Benchmarking Pre-trained Models and Low-Rank Adaptation Fine-Tuning for Remote Sensing Image-Text Retrieval

Abstract

Cross-modal retrieval between remote sensing (RS) images and natural language links high-level semantic information across visual and textual modalities. Although Vision-Language Pre-training (VLP) models usually perform well in general domains, their effectiveness in RS scenarios remains limited by domain gaps such as nadir-view imaging, complex textures, and multi-scale objects. This paper constructs a reproducible retrieval benchmark based on the full RSICD dataset and evaluates the zero- shot performance of CLIP, BLIP-ITM, SigLIP, and RemoteCLIP. RemoteCLIP is then used as the backbone model, and Low-Rank Adaptation (LoRA) fine-tuning is applied to examine the effects of different ranks and injection scopes on retrieval performance. The results show that the early training stage can effectively learn spatial information. Dual-tower LoRA fine-tuning improves the test-set Mean Recall@10 to 55.21% while updating only 1.35 million parameters, and the results remain statistically stable across different runs. The t-SNE visualization and qualitative case studies indicate that LoRA improves the embedding distribution and reduces semantic distance between matched image-text pairs. The qualitative analysis also reveals limitations in distinguishing fine-grained functional entities, which provides useful references and directions for future cross- modal RS retrieval research.