Slum image detection and localization using transfer learning: a case study in Northern Morocco

Abstract

Developing countries are faced with social and economic challenges, including the emergence and proliferation of slums. Slum detection and localization methods typically rely on regular topographic surveys or on visual identification of high-resolution spatial satellite images, as well as socio-environmental surveys from land surveys and general population censuses. Yet, they consume so much time and effort. To overcome these problems, this paper exploits well-known seven pretrained models using transfer learning approaches such as MobileNets, InceptionV3, NASNetMobile, Xception, VGG16, EfficientNet, and ResNet50, consecutively, on a smaller dataset of medium-resolution satellite imagery. The accuracies obtained from these experiments, respectively, demonstrate that the top three pretrained models achieve 98.78%, 97.9%, and 97.56%. Besides, MobileNets have the smallest memory sizes of 9.1 Mo and the shortest latency of 17.01 s, which can be implemented as needed. The results show the good performance of the top three pretrained models to be used for detecting and localizing slum housing in northern Morocco.