A Spatiotemporal Transformer for Satellite Retrieval of Lake Turbidity

Lake turbidity governs light availability, habitat quality and drinking water treatment cost, and most lakes worldwide have no routine ground monitoring. Satellite retrieval is therefore central to tracking it at scale. Published models report validation accuracy high enough to suggest the problem is solved. We show that this impression comes from how these models are tested. Using a new Sentinel-2 matchup dataset built from tight same-day pairings across thousands of lakes, we evaluate under four progressively harder protocols and find that apparent accuracy falls sharply once whole lakes and whole regions are held out. We then identify a failure mode specific to neural retrieval. Given latitude and longitude, a spatiotemporal transformer learns a location-indexed correction that is invisible during in-distribution validation and collapses completely on another continent. Removing three coordinate inputs repairs it at no measurable domestic cost. Our model then outperforms gradient boosting, the operational Dogliotti and Nechad algorithms, and every other baseline we tested on out-of-continent transfer, and our models take the best result on both spatial generalisation protocols. The result gives practitioners a concrete rule for building retrieval models that work in the ungauged regions where monitoring is most needed. We release the dataset, code and trained models.