RemoteCLIP-AF: a multimodal framework for enhanced remote sensing image processing and image-text retrieval

The rapid advancement of remote sensing technology has enabled high-resolution image acquisition of the Earth’s surface, essential for applications like environmental monitoring and urban planning. However, these images often suffer from quality degradation due to varying lighting conditions and sensor noise, complicating their effective use. This paper presents RemoteCLIP-AF, a multimodal framework that integrates a spatial-frequency joint feature enhancement paradigm to improve remote sensing image quality and facilitate image-text retrieval. By employing an improved Visual Transformer (ViT-AF) and a dual-domain feature fusion approach, RemoteCLIP-AF enhances image representation robustness against multiple distortions. Experimental results demonstrate significant improvements in retrieval accuracy across benchmark datasets (RSITMD, RSICD, and UCM), highlighting the model's potential in remote sensing applications and multimodal learning.