Uncertainty-Aware SegFormer for Satellite Image Segmentation Using Evidential Deep Learning

Abstract

Vision Transformers and other deep learning models are highly accurate in segmenting satellite images, but they tend to behave as overconfident black boxes and pose a safety risk for critical remote sensing applications. We present an Uncertainty-Aware SegFormer that combines Evidential Deep Learning (EDL) with a single forward pass to predict epistemic uncertainty. In the DeepGlobe Land Cover dataset, we reach a mean Intersection over Union (mIoU) of 65.81%. Through a series of experiments, we vary the weight of the Kullback-Leibler (KL) regularization, and find a trade-off: a larger regularization weight increases the correlation between uncertainty and error, but it also decreases spatial accuracy and calibration performance. There is no optimal setting for all three objectives. Compared to Monte Carlo (MC) Dropout, our single-pass EDL method provides better uncertainty-error localization, though MC Dropout achieves slightly better accuracy and calibration at the cost of 1.5 times more computing time. Furthermore, quantitative out-of-distribution (OOD) evaluation using synthetic corruptions confirms that the uncertainty of the model reliably increases as image corruption worsens. However, the ability to separate clean from corrupted images is constrained by a high baseline uncertainty caused by the KL training process. This research provides a systematic study of EDL tradeoffs in dense pixel-wise segmentation, revealing how current training mechanisms force a compromise between calibration, OOD detection, and accuracy.