Developing a Deep Learning Framework for Automated Extraction of Urban Features from Very High-Resolution Satellite Imagery in Nigerian Cities

Automated extraction of urban features from very high-resolution satellite imagery is critical for timely urban mapping in rapidly growing Nigerian cities where manual digitising cannot keep pace with development. This study develops a deep learning framework for automated extraction of buildings, roads, and vegetation from WorldView-2 and Maxar satellite imagery across Lagos, Kano, and Enugu metropolitan areas. A Convolutional Neural Network architecture based on the U-Net encoder-decoder structure is designed and trained using 12,000 labelled image patches across the three city typologies. Transfer learning from pre-trained ImageNet weights is applied to improve training efficiency under limited labelled data conditions. The model incorporates multi-scale feature extraction to address the varying spatial context of urban features across density gradients. Results demonstrate overall extraction accuracies exceeding 91 percent for buildings and 88 percent for roads, outperforming conventional object-based analysis approaches by 9 and 7 percentage points respectively. The framework's performance is evaluated across formal planned settlements, informal settlements, and peri-urban areas to characterise domain transfer limitations. The study contributes an original deep learning methodology specifically calibrated for the morphological characteristics of West African urban environments. Keywords: deep learning, urban feature extraction, very high-resolution, Nigerian cities, convolutional neural network