Deep learning-based mapping of total suspended solids in rivers across South Korea using high resolution satellite imagery

Managing total suspended solids (TSS) in inland water is crucial for maintaining water quality and aquatic ecosystems. To obtain spatial data, the conventional method of point sampling TSS is a time- and labor-intensive task. Thus, recent efforts have focused on using satellite imagery to estimate TSS. Previous studies have used empirical method and machine learning model with multispectral satellite bands; however, these are limited to specific study areas and the amount of data is insufficient. This study aimed to build a generalized estimating model that can generally estimate the spatial and temporal variations in TSS concentrations in an extensive region by integrating the Water Quality Monitoring System with Sentinel-2 satellite observations using deep learning algorithms. This would provide comprehensive coverage of continental terrain, enabling consistent acquisition of satellite data and computation of the TSS across four major rivers in South Korea. Deep learning algorithms can generally estimate TSS concentrations over large areas. We found that the convolutional neural network (CNN) model was more accurate than traditional regression and other models, with a Nash-Sutcliffe efficiency (NSE) of 0.758. These findings indicate that it is possible to estimate TSS concentrations in ungauged areas, allowing the acquisition of extensive and continuous spatiotemporal datasets for organic micropollutants. This would be of great assistance in monitoring micropollutants in the four major rivers in South Korea.