A Novel Active-Learning Based Residential Area Segmentation Algorithm

Gentrification and urban polarization has become one of the biggest societal issues of modern urbanization. While the apartment residential areas for affluent people are proliferating, poor people who cannot afford those areas are involuntarily forced to move out to other areas. In this paper, we proposed an active learning based residential area segmentation algorithm (AL-RASA) that can distinguish tenement area from the satellite images, which can show an imbalance of urban wealth. This paper used the Mask-RCNN as the backbone segmentation module, due to the high cost of labelling satellite image and the requirement of professional sociological knowledge, this paper used a method combining active learning to label the data efficiently and train the model. Finally, with the change detection model, this algorithm can produce an image of the change of tenement areas during last two decades. To better study residential area segmentation problem, this paper provided a high-resolution satellite images dataset of Seoul (South Korea).