Latent Representation Generation for Efficient Content-Based Image Retrieval in Weather Satellite Images Using Self-Supervised Segmentation

With the rapid advancement of satellite sensor technology, a huge volume of high-resolution image data sets can now be acquired, and efficiently representing and recognizing scenes within these high-resolution images has emerged as a critical challenge. In this paper, we propose AES (Auto-Encoder after Segmentation), a method for learning a latent representation containing high-level semantics of meteorological satellite images. AES uses W-Net, an unsupervised segmentation model, to simplify the images. Our approach significantly improves CBIR performance in the domain of weather satellite images. Experimental results on satellite weather images demonstrate that our proposed method outperforms traditional auto-encoders in terms of mean squared error (MSE). A lower average MSE indicates stronger similarity between the retrieved and query images, suggesting that our method better represents the underlying similarities within the data. Additionally, our approach demonstrates higher recall values for similarity assessment and enhanced retrieval accuracy, showcasing its effectiveness in extracting meaningful features for content-based image retrieval in meteorological satellite images.