Towards Robust Semantic Segmentation of Land Covers in Foggy Conditions

Abstract

When conducting land cover classification, it is inevitable to encounter foggy conditions, which degrades the performance by a large margin. Robustness may be reduced by a number of factors, such as aerial images of low quality and ineffective fusion of multimodal representations. Hence, it is crucial to establish a reliable framework that can robustly understand remote sensing image scenes. Based on multimodal fusion and attention mechanisms, we leverage HRNet to extract underlying features, followed by the Spectral and Spatial Representation Learning Module to extract spectral-spatial representations. A Multimodal Representation Fusion Module is proposed to bridge the gap between heterogeneous modalities which can be fused in a complementary manner. A comprehensive evaluation study of the fog-corrupted Potsdam and Vaihingen test sets demonstrates that the proposed method achieves a mean F1score exceeding 73%, indicating a promising performance compared to State-Of-The-Art methods in terms of robustness.