Training labels for satellite image segmentation are in practice derived from lower-resolution products and are therefore coarse and noisy. This thesis systematically studies how modern deep learning methods for semantic segmentation perform when trained on such labels: models are trained on the SEN12MS dataset with 500 m MODIS labels and evaluated on the manually verified test set of the DFC2020 contest. In 62 experiments we compare eleven models from three families (convolutional, transformer and hybrid architectures, and foundation models) and evaluate pretrained and frozen encoders, the robust loss functions GCE and SCE, and our own technique of excluding pixels near class boundaries. Pretraining turns out to be by far the most important factor; the foundation models FlexiMo and DEO-FM perform best, while a plain ViT with ImageNet weights trails them by only about 0.01 mIoU. A frozen encoder is competitive with full fine-tuning for transformer models, whereas robust losses and boundary exclusion bring no systematic improvement, because the label noise is structured rather than random. We further show that evaluation on noisy labels overestimates performance and ranks models incorrectly.
|
