Cross-Sensor Generalization in Soil Sealing Segmentation on Very-High-Resolution Imagery

Soil sealing, the permanent covering of land with impermeable materials, is a direct sign of urbanization and a recognized indicator of land degradation, monitored under EU policy in support of the target of no net land take. Producing accurate national-scale soil sealing maps is hard because of sensor heterogeneity: imagery comes from many platforms at different resolutions, band configurations, and radiometric characteristics, and a model trained on one sensor degrades sharply on an unseen sensor. This thesis asks whether sensor-agnostic deep learning can deliver dependable soil sealing maps across heterogeneous sources. Three Transformer-based architectures, trained on the same SPOT imagery (four-band RGB+NIR, 1.5m), are compared: SegFormer SAE, with a custom Spectral Awareness Embedding; CrossEarth, a DINOv2 foundation backbone fine-tuned end-to-end with a Mask2Former head; and DOFA, conditioned on each band’s wavelength. A convolutional U-Net, trained under the same budget, is included as a second architectural family. The models are integrated into a unified training, inference, and evaluation framework with sensor-agnostic preprocessing and a tiled predictor for large mosaics. All models are evaluated cross-sensor on five areas along the Italian coast (Ferrara, Ravenna, Venezia, Brindisi, Bari), each acquired by a different sensor at 0.5m-1m resolution, unseen during training. In domain, U-Net scores highest (mIoU 0.834), ahead of SegFormer-SAE (0.818), DOFA (0.773), and CrossEarth (0.764). Cross-sensor evaluation inverts this ranking: CrossEarth, last in domain, generalizes best (avg. mIoU 0.84), while U-Net (0.67), DOFA (0.66), and SegFormer-SAE (0.63) drop sharply, sometimes nearly failing to detect the sealed class while reporting high aggregate accuracy from class imbalance. In-domain performance does not predict cross-sensor transfer: foundation models with spectral agnostic representations offer a more reliable basis for multi-sensor monitoring.