Abstract
Remote sensing (RS) image retrieval is a crucial task for managing and exploiting the ever-expanding archives of Earth observation data. Hashing has been widely used for large-scale RS image retrieval due to its outstanding advantages in storage efficiency and search speed. Recently, existing supervised deep hashing methods have achieved remarkable progress, yet their reliance on large-scale, manual annotations limits their application in practice. To address this limitation, we propose self-distilled self-supervised hashing (SDSH), a label-free framework that learns compact binary hash codes directly from unlabeled RS images. To fully exploit the discriminative visual patterns of RS images across different scales, SDSH employs a teacherβstudent self-distillation strategy across global and local views. Meanwhile, an attention-based weighted pooling (AWP) module is introduced to address the issue of dispersed key information caused by large-scale scenes with multiple targets in RS images by highlighting informative regions. Extensive experiments reveal that SDSH not only sets a new state of the art (SOTA) in unsupervised RS hashing but also rivals supervised counterparts, demonstrating the potential of SDSH to provide an efficient and scalable paradigm for the RS image retrieval task.