| Extreme weather classification in high-latitude, data-sparse regions like Alaska is challenging due to rare events, severe class im-balance, and persistent gaps in ASOS station coverage. Single-modality models overfit to dominant classes, whereas prior multimodal methods often neglect the semantic richness of heterogeneous sources. To address this challenge, we propose a fusion framework combining: (i) hourly observations from 154 Alaskan ASOS stations, (ii) high-quality reference data retrieved from a central station, and (iii) semantic embeddings of LLM-generated narratives. The proposed lightweight pipeline converts 12-hour summary statistics into natural-language weather descriptions (via a local open-source LLM), encodes them with a sentence transformer, and then fuses with LSTM-extracted numeric features. Trained on 2018–2021 data and evaluated on 2022 (n = 730), the model achieves 91.8% accuracy and 0.87 macro-F1 across nine classes, with strong gains on rare events (e.g., extreme cold, or debris flow). This enabled retrospective labeling of 223,188 unlabeled windows, producing spatially consistent event maps. The obtained results provide evidence that this paradigm can effectively transform multimodal structured and unstructured signals into human-readable narratives, significantly improving tail-class recovery in sparse, observation-limited environments, thereby supporting retrospective climatological analysis and risk assessment |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.