Currently, the inversion of remote sensing satellite images of water environment indicators mostly stays in the indicators with active optical characteristics, while there is less research on the inversion of most water quality indicators with non-optical activity properties, weak scattering and absorption of optical radiation, the size of their concentration has little effect on the spectral characteristics of the water body, such as TOC(Total Organic Carbon). In this paper, based on Pingzhai Reservoir, a dammed river in the karst mountainous area, the inversion model of TOC concentration was established based on BP neural network (BPNN) and sentinel-2 satellite remote sensing images. The results showed that the single bands with high correlation with the measured TOC concentration data were two vegetation red-edge bands B6 (740 nm) and B7 (783 nm) and one NIR band B8 (842 nm), and finally b7, b6 + b7, b7 + b8, b7 × b8 were selected as the input layers of BPNN for modeling through the combination of the bands, and their Pearson’s coefficients were −0.667, −0.656, −0.655, − 0.675.The inverse model established could reach a minimum RMSE of 0.235 mg/L and a maximum R 2 of 0.889, which was superior to that of the conventional empirical model. Demonstrate the feasibility of a TOC inversion method based on Sentinel-2 data and BPNN to monitor TOC concentrations in Pingzhai Reservoir. The study successfully established a BP neural network inversion model of TOC concentration in Pingzhai Reservoir with low error, meanwhile, we analyzed the correlation between common water quality indicators and TOC in the reservoir, in which TOC showed significant positive correlation with WT and significant negative correlation with TN and EC, with Pearson’s coefficients of 0.655, −0.666, and −0.393, respectively. The article provides scientific theoretical foundation and technical support for water quality protection of water sources.