Remote estimates of total suspended matter in China’s main estuaries using Landsat images and a weight random forest model

Abstract

Total suspended matter (TSM) is a key parameter for coastal waters, and its transport in estuaries can regulate erosion-accretion and biogeochemical processes. However, most existing models were built to estimate TSM concentration in specific rivers or regions, such as the Yangtze River or Yellow River estuaries in China. Here, we propose a stable hybrid model to estimate TSM concentrations of different estuaries in China using local regression and a TSM mapping method based on the weight random forest (WRF) model. Two hundred thirty-eight in situ TSM samples were collected in six river stations from northern to southern China by the United Nations Environment Programme Global Environment Monitoring System. The surface reflectance for each sample was extracted from Landsat images considering the date and location of each TSM value in Google Earth Engine. One-variable, multiple stepwise linear regression (MSLR), and RF models were built to predict TSM concentration in global regressions (single model) across all calibration samples and local regressions (hybrid model) with a TSM threshold of 100 mg/L. A WRF model was used to map TSM concentrations based on the classification probability of a decision tree model. Our study showed that: (1) the Red band, band ratio, and band difference indices using the Red band as input were sensitive to the estuaries with low TSM concentrations, while the NIR band, band ratio, and band difference indices using NIR band were sensitive to highly turbid estuaries with high TSM concentrations; (2) the band difference index predicted TSM concentrations better in less turbid estuaries than the band ratio index; (3) the predictions of the RF model were better than those of the one-variable and MSLR models both for global regressions and local regressions — the local RF model had the best result, i.e., a validation R2 of 0.90 and an RMSE of 0.56 mg/L; and (4) the WRF model was more reasonable than the original RF model. TSM concentrations were overestimated using the original RF model. Our results suggested that the local regression was suitable for different estuaries with large differences in TSM concentration. The WRF model provided an effective approach for water quality parameter mapping.