Enhanced Building Footprint Extraction from High-Resolution Satellite Imagery Using U-Net++

An essential task in geospatial analysis, urban planning, and catastrophe management is building extraction from satellite photos. The effectiveness of the U-Net++ model for separating building footprints from high-resolution satellite photos is assessed in this work. By utilizing U-Net++'s sophisticated design, our method seeks to improve segmentation efficiency and accuracy. The model's performance was evaluated by comparing predicted masks with ground truth data on a variety of urban landscape datasets. The evaluation criteria consist of the following: accuracy, precision, recall, F1 score, mean Dice loss, and Intersection over Union (IoU) score. With a mean IoU score of 0.8199, mean recall of 0.9336, mean precision of 0.8696, and mean accuracy of 0.9014, the U-Net++ model performed well. These outcomes show how reliable the model is in drawing precise lines around building boundaries in complex urban environments.