Informal settlement mapping from very high-resolution satellite data using a hybrid deep learning framework

Informal settlements are becoming increasingly common in the global south, and their locations are often inaccurately represented in government statistics and maps. In remote sensing (RS), classifying informal settlements from very high-resolution satellite images presents a significant challenge due to their unique spectral signatures. This article proposes a novel hybrid framework, multi-shape-multisize-morphological profiles-dual-channel convolutional neural network (MSh-MSi-MP-DC-CNN) architecture, which fully exploits the feature extraction capabilities of both morphological profiles (MP) and convolutional neural networks (CNN) for detecting informal settlements from RS images. Three different images were used to evaluate the proposed MSh-MSi-MP-DC-CNN architecture. One was taken by the WorldView-2 dataset (1.82 m resolution) of Tiruppur in India, and the other two were taken by the WorldView-3 dataset (0.31 m resolution) of Kibera in Kenya and Medellin in Colombia. By extracting both spatial and structural features, the proposed CNN architecture proves robust and scalable for RS images of complex urban slum environments. It also generates significantly higher classification accuracy with less computing time compared to other state-of-the-art CNN methods considered in this work.