DARAD: Dual Adapters and Ranking-Aware Distillation for Continual Remote Sensing Image-Text Retrieval

2026-08-06

AI Summary

This work addresses the challenge of continual remote sensing image-text retrieval, where scale variations and distribution shifts distort the cross-modal alignment space, severely degrading the performance of existing continual learning methods. To mitigate these issues, the authors propose the DARAD framework, which employs a spatial fusion adapter to handle visual scale changes and a multi-expert semantic routing text adapter to suppress embedding drift. Furthermore, a bidirectional ranking-aware knowledge distillation mechanism is introduced, leveraging a frozen teacher model and historical anchor points to effectively preserve prior cross-modal ranking structures. Experimental results demonstrate that DARAD significantly outperforms state-of-the-art methods under multi-stage continual retrieval protocols, achieving a strong balance between adaptation to new tasks and retention of historical retrieval performance.

Abstract

With the rapid growth of Earth observation technologies, remote sensing archives are rapidly expanding, making remote sensing image-text retrieval (RS-ITR) increasingly important. However, continual RS-ITR remains challenging because scale variation and distribution shifts in RS aggravate cross-modal alignment space distortion, making it difficult for existing continual learning (CL) methods to support reliable continual retrieval. To address this challenge, we propose DARAD, a dual-adapter and ranking-aware distillation framework that preserves the historical cross-modal ranking structure while learning new visual and textual concepts from evolving archives. Specifically, the visual branch introduces a spatial fusion adapter, which integrates coarse regional cues and fine-grained patch cues to accommodate RS scale variation while anchoring visual updates to the pretrained alignment space. The textual branch employs multi-expert semantic routing, which separates shared textual semantics from semantically specialized residuals to absorb newly emerging descriptions while constraining global text embedding drift. Furthermore, bidirectional ranking distillation uses a frozen teacher model and historical anchors to preserve the historical cross-modal ranking structure, thereby mitigating alignment space distortion across continual stages. Experiments under a multi-stage continual retrieval protocol show that DARAD achieves superior performance over existing CL methods, improving adaptation to newly arrived data while maintaining effectiveness on historical data.

Problem

• continual learning
• remote sensing image-text retrieval
• cross-modal alignment
• distribution shift
• scale variation

Innovation

• dual adapters
• ranking-aware distillation
• continual learning
• remote sensing image-text retrieval
• cross-modal alignment