Open-Set Black-Box Domain Adaptation for Remote Sensing Image Scene Classification

Abstract

Domain adaptation (DA) has recently made tremendous progress in remote sensing image scene classification. Particularly, open-set DA (OSDA) has attracted increasing attention, wherein the target domain includes unknown classes. However, existing OSDA methods assume that the source samples or the parameters of the source model are available, which is not practical due to concerns about digital privacy and portability issues. Addressing this, we investigate a more realistic and challenging open-set DA scenario for remote sensing image scene classification, where the unlabeled target domain is only provided with a black-box source predictor (i.e., only model predictions are accessible). To address this problem, we devise an Open-set Knowledge distillation framework with neighboRhood similarity regularization and uncertAinty modeling called OKRA. Specifically, we introduce a neighborhood similarity regularization to facilitate the open-set knowledge distillation (KD) using local neighborhood information. Furthermore, we propose an energy-based uncertainty modeling (UM) strategy for open-set recognition, which can effectively discriminate known and unknown target data without any thresholding. Empirical results on six cross-scene scenarios built from three datasets verify that OKRA is effective and practical for remote sensing image scene classification, outperforming existing data-dependent OSDA methods by a large margin.