Updating Land-Cover Maps by Iterative Difference Learning Network

Abstract

Multi-temporal remote sensing image classification aims to exploit the available information of image in the source domain to classify target domain. Since the manual labeling is time-consuming and labor-intensive, it is unrealistic to have enough labels for all images of the time series. By analyzing the difference information of multi-temporal images, the labels of unchanged region can be transferred form source domain to the target domain. In order to further utilize the difference information and learn a robust classifier, we propose an iterative difference learning network (IDLnet) to update land-cover maps in this paper. The proposed method aims at optimizing the process of label transfer by analyzing results of classifier and to fine-tuning it with a series of dynamic training sets. In proposed method, we first utilize the source domain data to initialize a training set and train a classifier to classify both the source and target domains. The change detection (CD) is applied on the ground image datasets and the classification result. Then the transfer learning (TL) is employed to transfer the unchanged information to fine-tuning network. We detect changes of the classification result images again and fuse the previous CD results. Finally, the accuracy cannot be improved after several iterations of fine-tuning the classifier.