多模態機器學習架構於大台北地區劇烈降水預報之應用

Abstract

我們提出一個創新的多模態機器學習架構，將機器學習轉型為科學發現工具，並以大台北地區的劇烈降水預報作為測試。我們的研究整合了多種氣象資料，包括再分析資料、雷達資料、衛星影像和探空資料，透過四個已發表和發表中的研究，探討各種資料型態各自在降水預報中的應用與限制。 在個別資料模態研究方面，本論文指出機器學習結合再分析資料能有效預測鋒面、颱風和豪雨事件，其中豪雨預報命中率達78-83%。在雷達資料方面，基於深度學習的體到點（Volume-to-Point, VTP）架構在強降雨偵測上表現卓越，平均命中率達0.8，優於傳統方法。而針對衛星影像的表徵學習顯示，卷積自動編碼器（convolutional autoencoder, CAE）在多種天氣事件分類中表現最佳，但在豪雨預測方面受限於單一時間點資料的不足。探空資料研究則證明，低成本的storm tracker探空儀器經過校正後，能在低對流層提供可靠的高時間-空間解析度觀測，有助於理解與劇烈降水相關的深對流發展。 論文進一步測試了整合三種資料（再分析、雷達、衛星）的多模態預報架構，初步結果顯示，次日劇烈降水預報的命中率為0.71，且再分析資料在預報中顯示出較高的重要性 。總體而言，本研究強調了多模態機器學習在提升劇烈降水預報準確性方面的巨大潛力，並開闢了利用模型可解釋性進行科學探索的新途徑 。

This dissertation presents an innovative multimodal machine learning framework designed to enhance severe rainfall forecasting in the greater Taipei area, thereby transforming machine learning into a tool for scientific discovery. The research integrates diverse meteorological datasets, including reanalysis data, radar data, satellite imagery, and sounding data, exploring their applications and limitations in precipitation forecasting. In the individual data modality studies, the thesis demonstrates that machine learning combined with reanalysis data can effectively predict frontal, typhoon, and heavy rainfall events, achieving a hit rate of 78-83% for heavy rainfall forecasts. For radar data, the deep learning-based "volume-to-point" (VTP) architecture shows significant superiority in detecting heavy rainfall, with an average hit rate of 0.8, outperforming traditional methods. Representation learning from satellite imagery reveals that the Convolutional Autoencoder (CAE) performs best in classifying various weather events; however, its effectiveness in predicting heavy rainfall is limited by the relevancy between the daily infrared images and small-scale convective cells. The sounding data study confirms that, after calibration, the low-cost Storm Tracker radiosonde can provide reliable, high spatio-temporal resolution observations in the lower troposphere, aiding in the understanding of deep convection development associated with severe rainfall. The dissertation further tests a multimodal forecasting framework integrating three data types (reanalysis, radar, satellite). Preliminary results show a next-day severe rainfall prediction hit rate of 0.71, with reanalysis data demonstrating higher importance in the forecast. Overall, this research highlights the significant potential of multimodal machine learning in enhancing the accuracy of severe rainfall forecasting and opens new avenues for scientific exploration through model interpretability.