Semantic Segmentation in Satellite Imagery: An Attentive U - Net Approach

Semantic segmentation is a computer vision technique that assigns a distinct class label to every pixel in an image, generating a segmentation map where each pixel belongs to a specific category. This enables various applications, such as automatic and in-depth land cover and land use analysis, tracking urbanization, deforestation, road, and building detection, disaster response, traffic management, city planning, and environmental monitoring. Encoder-decoder-based architectures excel in semantic segmentation tasks than traditional CNN-based methods because they can extract hierarchical features, combine contextual and spatial information, and accurately reconstruct segmentation maps. In this paper, Attention U-Net, where the attention mechanism is introduced to the modified version of U-Net, is implemented for semantic segmentation in satellite images. Instead of just focusing on extracting features, the proposed method emphasizes making accurate segmentation output. The segmentation quality of each model was assessed using performance metrics, including accuracy, precision, recall, F1-score, and IoU score. The findings suggest that the Attention U - Net model surpasses other variations, attaining a higher IoU score of 0.7959 and exhibiting strong segmentation abilities across several land cover categories, such as water bodies, land, roads, buildings, and vegetation. Furthermore, by visual analysis of the results, it becomes apparent that the Attention U-Net is remarkably proficient at accurately capturing intricate details in satellite imagery,