Evaluation of Machine Learning Algorithms for Urban Building Extraction from High-Resolution Satellite Imagery in Kano City

Automated building extraction from high-resolution satellite imagery is a critical input to urban population estimation, infrastructure planning, and property tax enumeration in rapidly growing Nigerian cities, yet the performance of machine learning algorithms for this task in the morphologically complex urban environments of northern Nigeria has not been systematically evaluated. This study evaluates the performance of four machine learning algorithms, namely Support Vector Machine, Random Forest, Convolutional Neural Network, and U-Net deep learning architecture, for automated building extraction from 50-centimetre Pleiades-1 imagery covering a 20-square-kilometre area of Kano City. A manually digitised reference dataset of 12,400 building polygons is used as ground truth, with 70 percent allocated to training, 15 percent to validation, and 15 percent to testing. Performance metrics including precision, recall, F1 score, intersection over union, and boundary delineation accuracy are computed. The U-Net architecture achieves the highest F1 score of 0.89 for building detection, while Random Forest demonstrates the best boundary delineation performance for regular building shapes. All algorithms perform less accurately on compound wall structures and mud-brick buildings with low spectral contrast against bare soil, which are prevalent in the old city densely built areas. The study provides algorithm selection guidance for automated building extraction applications in Kano and comparable North West Nigerian city environments, and identifies pre-processing strategies to address the identified accuracy limitations. Keywords: building extraction, machine learning, high-resolution imagery, Kano, U-Net.