U.S.-Wide Suspended Sediment Concentration Estimation Using Machine Learning and Sentinel-2 Data

Suspended sediment concentration (SSC) is measured at relatively few river sites, which limits the coverage available for large-scale assessment. We built a nationwide Sentinel-2–USGS matchup dataset and used it to test how well machine learning can predict SSC between 0 and 10,000 mg/L. A three-gate pipeline combining temporal matching, adaptive spatial extraction, and cloud-screened water-pixel quality control produced 25,474 quality-passing matchups at 1170 sites, of which 25,095 observations were retained for modeling. Every site–image group was kept entirely within either the training or test set. A single unified model reached R2 = 0.399 across the full range. Because the dataset is strongly imbalanced across concentration ranges, we also examined performance by regime. Conditional models that were given the observed regime reached pooled R2 = 0.572 for three regimes and 0.907 for five, yet the within-regime R2 in the five-regime case ranged only from −0.243 to 0.389. The high pooled value therefore reflects mostly what is gained by knowing the concentration range, not skill at predicting within it. When the regime had to be predicted instead, a two-stage workflow reached R2 = 0.405, which is essentially the unified result, although MAE improved from 187.93 to 156.83 mg/L and the share of predictions within ±30% relative error rose to 33.62%. Both regime identification and within-regime prediction limit performance. A complementary complete site-level holdout retained R2 = 0.379 but showed higher error, indicating that transfer to entirely unseen monitoring locations is more difficult than prediction for new site–image conditions when sites may already be represented during training.