Multi-layer Domain Generalization for the Semantic Segmentation of Optical Remote Sensing Images

Supervised learning is the dominant training paradigm of semantic segmentation. However, it assumes that the deployment (or target) data follow the same distribution as the training data. And unfortunately this assumption is often violated in practice, leading to the infamous domain shift, coupled with degraded segmentation performances. Domain generalization addresses this issue, by exploiting the source (or training) domains, so as to maximize performance with unseen target data. This paper focuses on domain generalization for the semantic segmentation of optical remote sensing images through an adversarial strategy and investigates the concurrent use of feature maps from multiple layers, and proposes enforcing progressive domain invariance from earlier to latter network layers. The proposed approach is validated with the FLAIR dataset, where it achieves superior performance w.r.t. state-of-the-art studies.