Attention and classifier-constrained -based deep adversarial domain adaptive method for remote sensing image classification

To address the high cost problem of manual annotation of remote sensing data and negative migration caused by the feature distribution discrepancy between different domains, in this paper, a novel deep adversarial domain adaptive method based on attention and classifier-constrained strategy for remote sensing image classification is proposed. Firstly, for the various migratability of different image regions, and the corresponding low migratability regions will cause negative migration during the training process, a new adversarial method based on mixed attention mechanism is given so that the network can learn which parts need to be paid attention automatically during the migration process; Secondly, due to the fact that the difference of the classes spatial distribution between the source domain and the target domain, an adaptive metric module are adding to the adversarial domain adaptation model, which measures the distance between source domain and target domain data by the maximum mean difference of the multiple kernels, furtherly, attempt to align the feature distributions of the two domains on the basis of the adversarial domain adaptation model thereby improving the classification performance of the model. Lastly, to address the problem that remote sensing sample data is difficult to obtain and is often a subset of the actual application scenarios, which leads to unable to identify the new labels and poor generalization ability, we introduce the maximum classifier difference structure to adapt the cross-domain edge distributions and to emphasize the each domain's respective characteristics importance simultaneously. A series of extensive experiment results based on the UC Merced dataset, AID dataset and the NWPU-RESISC45 dataset are conducted to show that the proposed approach in this paper effectively improves the classification performance can be comparable with the methods of the state-of-the-art.