Open-Set Domain Adaptation for Hyperspectral Image Classification Based on Reconstruction Discrepancy and Feature Alignment

For hyperspectral image classification, domain adaptation algorithms often assume that the source domain and target domain share the same label space, and thus classify all samples in the target domain as one of the classes in the source domain. In real-world settings, the target image may contain unknown classes that have not been observed in the source image. Open set domain adaptation (OSDA) aims to classify the target samples as known source classes or unknown target-specific classes, thereby achieving more accurate and complete classification results. In this paper, we propose a new OSDA method based on reconstruction discrepancy and feature alignment (RDFA). Identifying unknown target samples requires measuring their similarity to the source domain. RDFA achieves this purpose by employing a reconstruction module trained on the source data. The reconstruction discrepancy of target samples can be used to calculate the similarity weights, which indicate how likely they belong to the known classes from the source domain. The similarity weights are then assigned to the target samples to train an extended classifier capable of predicting unknown classes. To classify the known classes in the target domain, we adapt both domain-level adversarial adaptation and class-level centroid alignment strategies to open set scenario to obtain domain-invariant features of the known classes. Therefore, the classifier trained on the source domain become suitable for known classes of target domain. Experiments with multiple cross-domain hyperspectral remote sensing images demonstrate the effectiveness of the proposed approach.