Extracting Building Footprints from Satellite Images using Convolutional Neural Networks

Abstract

Extracting building footprints from high resolution satellite(HRS) imagery has been an extremely important topic of research in the field of remote sensing. The problem of extracting building footprints is challenging due to a high variation of building shapes, sizes and appearances. This paper presents an approach for extracting and detecting building footprints from satellite images. The proposed technique is capable of extracting buildings having different shapes, sizes and textures. We use a Convolutional Neural Network(CNN) architecture which gives pixel-wise predictions. Due to a large variation in the satellite images, we pre-process the images before passing them to the network. Finally, we apply some post-processing algorithms to regularize the extracted building footprints and also remove some false-positives. The experiments were performed on a set of high resolution satellite images from Bing. The performance of the developed technique was assessed by comparing the results of prediction with the manually created reference data by using metrics like Precision, Recall and F1-Score. Pixel based and Object based evaluations of the test images with diverse characteristics produced an overall precision of 84.7% and 99% respectively.