Adaptive Spatial Modeling and Multi-Scale Attention Aggregation for Semantic Segmentation of Remote Sensing Images

As remote sensing technology advances and high-resolution sensors are deployed, the analysis of high-resolution remote sensing images encounters challenging issues, such as intra-class variability, and inter-class similarity, due to the unobvious appearance and complicated background. The high-resolution remote sensing images pose critical issues for precise semantic segmentation due to these fundamental issues. To this end, we propose an adaptive hierarchical aggregation network. Different from the previous approach of modeling semantic interdependencies solely within a single feature map in the channel dimension, we introduce multi-scale channel-wise feature aggregation to capture global contextual dependencies. Meanwhile, we design an adaptive detail-aware module to model long-range spatial dependencies, adaptively extracting effective details from different levels as compensation. Additionally, this module utilizes depth-wise separable convolution to selectively filter out irrelevant spatial details from lower-level feature maps. The outputs of the two modules are fused layer by layer to achieve refined feature representation, thereby generating precise pixel-level segmentation results. Our experiments are conducted on the Vaihingen dataset and the Potsdam dataset. The experimental results demonstrate that the proposed algorithm surpasses several existing state-of-the-art approaches.