Prediction of Turbidity and TDS in Dam Reservoir from Multispectral UAV-Drone and Sentinel-2 Image Sensors Using Machine Learning Models

Abstract

This study presents results on the utility of DJI P4 Multispectral (DJI-PH4) UAV-Drone and Sentinel-2 MSI (S2-MSI) satellite datasets for the retrieval of Turbidity and Total Dissolved Solids (TDS) using empirical linear regression (ELR), XGBoost (eXtreme Gradient Boosting) and Random Forest Regression (RFR) machine learning (ML) models. For the case study of Gaborone dam in Botswana, 21 water sampling points were correlated with the corresponding spectral reflectances from DJI-PH4 and S2-MSI imagery. For the estimation of Turbidity, XGBoost gave the best prediction results with average training accuracy of R2 = NSE = 0.999, MAE=0.001 NTU, RMSE = 0.001 NTU and PBIAS = 0.1% for both the DJI-PH4 and S2-MSI sensors. XGBoost performed better than ELR and RFR at the model training phases, however its prediction of Turbidity in testing was lower than ELR but nearly same as RFR. In predicting TDS from both sensors, XGBoost had the highest performance with equivalent accuracy measures as for the prediction of Turbidity. Both the training and testing results for the estimation of TDS is accurate from the sensors, with ELR marginally outperforming the XGBoost and RFR in the testing phase with R2 = 0.998, MAE=0.338 mg/L, RMSE = 0.435 mg/L and NSE = 0.858. For the prediction of Turbidity, all the ML models gave good training results from the drone and Sentinel-2 data except for RFR in the case of Sentinel-2. The introduction of ensemble ELR-XGBoost model significantly improved the prediction of the water quality parameters from the drone and Sentinel-2 datasets. With the potential of providing high-frequency and large spatial coverage observational data in the near-real-time mode, the results of this study demonstrate the applicability of UAV-drone for the retrieval of Turbidity and TDS physical water quality parameter in dam reservoirs.