Digital Satellite Image-Based Automatic Building Segmentation Using Pixel Attention Half-UNet

Building detection is important in geographic information updates and urban and regional planning. Building detection is more efficient when done automatically. However, automatic building detection is still a challenge in Geoinformatics. Buildings with the same colour, texture, and shape as their surroundings pose a challenge in automatic building detection. Due to their complexity, some traditional methods for automatic building detection do not have optimal performance. A neural network method called Pixel Attention Half-UNet is proposed to overcome this challenge. The proposed method is based on the Half-UNet framework, utilizing Pixel Attention. Pixel Attention Half-UNet was trained and tested using high-resolution satellite images from the Massachusetts Buildings Dataset. The proposed method gives F1-Score and Intersection-over-Union (IoU) results of 82.95% and 70.94%, respectively. The proposed method performs better than previous methods, namely DeepLabv3+ (Xception), RefineNet (ResNet50), U-Net, Half-UNet, SegNet, and Bayesian-SegNet. This method also provides a lower parameter, which is 0.26 million.