Remote sensing object retrieval (RSOR) aims to identify images of the same object from large remote sensing image databases. However, existing RSOR methods often rely on simple and distribution-agnostic distance metrics, for example, Euclidean distance, that treat all the feature dimensions equally, making them susceptible to the influence of spurious features and vulnerable to variations such as noise, resolution degradation, and rotation variations. Inspired by the Neyman-Pearson theorem, we propose a generalized likelihood ratio test-based deep metric learning (GLRT-DML) approach for RSOR. GLRT-DML is a distribution-aware metric learning framework that leverages distribution information from feature embeddings across the dataset to suppress the influence of spurious features while emphasizing informative ones. This process facilitates the construction of a robust metric space for RSOR. Extensive experiments on ship, aircraft, and vehicle retrieval tasks demonstrate the superior performance of the proposed GLRT-DML approach compared with state-of-the-art RSOR methods.
Zhang et al. (Thu,) studied this question.