoost and CATBoost were respectively determined as the most accurate base and meta models for the estimation of DO levels. Using the XGBoost-CATBoost SE-ML model with k-fold cross-validation, Landsat-9 exhibited the best outcome with DO prediction accuracies of R2 = 0.908, NSE = 0.881, RMSE = 0.126 mg/L and MAE = 0.098 mg/L, as compared to Sentinel-2 (R2 = 0.895, NSE = 0.842, RMSE = 0.137 mg/L and MAE = 0.111 mg/L), and UAV-drone (R2 = 0.864, NSE = 0.807, RMSE = 0.166 mg/L and MAE = 0.129 mg/L). With marginal accuracy differences compared to Landsat-9 and Sentinel-2 satellite sensors, the experimental results demonstrated the potential of UAV multispectral data for rapid, continuous and cost-effective approach towards monitoring water quality in dam water reservoirs.