Machine-Learning Code for U.S.-Wide Sentinel-2–USGS Suspended Sediment Concentration Estimation
This software release contains the four principal site–image-grouped machine-learning workflows used to develop, validate, and evaluate suspended sediment concentration (SSC) models in the associated study, “U.S.-Wide Suspended Sediment Concentration Estimation Using Machine Learning and Sentinel-2 Data.” The included workflows are: (1) a unified single-stage model across the full 0–10,000 mg/L modeling range; (2) a three-regime conditional reference in which the observed SSC regime is supplied before regression; (3) a five-regime conditional reference in which the observed SSC regime is supplied before regression; and (4) a fully predictive two-stage classifier–regressor workflow in which the SSC regime is predicted from the available inputs before the corresponding regime-specific regressor is applied. The two principal feature configurations follow the terminology of the manuscript: the original predictor set, consisting of Sentinel-2 reflectance summaries, spectral indices, matchup-context variables, latitude, longitude, and state; and the spectral + hydrologic interaction set, which additionally incorporates the hydrologic predictors used in modeling and selected spectral–hydrologic interaction terms. The scripts use the processed Sentinel-2–USGS SSC matchup dataset available separately on Zenodo at https://doi.org/10.5281/zenodo.22298970. The public dataset contains 25,474 quality-passing matchups; modeling-domain filtering retains 25,095 observations for the principal analyses. Site–image-group validation uses a grouping key based on site_no_norm + S2_image_id so that the same site–image group is not shared between the outer training and test sets. The three-regime conditional reference uses 5,088 grouped test observations. The unified, five-regime conditional-reference, and fully predictive two-stage workflows use the same 5,012-observation grouped test split. This is a source-focused software release. Earlier row-level development experiments, supporting descriptive/diagnostic analyses, large trained-model binaries, and complete historical output trees are not redistributed. Associated study authors: Sathvik Reddy Nookala, Jennifer G. Duan, and Kun Qi, University of Arizona.