Zero-Shot Remote Sensing Image Classification Algorithm Based on Improved CLIP Model

Zero-shot remote sensing image classification is of great significance in earth observation and environmental monitoring, as it can identify and classify new land cover types without labeled data, significantly reducing the cost of manual annotation. However, traditional methods often rely on subjective attribute definitions, limited semantic embeddings, complex generative model training, high-quality graph construction, and costly cross-modal learning, leading to performance limitations. This paper proposes an innovative approach based on the CLIP (Contrastive Language-Image Pre-training) model. The method first utilizes the CLIP model to make initial predictions on remote sensing images, generating pseudo-labels and constructing a cache. Then, it calculates the similarity between the test images and the images in the cache to obtain the classification probability of the test images. Finally, it combines the similarity results with the initial CLIP predictions in a weighted manner to form the final classification result. The experimental results on four typical remote sensing datasets demonstrate that the method proposed in this paper can improve the accuracy of zero-shot remote sensing image classification.