Examining Model Generality of Instance Segmentation for Building Mapping in Satellite Images - Case Study for Tokyo and Bangkok

In this study, we created a building extraction model from satellite images of Tokyo, which is updated more frequently than developing countries and has abundant data for training, and examined the possibility of extrapolating it to Bangkok, one of the megacities in developing countries. A deep learning model using Meta's detectron2 library was used as the model for building extraction. The results of extrapolating to Bangkok with the model built in Tokyo showed that both IoU and Building Extraction Rate were more than 60% accurate. It was confirmed that this model can be extrapolated with the same performance as the model constructed in Bangkok for the extraction of buildings in Bangkok. This study also conducted a field survey in Bangkok to identify the causes of the obstacles to improving the accuracy of extrapolating the Tokyo model to Bangkok. The results revealed that the roof color and vegetation around the building, which are unique to Bangkok, affected the performance of the model.