Automatic building footprint extraction from very high-resolution imagery using deep learning techniques

Abstract

Building footprint maps are useful for urban planning, infrastructural development, population estimation and disaster management. With the availability of very-high resolution satellite imagery, remote sensing community is pursuing automatic techniques for extracting building footprints for cities with varried building types. Recently, CNNs (Convolutional Neural Network) have been successfully applied for extraction of building footprint from satellite imagery. In this paper, we propose a novel CNN architecture termed UNet-AP inspired by UNet and the concept of Atrous Spatial Pyramid Pooling, for automatic extraction of building footprint from very-high resolution satellite imagery. We demonstrate extraction of building footprints from Cartosat-2 series 4-band (Blue, Green, Red and Near-Infrared) multispectral satellite imagery, pan-sharpened using the panchromatic image with less than 1-meter resolution. We also compare the performance of our proposed architecture with baseline implementation of recently proposed UNet and SegNet architectures. We present a comparative assessment of the architecture performance across different types of urban settlement classes such as dense built-up areas, slums and isolated buildings. We demonstrate that our proposed architecture outperforms SegNet and UNet in terms of overall mean intersection over union (0.75 vs 0.70 and 0.68 for UNet and SegNet respectively) and delivers consistent improvement across all three settlement classes.