Automated building extraction from very high-resolution remote sensing data with deep learning approaches

Abstract

Building extraction from very high-resolution satellite images is an important task due to the various different usage of the extracted information such as population estimation, city planning, and disaster management. Manual extraction of the buildings is a labor-intensive task that is prone to human-induced errors ad mistakes. Index-based and classical machine learning approaches remain insufficient due to diversity in building geometries, changes in reflectance values, and similar properties with other objects. Recently deep learning-based approaches show promising developments and results for this task. Unet architecture is one of the most popular deep learning architectures for building extraction within the scope of semantic segmentation. This study aims to automatically detect buildings by using the Unet architecture. The Unet model was trained twice with the same hyperparameters and Resnet50 backbone on 50 epochs initially with the Massachusetts building detection dataset and secondly with a combination of Massachusetts and Inria datasets to perform a comparative evaluation. According to the independent testing results with data from Massachusetts, Inria, Pleiades and Google Earth, both datasets provided satisfactory IoU scores ranging between 0.71 and 0.89, except for the first dataset testing with Pleiades images that provide a 0.51 IoU score.