UNet with global feature attention for multi-scale feature fusion in building extraction from satellite and aerial imagery

Extracting building footprints is a complex and major process in geospatial object detection and recognition from remote sensing imagery. Because of the complex interaction among a building’s body and its boundaries in remote sensing images, many of the most advanced deep learning (DL) models have problems in achieving high accuracy footprint representation and often have performance limitations. To address these challenges, this work presents a UNet-based segmentation model enhanced with a Global Feature Attention Module (GFAM). The GFAM is designed for integrating local structural details and global contextual dependencies. This allows the network to show discriminative building features and eliminate irrelevant background noise. Integration of this attention module in the UNet, the suggested model captures multi-scale spatial features and improves feature representation. The framework was evaluated on benchmark satellite image datasets, and its performance was experimentation I assessed using two benchmark datasets and attained accuracy values of 99.55% (Inria) and 99.53% (Spacenet-7). Thus, the suggested model demonstrates the potential of context aware feature learning for large scale and reliable building extraction from satellite and aerial imagery.