Accurate estimation of suspended sediment concentration integrated remote sensing information and a novel stacking machine learning model

Suspended sediment concentration (SSC) in rivers significantly impacts the preservation of the ecological environment and the exploitation of water resources. The advancement of remote sensing technique offers a robust approach for monitoring SSC. However, the complexity of watersheds and the surrounding environment present a new challenge for accurate estimation of SSC. To address this limitation, this study proposes a new stacking model considering Multilayer Perceptron and Light Gradient Boosting Machine with Elastic Net algorithm (MLEN), and integrates remote sensing information for precise estimating SSC. The Tree-structured Parzen Estimator method was adopted to optimize hyperparameters, the MLEN model was trained by reconstructed datasets combining surface reflectance from high-quality Landsat remotely-sensed images over 30 years, with environmental factors including precipitation, temperature, wind, and surface pressure from ERA5 dataset, as well as discharge and SSC data from USGS five hydrographic stations of the Middle Rio Grande River Basin in the United States. Those stations were selected with over 30 years of available data and nearby gauged stream widths of at least 90 m to ensure local characteristics and reliable satellite sampling. Moreover, the contribution of features on estimating SSC was also discussed in detail. The results show that compared with the individual models, the MLEN model achieved best accuracy in estimating SSC. Furthermore, the MLEN model also outperformed the other five machine learning algorithms (R2 = 0.80, RMSE = 0.44, and MAPE = 0.30). It indicates the MLEN model can effectively predict SSC in complex, long-term, and time-varying watersheds with readily available hydrographic data.