How can we retrieve past weather cases that resemble a specific region in a satellite image, without cropping and recomputing embeddings? Analyzing similar cases is a key approach to improving weather forecasting accuracy. Especially, local atmospheri...
How can we retrieve past weather cases that resemble a specific region in a satellite image, without cropping and recomputing embeddings? Analyzing similar cases is a key approach to improving weather forecasting accuracy. Especially, local atmospheric patterns---such as cloud formations or storm centers---often determine event dynamics, making region-specific search crucial. To improve efficiency, forecasting agencies typically store satellite images as embeddings generated by vision models. However, conventional vision models are not specialized for capturing meteorological patterns and typically generate a single embedding per image, making region-specific search inefficient at scale. To address this, we propose \textbf{RaP-Weather}, a Region-aware Patch-level embedding and retrieval framework that enables patch-level similarity search using precomputed features. Our method modifies a Vision Transformer (ViT) backbone to retain spatial information and fine-tune the model with a triplet loss to capture meteorological similarities. Experiments on a large-scale satellite dataset demonstrate that our approach achieves accurate and efficient region-focused retrieval without additional forward passes.