Disaggregated data on housing and household economic conditions, particularly in developing countries, are often inaccessible to urban planners and local users, and conducting independent in-person surveys is prohibitively expensive. This impedes the development of precisely informed policies and targeted infrastructure investments, services, and incentives that are essential for fostering equitable and sustainable cities. Using Kigali and Musanze cities in Rwanda, this study introduces a novel deep learning-based approach that combines limited expert annotation, self-training, and instance image segmentation to generate building-level housing wealth data. Our experiment included state-of-the-art instance segmentation based on You Only Look Once (YOLO) and Masked-attention Mask Transformer (Mask2former) models. To check the validity of the predictions, over ten thousand samples from the official cadastre-based property taxation database were used to evaluate whether the observed patterns in actual property values align with the model’s predictions. Importantly, our method successfully detected approximately 70% of the existing buildings within their respective perceived wealth classes. Buildings predicted to belong to the high-wealth class exhibited mean property values 2.5-times higher in Kigali and 1.7-times higher in Musanze than those classified as low-wealth, thereby demonstrating the model’s ability to capture spatial patterns of housing wealth. The predicted maps reveal that housing wealth growth in urban cores parallels increased low-wealth housing in peri-urban areas, a likely result of gentrification, an insight that could inform more inclusive urban development strategies and support policies aimed at curbing urban sprawl. Our findings demonstrate that combining minimal expert labeling of economically meaningful visual features with self-training and multi-class instance segmentation offers an effective and rapid approach for generating precise planning data in contexts that lack detailed local census information. Moreover, while leveraging advanced deep networks, our methodology was designed to be simple and easily replicable.