SCAL: A Semantic-Consistent Adaptive Alignment Learning Framework for UAV Remote Sensing Cross-Modal Retrieval

The remote sensing cross-modal retrieval task aims to bridge the modality gap between visual data and textual descriptions. Among various remote sensing platforms, unmanned aerial vehicles (UAVs) have emerged as a prominent paradigm due to their flexible deployment and fine-grained observation capability, making UAV image-text retrieval a particularly important yet challenging setting. However, existing methods still suffer from significant limitations in UAV scenarios: on the one hand, most approaches rely on coarse-grained cross-modal alignment, and on the other hand, they are easily affected by background-dominated visual features. These issues hinder the establishment of reliable semantic associations between UAV images and text descriptions. To overcome these limitations, we propose a Semantic-Consistent Adaptive Alignment Learning (SCAL) framework that integrates two complementary modules: Confidence-Scaled Text-Conditioned Local Alignment (CSTLA) module and Contextual Semantic Alignment with Background Suppression (CSABG) module. Specifically, the CSTLA module refines fine-grained cross-modal interactions by performing textconditioned attention weighting and confidence-adaptive scaling, while the CSABG module maintains cross-modal consistency by suppressing background-dominated samples and transferring semantic relational structures from an EMA-based teacher model. Extensive experiments on two UAV image-text benchmark datasets demonstrate that the proposed SCAL framework achieves state-of-the-art retrieval performance, significantly enhancing fine-grained cross-modal alignment and overall semantic consistency.