Unsupervised domain adaptive classification for hyperspectral remote sensing by adversary coupled with distillation

Unsupervised Domain Adaptive (UDA) classification aims to categorize the target domain scenes without labeled samples using knowledge from the source domain data with the labeled samples. Thus, UDA classification is one of the important cross-scene classification methods in the field of Hyperspectral Image Classification (HSIC). The existing domain adaptive classification methods for hyperspectral remote sensing data mainly utilize the adversarial training mode to achieve the feature alignment between the target and source domains. The popular UDA approach with the local alignment condition of dual domains generates acceptable classification accuracy. However, the key issue of the sufficient transfer of the source domain knowledge to the target domain is not considered. An unsupervised domain adaptation classification by adversary coupled with distillation is proposed in this paper for the unsupervised HSIC to effectively extract and transfer source domain knowledge. In the proposed framework, the dense-base network with convolutional block attention module is presented to extract abundant features for the representation of the source and target domain categories. In the source domain training, a self-distillation learning schema is adopted to reduce the class-wised difference by matching the predictive distribution of the same class samples. The self-distillation regularization constraint is increased between the samples of the same category in the source domain to reduce the intraclass difference of the classification subspace and improve the knowledge expression accuracy of the source domain classification model. Thus, the capability of the adaptive classification model to refine the source domain supervision knowledge is improved. In addition, a novel mechanism of adversarial training coupled with distillation knowledge is presented to guarantee the complete transfer of source domain knowledge to the target domain scene with feature alignment. Moreover, dual classifiers are employed in the adversarial training process to eliminate the prediction effect of the confused samples. The maximum and minimum discrepancies of the dual classifiers during the adversarial training rapidly promote the feature alignment without confusion. Thus, knowledge distillation is conducted to improve the recognition capability of the network in the domain while ensuring the complete transfer of hyperspectral source domain knowledge in the feature alignment process to improve the knowledge acquisition capability of the model in the target domain. Finally, the unsupervised classification of HSIs in the target domain is completed after the knowledge transfer. The experiments for HSI cross-scene image classification are conducted on four hyperspectral remote sensing scene datasets, including Pavia University, Pavia Center, Houston 2013, and Houston 2018. Results demonstrate that the proposed model is superior to other hyperspectral domain adaptive methods. Under the same sample conditions, the classification accuracy achieves 91.75% (Pavia University to Pavia Center), 74.41% (Pavia Center to Pavia University), 70.68% (Houston 2013 to Houston 2018), and 67.76% (Houston 2018 to Houston 2013). In addition, the ablation study illustrates that the final classification accuracy of the unsupervised HSIC is improved with the self-distillation and the distillation loss in the adversarial training model. The parameters with different weights and temperatures are analyzed in the experiments with variations of values. The validity of the method is verified by all the mentioned experimental results and analyses. © 2024 Science Press. All rights reserved.