Hierarchical Visual Adaptation and Textual Inversion for Enhanced Remote Sensing Image-Text Retrieval

Image-text retrieval in remote sensing has become an important research area due to the growing need for semantic understanding of aerial images and their corresponding textual descriptions. However, establishing a strong alignment between these two modalities remains a significant challenge. Most existing models are not easily extended to the remote sensing domain and often fail to capture fine-grained details, spatial distributions, and domain-specific semantics. To overcome these limitations, a dual-enhancement framework, CLIP with extended textual and image adaptation (CLIP-XTI), is proposed to enrich mid-level visual representations by inserting a Hierarchical Feature Adapter (HFA) at selected intermediate stages of the vision transformer to refine visual representations and by incorporating image-guided pseudo-token embeddings into textual descriptions to enhance textual representations. By preserving the pretrained CLIP backbone and incorporating lightweight trainable adaptation modules, the proposed framework adapts CLIP to the remote sensing domain while maintaining competitive performance on widely used RSITR benchmarks. The results demonstrate that enriched textual embeddings and adaptive feature alignment effectively reduce the cross-modal gap in remote sensing retrieval.