Deep learning-based automatic building types classification for transport planning

Abstract

The present study proposes a building classification algorithm that uses deep learning techniques, namely object detection and image segmentation, to distinguish between residential and commercial structures. The algorithm is trained using images from the ISPRS Potsdam and Spacenet 3 (Vegas) datasets. According to the model's results, the model has obtained high precision, recall, and mean average precision (mAP) values for both classes. Despite the high-performance yield of the model, the robustness the model can be improved by expanding the training dataset to include more building images from diverse locations on Earth.