Evaluation of machine learning methods for forecasting turbidity in river networks using Sentinel-2 remote sensing data

Abstract

Turbidity is an important indicator of river water quality and of great interest to improve the data acquisition methods and efficiency of decision support systems for sustainable ecosystem management. However, river water quality monitoring stations are very expensive to operate and maintain and lack spatial coverage. Therefore, this study takes advantage of the vast spatial coverage of remote sensing datasets from satellites to provide a more efficient hybrid system with comprehensive coverage of both spatial and temporal changes in water quality across a vast river network. Spectral bands from Sentinel-2 were analyzed using machine learning algorithms, namely XGBoost, Random Forests, GMDH, Support Vector Regression, k-Nearest Neighbors and Least Absolute Shrinkage and Selection Operator to model turbidity, using data from twelve monitoring stations across the Mississippi River, USA. Results show that considering the individual monitoring stations, the ML algorithms for turbidity modeling were satisfactory at locations with a larger range and standard deviation of turbidity values, achieving a mean R² value of 59.5 %. Tree-based models were the best overall approach, often ranking as the best or second-best performing model. Using all the samples from the monitoring stations, the XGBoost provided a superior output for turbidity modeling, reaching R² equal to 75.7 %. This represents an improvement of over 16 % compared to the average metric value for the individual stations. A comprehensive comparison with the literature found that the models implemented using this study's methodology could provide competitive results, deeming it as an alternative for turbidity modeling from remote sensing data.