HigherNet-DST: Higher-Resolution Network With Dynamic Scale Training for Rooftop Delineation

High-definition (HD) maps of building rooftops or footprints are important for urban application and disaster management. Rapid creation of such HD maps through rooftop delineation at the city scale using high-resolution satellite and aerial images with deep-learning methods has become feasible and has drawn much attention. However, the scale variance issue in rooftop delineation limited the overall performance. Existing methods exhibit considerably poor performance in the rooftop delineation of small buildings. In this article, we propose a new method, namely the higher-resolution network with dynamic scale training (HigherNet-DST) to overcome the scale variance problem in rooftop delineation. Specifically, dynamic scale training (DST) is applied in the model training phase to reduce the negative impact of scale variance. Then, a scale-aware backbone, namely the Higher-Resolution Network, is adopted to enhance the feature representation. Finally, the high-resolution supervision targets are used to further boost the delineation performance. Our method was tested on four publicly accessible building datasets and the results demonstrated that our method achieved the highest performance in rooftop delineation among the existing methods. Extensive experiments showed the superior performance of our method with an average precision (AP) of 68.5% on the AICrowd Building Dataset and an intersection of union (IoU) of 82.6% on the Inria Building Dataset, respectively, which surpassed many state-of-the-art (SOTA) methods. On the WHU Building Dataset and the Waterloo Building Dataset (WBD), our method also achieved the highest performance among the benchmarked methods, showing the high performance of our method for building boundary delineation.