High-density urban cities contain numerous similar buildings positioned in close proximity. Building detection from high-spatial-resolution remote sensing imagery in such scenes remains a challenge in computer vision and remote sensing urban applications. The integration of traditional segmentation algorithms and a novel neural network is an effective approach for such challenging settings. Inspired by the recent success of deep-learning-based edge detection, a new building detection method aiming at accurate boundaries is proposed. In accordance with the characteristics of buildings and their border, this study improves the network structure and integrates the network with bottom-up watershed segmentation to improve boundary precision and classification accuracy.First, two auxiliary labels, namely, the building boundary and parting line, are derived from the original dataset through data preprocessing. Second, the newly proposed building detection frame called ICT-Net is improved by modifying its structure and loss function in accordance with the two auxiliary labels to obtain the probability of three classes. Lastly, a post-process integrating watershed segmentation with gradient-boosted regression trees is employed to achieve high-accuracy building detection. Specifically, a probability feature map is generated by merging the probability of three classes. Watershed segmentation with building marker thresholds is applied to obtain building instances from the probability feature map. Then, the building probability of each building instance predicted by gradient-boosted regression trees is used to select building instances, resulting in building detection results. Parameter selection is also implemented.The performance of the proposed method is validated on the INRIA dataset, which provides aerial orthorectified color imagery with a spatial resolution of 0.3 m and with corresponding ground truth labels for two semantic classes: building and not building. Experimental results suggest that data preprocessing and the application of boundary loss can obtain an improvement of 1% in terms of the Intersection over Union (IOU) of building detection. The post-process can take full advantage of probability information from the network, thereby effectively optimizing the building boundary. The post-process brings an improvement of 10.5% in terms of building instance recall compared with the results of the neural network. Our study achieves a building instance recall rate that is 22.9% higher than that of the original ICT-Net.A novel building detection method based on a boundary-regulated network and watershed segmentation is proposed in this study. Experimental results reveal the advantages of the enhanced-boundary-oriented data preprocessing and modified neural network and demonstrate that the proposed method can further improve prediction accuracy on a network basis. However, the excellent performance of the proposed method largely depends on parameter selection, and further improvements should be made in the future.

remote sensing;building detection;deep neural network;boundary loss;watershed segmentation;instance segmentation