遥感对地观测中普遍存在多平台、多传感器、多角度的多源数据,为遥感场景解译提供协同互补信息。然而,现有的场景解译方法需要根据不同遥感场景设计训练模型,或者对测试数据标准化以适应现有模型,训练成本高响应周期长,已无法适应多源数据协同解译的新阶段。跨域遥感场景解译将已训练的老模型迁移到新的应用场景,通过模型复用以适应不同场景变化,利用已有领域的知识来解决未知领域问题。本报告以跨域遥感场景解译为主线,综合分析国内外文献,结合场景识别和目标识别两个典型任务,论述近年来的国内外研究现状、前沿热点和未来趋势,梳理总结跨域解译的常用数据集和统一的实验设置。

Multi-source data with multiple platforms, sensors, and perspectives are common in remote sensing earth observation, providing synergistic and complementary information for remote sensing scene interpretation. However, the existing scene interpretation methods need to design training models according to different remote sensing scenes or standardize the test data to fit the existing models, with high training cost and long response period, which are no longer able to adapt to the new stage of collaborative interpretation of multi-source data. Cross-domain remote sensing scene interpretation migrates old trained models to new application scenarios, adapts to different scene changes through model reuse, and utilizes existing domain knowledge to solve unknown domain problems. This report focuses on cross-domain remote sensing scene interpretation, comprehensively analyzes domestic and international literature, discusses the current status of domestic and international researches in recent years, cutting-edge hotspots and future trends, and summarizes the commonly used datasets and unified experimental setups for cross-domain interpretation in combination with the two typical tasks of scene identification and target recognition.