A Residual Attention U-Net Approach for High-Resolution Satellite Image Segmentation

Semantic segmentation of high-resolution satellite images is a critical task in remote sensing, enabling detailed analysis for applications such as urban development, environmental monitoring, and disaster management. This paper introduces a Residual Attention U-Net architecture that integrates residual blocks for effective feature extraction, improving segmentation accuracy. To facilitate focused attention on relevant spatial regions, attention gates are also included. These improvements allow the model to address the challenges of complex and diverse aerial imagery by maintaining essential spatial details and reducing background noise. Our method is evaluated on a high-resolution aerial dataset that includes multiple land cover categories, such as water bodies, roads, buildings, vegetation, and open land areas, allowing for a comprehensive assessment across different types of terrain. The Residual Attention U-Net achieved strong performance across multiple evaluation metrics including Accuracy and Intersection over Union (IoU), with an overall IoU score of 81.07% and an accuracy of 92.33%. Comparative analysis with other U-Net variants demonstrates the superior capability of our model to capture intricate details and boundaries within satellite images, making it a valuable tool for precise segmentation in remote sensing applications. These findings underscore the potential of the Residual Attention U-Net in providing high-accuracy segmentation for high-resolution satellite imagery, paving the way for its use in various geospatial and environmental studies.