Multi-branch regression network for building classification using remote sensing images

Convolutional neural networks (CNN) are widely used for processing high-resolution remote sensing images like segmentation or classification, and have been demonstrated excellent performance in recent years. In this paper, a novel classification framework based on segmentation method, called Multi-branch regression network (named as MBR-Net) is proposed. The proposed method can generate multiple losses rely on training images in different size of information. In addition, a complete training strategy for classifying remote sensing images, which can reduce the influence of uneven samples is also developed. Experimental results with Inrial aerial dataset demonstrate that the proposed framework can provide much better results compared to state-of-the-art U-Net and generate fine-grained prediction maps.