An Unsupervised Domain Adaptation Method for Multi-Modal Remote Sensing Image Classification

Labeling remote sensing data for classification is costly and time-consuming in practical applications, while sufficient and representative labels are critical for achieving a high accuracy. Transfer learning emerges as an effective method for this issue by reusing samples from other domains. In this paper, we propose an unsupervised domain adaptation method which can align the marginal distribution and conditional distribution in source and target domain at the same time. Our method treats the importance of the marginal and conditional distribution discrepancies at different levels and maps the feature sets of source domain and target domain into Reproducing Kernel Hilbert Space (RKHS) to obtain similar feature sets. In particular, we apply the proposed method on the multi-modal remote sensing data including pixel-wise overlaid Orthophoto and Digital Surface Models (DSM). With experiments containing images of different cities with highly distinguishable land-cover patterns as source and target domain, we demonstrate that, as compared to several state-of-the-art domain adaptation (DA) algorithms, our method can achieve a satisfactory performance on the target domain by a simple statistical classifier trained only by samples in the source domain.