When the Clouds Move In: How Simulated Cloud Occlusion Degrades Satellite Land-Cover Classification — and What Recovers It
Operational land-cover products are validated, almost without exception, on clear-sky imagery — yet the imagery that arrives in the wet season is not clear. We quantify what that gap costs on the EuroSAT benchmark (Sentinel-2 RGB, 24,300 tiles, 10 classes). A compact convolutional network trained only on clear tiles reaches 83.6% test accuracy, but it does not degrade gracefully when clouds arrive: under simulated cloud occlusion it falls to 35.6% at 10% coverage and 13.8% at 25%, and at 50% coverage it collapses almost entirely onto a single bright-scene class, predicting Industrial for 2,396 of 2,399 test tiles. Training the same network with cloud augmentation costs 7.3 points of clear-sky accuracy (76.2% vs 83.6%) and buys back nearly all of the loss: 74.6% at 10% coverage, 68.6% at 50% and 58.0% at 75%. The classes that fail first are the thin, linear ones — Highway and River are effectively erased at high coverage even for the cloud-trained model — while area classes such as Forest and Residential persist. A simple brightness-based cloud estimator tracks true coverage closely (r = 0.98), but within a fixed coverage level it adds nothing to abstention decisions beyond the model's own confidence: answering only the most confident 43.8% of tiles at 50% coverage restores accuracy on answered tiles to 89.9%. We argue that clear-sky validation materially overstates the operational accuracy of land-cover classifiers, that cloud augmentation is a cheap and effective first defence with a real but bounded clean-sky cost, and that abstention — not prediction — is the correct behaviour for the most occluded tiles.
Authors
- Saagar Sengupta
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-13
- DOI
- https://doi.org/10.5281/zenodo.24208972
- Primary Topic
- Remote-Sensing Image Classification
- Type
- preprint