This paper describes a deep learning approach to semantic segmentation of very high resolution remote sensingimages. We introduce RLFCN, a fully convolutional architecture based on residual logic blocks, to model theambiguous mapping between remote sensing images and classification maps. In order to recover the outputresolution to the original size, we adopt a special way to eciently learn feature map up-sampling within thenetwork. For optimization, we employ the equally-weighted focal loss which is particularly suitable for the taskfor it reduces the impact of class imbalance. Our framework consists of only one single architecture which istrained end-to-end and doesn't rely on any post-processing techniques and needs no extra data except images.Based on our framework, we conducted experiments on a ISPRS dataset: Vaihingen. The results indicate thatour framework achieves better performance than the current state of the art, while containing fewer parametersand requires fewer training data.
展开▼