An Efficient Annotation Method for Big Data Sets of High-Resolution Earth Observation Images

High-resolution earth observation images have increased dramatically because of the increasing of remote sensing satellites. Researchers must do large-scale target annotations to meet the training needs of deep neural network based model. However, most existing datasets contain an insufficient number of annotated samples, due to the inefficient manual annotation process which reason lies in the large number of remote sensing images, huge size, numerous targets, and high accuracy requirements. This paper proposed an efficient annotation method for big data sets of high-resolution earth observation images, in which the annotation process is divided into two parallel sub-processes, fast panchromatic image labeling and multi-spectral image fusion. Automatic scale transform is utilized for annotation of fused imagery. Experimental results show that the proposed method could improve the accuracy and efficiency of target labeling. Mask-RCNN and Faster-RCNN based target detection results demonstrate the validity of the big dataset annotated via our method.