Attention Mechanisms in Semantic Segmentation of Remote Sensing Images

Abstract

This research addresses the persistent challenge of accurately segmenting complex, high-resolution aerial and satellite imagery. The primary objective is to enhance semantic segmentation performance by leveraging advanced attention mechanisms within deep convolutional neural network (CNN) architectures. Unlike traditional segmentation approaches that often struggle with heterogeneous landscapes and intricate object boundaries, our methodology systematically integrates channel attention, spatial attention, and transformer-based attention modules into standard encoder-decoder CNN frameworks. Specifically, channel attention focuses on strengthening feature representation by adaptively recalibrating channel-wise responses, while spatial attention guides the network to prioritize salient regions across spatial dimensions. The transformer-based attention component captures long-range dependencies, enabling more coherent global context aggregation, which is crucial in remote sensing scenes characterized by scale variation and spatial complexity.

The proposed approach is evaluated on benchmark datasets widely acknowledged in the remote sensing field, including ISPRS Potsdam, Deep Globe Land Cover Classification, and Space Net. These datasets offer diverse urban and rural scenes, challenging the segmentation models to generalize across variable geographic and environmental contexts. Empirical results demonstrate that our proposed attention-infused models consistently outperform baseline CNN architectures in both overall segmentation accuracy and the delineation of fine-grained boundaries. For instance, on the ISPRS Potsdam dataset, our best-performing model achieves a 3.8% absolute improvement in mean Intersection-over-Union (mIoU) compared to established baselines.