Res2-Unet, a New Deep Architecture for Building Detection From High Spatial Resolution Images

Accurate large-scale building detection is significant in monitoring urban development, map updating, change detection, and digital city establishment. However, due to the complicated details of background objects in high spatial resolution remotely sensed images, the models proposed in building detection are still not performing satisfactorily. Particularly, such issue lies in the small buildings, which are easily to be omitted, and the pixels in the bounding area of each building instance can be especially confusing with the background objects. Aiming to deal with such problem, we propose Res2-Unet to employ multi-scale learning at a granular level, rather than the commonly used layer-wise feature learning, to enlarge the scale of receptive fields of each bottleneck layer. It replaces the widely used 3
$ \times $
3 convolution on n-channel feature maps with a set of smaller groups, which are organized in a hierarchical structure to enlarge the scale-variability. The general framework is an end-to-end learning network, taking a typical semantic segmentation network structure with encoders to encode the input image into feature maps and decoders to decode the feature maps into binary segmented result image. Moreover, to enhance the building boundary generation ability of our model, a boundary loss function is proposed to improve the detection performance. The proposed framework is evaluated on three public datasets, Massachusetts building dataset, WHU East Asia Satellite dataset and WHU Aerial building dataset. It is compared with the published performances and has achieved the state-of-the-art accuracies. That verifies the robustness of the proposed framework.