Enhancing suspended sediment concentration retrieval by integrating thermal infrared and optical bands of Landsat-8 and machine learning algorithms

Suspended Sediment Concentration (SSC) variabilities sizably affect the bio-optical conditions in water bodies. Remote sensing plays a key role in spatiotemporal monitoring of the SSC in inland and coastal waters. However, existing studies on remote sensing of SSC focus on optical domains, i.e. visible and near-infrared bands. Here, we assume thermal infrared bands can potentially contribute to retrieving SSC, inspired by sediments' heat-trapping effect demonstrated in aquatic research. Thus, we propose to incorporate the thermal infrared bands of Landsat-8 along with the optical bands for the estimation of SSC using machine learning models. In this study, 85 Landsat-8 images collected from the Google Earth Engine (GEE), coupled with in-situ observations from the United States Geological Survey (USGS) gauging stations, were employed to estimate SSC variabilities in Missouri-Mississippi Rivers. The relationship between in-situ SSC and water temperature revealed a promising correlation (R = 0.7), motivating our assumption of incorporating thermal infrared bands. Therefore, we employed both optical and thermal infrared bands of Landsat-8 imagery as predictors for the first time to estimate SSC. Three machine learning algorithms of Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Machine (SVM) were considered. The results demonstrated that incorporating thermal infrared bands improved the performance of the models (average improvements in R-2 for similar to 9.6% and in RMSE for similar to 18.9%) with respect to the use of optical bands only.