Multidomain Adaptive Unsupervised Learning for Cross-Scene Hyperspectral Image Classification Based on a Generative Network

Abstract

Cross-scene hyperspectral image classification has achieved favorable outcomes in the domain adaptation of deep learning. However, transferring the sample features learned from the source domain (SD) to the target domain (TD) remains challenging due to the interdomain differences. To address this issue, we propose a multidomain adaptive unsupervised learning algorithm (MDAULA) based on a generative network, which comprises three core components. First, the architecture consists of two branches, representing SD and TD, respectively. Each branch is embedded with a generation module based on contrastive learning in order to form the generation domain (GD). GD is composed of the spectral and spatial features obtained by decoupling the original data (SD or TD). GD is fused with the original data of the corresponding branch to form a mixed domain (MD). GD and MD are designed to minimize the differences of domain adaptation during the prediction process. Thus, GD, MD, the SD, and the TD are simultaneously learned, enabling multidomain integration. Second, we introduce a novel domain adaptive method for multilevel feature correlation alignment loss. It can orderly and hierarchically align the means and covariances of the features in the source and TD, explicitly model the fine-grained distribution differences between feature channels, so that the model can balance the importance of features at different levels adaptively. Finally, we design an evaluation strategy for the most uncertain samples in the TD. By calculating the entropy of these samples, we identify difficult samples and assign corresponding weights to further train and fine-tune the model. To validate the effectiveness of the proposed method, we conducted a series of comparative and ablation experiments. Experimental results on three cross-scene datasets demonstrate that MDAULA outperforms existing domain adaptation methods.