Class-Wise Adversarial Transfer Network For Remote Sensing Scene Classification

In this paper, we proposed a class-wise adversarial transfer network (CATN) for remote sensing scene classification. A class-wise discriminator is used to achieve conditional alignment on a per class basis. The CATN employs deep convolutional neural network to learn domain invariant classification layer features in a class-conditional manner. The classification probability output of the target data is utilized to determine the weights of the class adversarial losses. The proposed method can promote positive adaptation, and does not need labeled instances in the target domain. The experiments results using two remote scene image data sets with different resolution indicates its good performance.