Semi-automatic Annotation Method for Semantic Segmentation of Synthetic Aperture Radar Images

With the availability of public high-resolution remote sensing datasets, it has become a trend to realize land cover classification using deep learning methods. However, supervised semantic segmentation requires an accurate label at the pixel level as the ground truth of image. This is a challenge to remote sensing data with big data characteristics, and manual annotation is challenged with a heavy workload and low efficiency. In this paper, a semi-automatic annotation method is proposed and verified on synthetic aperture radar (SAR) images obtained by Gaofen-3 (GF-3) and the Chinese Aeronautic Remote Sensing System (CARSS). The method uses semantic segmentation to classify the homologous images initially, then iterates the coarse classification results as ground truth. The coarse classification results of multiple iterations will be corrected and added to the base dataset. The whole system is updated positively, and the model is constantly optimized to improve the performance of the auxiliary semi-automatic annotation model. Experimental results demonstrate that the pixel accuracy of the final classification results can reach higher than 80%.