An Object-Level Multi-Source Transfer Learning Method Integrating Optical and SAR Features: A Case Study of Gaofen, Ziyuan, and Sentinel-1 Satellites

In remote sensing earth observation applications, there is often a trade-off between automated processing and improved observation accuracy. A key issue in this contradiction is the underutilization of rich knowledge embedded in historical remote sensing data. To address this, this paper proposes an object-level multi-source transfer learning method (OMSTL) that effectively leverages the knowledge in historical remote sensing data for unsupervised land object discrimination. This method effectively integrates multi-source heterogeneous features and represents land object knowledge, enabling reliable knowledge transfer. Multi-source transfer learning utilizes knowledge from remote sensing images at different times and locations, significantly enhancing transfer stability. The method's effectiveness and potential were demonstrated through experiments using five optical remote sensing images (GF-1B, GF-1D, GF-6, GF-7, ZY-3) and their corresponding SAR images (Sentinel-1).