U-Net Cracks on Raw Sentinel-2

The main failure mode here is probably not the U-Net itself; it’s the jump from cloud-free, curated composites to messy real-world Sentinel-2 scenes with different illumination, viewing geometry, and atmospheric conditions.

• –Stacking dates can help, but only if you first handle clouds and temporal mismatch; otherwise you add noise instead of signal
• –Normalize aggressively for radiometry and geometry, and expose the model to sun/view-angle variation during training with augmentation or metadata-aware inputs
• –Fine-tune on a small, labeled target-domain set from the same Sentinel-2 pipeline you plan to use in production
• –Consider multi-temporal compositing, cloud masks, and possibly Sentinel-1 fusion if clouds are a recurring problem
• –Calibrate confidence separately; “almost zero confidence” often means the model is out of distribution, not that segmentation is impossible