A reproducible baseline framework for landsat satellite-based water quality modeling

Remote sensing can monitor water quality at a large scale and low cost, but many existing models are narrowly tuned and hard to generalize. To the best of our knowledge, this study introduces a reproducible, end-to-end evaluation framework for Landsat-based water quality modeling. It compares machine learning algorithms, water quality indicators, cloud-cover thresholds, temporal matching windows, and water body types using a single standardized dataset. Using AquaSat, with over 600,000 matched in situ and Landsat observations, we evaluated 27,720 model configurations across eleven algorithms, six indicators, seven cloud-cover thresholds, five water body types, and temporal offsets up to ± 48 h, under one harmonized hyperparameter configuration on the Landsat 5, 7, and 8 family. Every configuration was evaluated twice: under a random splitting protocol and 50 km spatially blocked validation. Under the random protocol, 95.8% of the held-out observations came from sampling sites also represented in the training data, while under spatial blocking, no sampling site was shared between the training and test sets. The median r 2 decreased by 0.087 between the two. Three of the six indicators retain meaningful skill under blocked validation: the best models for total inorganic sediment, Secchi depth, and dissolved organic carbon average an r 2 of 0.403, while chlorophyll-a and percent sand showed little or no predictive skill. These results indicate that poor blocked-validation performance primarily reflects limited spatial transfer rather than failure to fit the training data: 98.7% of the configurations with a negative held-out r 2 fitted their training data successfully, and a hyperparameter search across 15,578 paired comparisons, pooled over both splitting schemes, moved held-out performance by −0.001. Sensitivity analyses show that the effects of temporal matching and cloud cover are indicator-specific, and permutation importance on the held-out data shows that the visible bands dominate for every model with predictive skill. The accompanying open data and code support reproducibility, and we demonstrate operational use by applying the baseline models directly to independent Landsat imagery without retraining or local calibration. By measuring, rather than assuming, the cost of the validation design, this work provides a transparent, more conservative baseline against which future Landsat-based water quality models can be compared.