An Appraisal of Deep Learning Algorithms in Automatic Building Footprint Extraction from High-Resolution Satellite Image in Parts of Akure, Nigeria

Authors: Nzelibe, I.U. And Akinboyewa, T.E.

DOI Info: http://doi.org/10.5281/zenodo.12599520

ABSTRACT

The accuracy of spatial objects extracted from raster images using automatic feature extraction algorithms remains a problem in the field of remote sensing. This study assesses the performance of three existing deep learning algorithms, viz: DeepLabV3, Pyramid Scene Parsing Network (PSPNET), and U-Network (U-NET) in extracting Building Footprints (BFP) from High-Resolution Satellite Images (HRSI). The assessment was performed on two study sites, High Building Densities (HBD) and Low Building Densities (LBD) areas, both located in the city of Akure, Ondo State, Nigeria. The HRSI used is the Google satellite image. Ground truthing data used as reference was derived by manually digitizing the HRSI after preprocessing. Results from the study reveal that UNet recorded the highest overall classification accuracy having values of 90.7 % and 97.6 % in HBD and LBD sites respectively. In terms of the area, the U Net recorded a Root Mean Square (RMS) error of 0.3 m2 and 0.2 m2 in HBD and LBD sites respectively. In terms of positional accuracy, the UNet recorded a mean Euclidean BFP centroid displacement of 1.4 m and 0.9 m in HBD and LBD sites respectively. The UNet model indicated the best accuracy compared to DeepLabv3, and PSPNet in automatic BFP extraction, performing better in the LBD sites compared to the HBD sites. Based on the study, the UNET model is found to be best suited for building types both in HBD and LBD development in and around the city of Akure, Nigeria.

Download Full Text