Building Extraction from High-Resolution Satellite Images using 2D-Attention Mechanism with Deep Learning

Abstract

Building extraction from remote sensing satellite images is very useful for the urban monitoring and its planning. Several methodologies based on CNN are proposed in past for building extraction. Many of them have also used the skip connections for better propagation of information between different layers. But the suppression of the irrelevant information from earlier layers helps to focus on relevant information and improve the building extraction performance. For this, our work applies the 2D-attention mechanism in one of the state-of-art model, i.e., Unet for extracting buildings from high-resolution satellite images. To further improve its results, the work investigates the optimal deep learning hyper-parameters through various experimentations with activation, loss function, ImageNet weights and various backbones, i.e., ResNet152_V2, VGG19. The work uses the Satellite dataset I (global cities) from WHU repository. The results show that our approach, i.e., 2D-Attention based Unet model along with ImageNet weights, ReLU activation and IoU loss function has better building extraction performance and can be utilized for societal perspective.