Abstract:
Content-based remote sensing image retrieval (CBRSIR) is one of the important techniques for the mining and analysis of big remote sensing data. Recently, unified-source ...Show MoreMetadata
Abstract:
Content-based remote sensing image retrieval (CBRSIR) is one of the important techniques for the mining and analysis of big remote sensing data. Recently, unified-source CBRSIR (US-CBRSIR) has been extensively studied and explored, but there has been little attention on cross-source CBRSIR (CS-CBRSIR). Although there is motivation for exploration of CS-CBRSIR for the continually increasing multisource remote sensing data, CS-CBRSIR suffers data drift due to multisource data. In this article, we explore to explicitly address the problem by mapping the source domain to target domain and propose an image translation-based framework for CS-CBRSIR. On the one hand, a novel cycle-identity-generative adversarial network (CI-GAN) is proposed based on the cycle-GAN. In addition to the generator and discriminator, a pretrained classifier, identity module, is designed to further boost the discriminative ability of translated images and facilitate the implementation of feature extraction and similarity measure. On the other hand, to alleviate the impact of style difference between the generated and real images, translated image augmentations and label smoothing regularization (LSR) are adopted to enhance training and contribute toward generation of a robust feature extractor. Extensive experiments on a public data set and a detailed ablation study confirm the effectiveness of our approach.
Published in: IEEE Transactions on Geoscience and Remote Sensing ( Volume: 58, Issue: 7, July 2020)