Building Segmentation from Remote Sensing Data using Enhanced U-Net

Abstract

In this paper, a simple boundary-enhanced network and a new multi-task loss function are proposed for building segmentation from multiple remote sensing sources. The experimental results on MapAI-challenge dataset demonstrates that our network can segment buildings from remote sensing data, especially on the image & laser multi-source dataset. The mean score of IoU and BIoU is 0.8995 for task 1 and 0.9155 for task 2 on the MapAI validation dataset and 0.5239 for task 1, and 0.7038 for task 2 on the MapAI test dataset.