Machine Learning-Based Estimation of Suspended Sediment Concentration along Missouri River using Remote Sensing Imageries in Google Earth Engine

Abstract

Estimation of Suspended Sediment Concentration (SSC), regarded as a crucial component of hydrological and ecological processes, can provide a better understanding of water quality. This study aims to use Sentinel-2 (S2) level-2A (L2A) images with less than 1% cloud coverage and supervised machine learning-based regression models to estimate SSC along the Missouri River. The model gets the reflectance values of different spectral bands and predicts the corresponding SSC value for each water pixel. Time-series data of three different ground measuring stations and surface reflectance values of the closest pixel to each station are used to train and validate the model. Two popular regression models, Support Vector Regression (SVR) and Random Forests (RF), are trained, validated, and compared online in the Google Earth Engine (GEE) processing platform by using 68 satellite images, without the need to be downloaded. The results demonstrated that the RF model with a root mean square error (RMSE) and mean absolute error (MAE) of 59.521 and 46.493 mg/L outperforms the SVR model. Moreover, the RF model resulted in a higher correlation between the real and predicted SSC values (R2 = 0.79 and Pearson's r = 0.877). Also, the two short wave infra-red (SWIR) and red bands play more considerable roles in SSC estimation using S2L2A images than other bands.