Transfer Learning for Optical and SAR Data Correspondence Identification with Limited Training Labels

Abstract

Recent advancements in sensor technology have reflected promise in collaborative utilization; specifically, multisource remote sensing data correspondence identification attracts increasing attention. In this article, a domain-transfer learning based generative correspondence analysis (DT-GCA) scheme is proposed, which enables identifying corresponding data in optical and synthetic aperture radar (SAR) images with small-sized reference data. In the proposed architecture, an adversarial domain-translator is investigated as general-purpose domain transference solution to learn cross domain features. The optical-aided implicit representation, which is regarded as the clone of SAR, is adopted to estimate the correlation with SAR images. Particularly, the designed GCA integrates optical-generated features with SAR tightly instead of treating them separately and eliminates the discrepancy influence of different sensors. Experiments on cross-domain remote sensing data are validated, and extensive results demonstrate that the proposed DT-GCA yields substantial improvements over some state-of-the-art techniques when only limited training samples are available.