Deep learning in remote sensing image matching: A survey

Abstract

Deep learning demonstrates significant potential in enhancing the techniques of remote sensing image (RSI) matching. The current review delves into the incorporation of deep learning in RSI matching methods. Four predominant strategies are elucidated: area-based matching, feature-based matching, regression-based matching, and unsupervised learning methods. Area-based strategies concentrate on the quantification of similarity among image regions through sophisticated deep networks. Conversely, feature-based strategies are designed to detect, describe, and correspond salient features via comprehensive end-to-end networks. Regression-based matching methods leverage labeled data to train networks to identify correspondences. Unsupervised methods directly learn matching transformations in an end-to-end manner without labels. For each approach, representative methods, network architectures, loss functions, and modules are analyzed. Current challenges and future directions are provided, including needs for unified datasets, cross-modal loss functions, and end-to-end matching networks. This review offers researchers and practitioners systematic insights into deep learning advances for RSI matching. The discussion of methods, techniques, and research directions provides valuable reference for future research and application development in this important area.