Efficient Building Inventory Extraction from Satellite Imagery for Megacities

Abstract

Accurate building inventories are essential for city planning and disaster risk management. Traditionally generated via census or selected small surveys, these suffer from data quality and/or resolution. High-resolution satellite imagery with object segmentation provides an effective alternative, readily capturing large extents. This study develops a highly automated building extraction methodology for location-based building exposure data from high (0.5 m) resolution satellite stereo imagery. The development relied on Taipei test areas covering 13.5 km2 before application to the megacity of Jakarta. Of the captured Taipei buildings, 48.8% are at one-to-one extraction, improving to 71.9% for larger buildings with total floor area >8000 m2, and to 99% when tightly-spaced building clusters are further included. Mean absolute error in extracted footprint area is 16% for these larger buildings. The extraction parameters are tuned for Jakarta buildings using small test areas before covering Jakarta's 643 km2 with over 1.247 million buildings extracted.