Satellite-driven monitoring of suspended sediment concentration in large river systems: an application to the Loire River, France

Suspended sediment concentration (SSC) is a key indicator of river morphodynamics and water quality, yet long-term spatial monitoring remains limited because in situ measurements are sparse and discontinuous. By integrating field observations with multispectral Landsat imagery, this study develops a satellite-based framework to estimate SSC along the Loire River (France) over the period 2005–2025 in low to medium flow conditions with an SSC below 100 mg/l. Reflectance data from Landsat 5, 7, 8, and 9 were extracted using Google Earth Engine and matched with in situ SSC measurements within a ± 5-day temporal window. A feed-forward neural network was trained to predict SSC from selected spectral bands (B2, B3, B4, B5, and B7), with separate models developed for low, intermediate, and high turbidity regimes to improve predictive stability, and then recombined to obtain a single model. This complete neural network achieved high performance ( R 2 = 0.941, RMSE = 3.78 mg/l, PBIAS = 6.53%). To enable large-scale operational mapping, the neural network predictions were approximated through a compact surrogate model based on quadratic feature expansion and ridge regression, maintaining strong accuracy in reproducing observed SSC ( R 2 = 0.890, RMSE = 5.17 mg/l, PBIAS = 9.92%). The surrogate model was implemented in Google Earth Engine to generate spatially continuous SSC maps along the river corridor. The proposed approach demonstrates the capability of combining machine learning and optical satellite imagery to efficiently reconstruct SSC patterns in the Loire River and provides a scalable framework for monitoring suspended sediment dynamics in large river systems, in particular during low to medium flow conditions.