本篇論文使用基於transformer的深度學習模型TSViT(Temporal Spatial Vision Transformer[3])以及其他深度學習模型，以三種面向探討在只使用向日葵8號的衛星光譜影像的情況下，降雨量預測之可行性。資料為2017年6月其中十天之向日葵8號衛星影像以及交通部中央氣象局劇烈天氣監測系統QPESUMS產品之歷史一小時降雨資料。 由於現今的降雨預測產品大多需要許多的資料做為產品內預測模型的輸入，包含但不限於各類氣象衛星影像、地表雷達回波資料、各地區氣象站所蒐集的資料以及歷史資料等等，並且需要多位專業氣象人員依據不同的天氣狀況選擇合適的降雨預測產品和模型輸入，整個流程耗時費力。如能單只使用氣象衛星光譜影像就可獲得相對準確的降雨量，將會節省許多人物力於此。 本文依據手上的資料做了三種實驗(1)一般化模型(2)近即時模型-相同空間相近時間以及(3)近即時模型- 相同時間相鄰空間。就實驗結果而言，僅(2)具較高之可行性。;This study investigates the feasibility of rainfall prediction using only Himawari-8 satellite spectral imagery. The research explores the potential of the deep learning model TSViT (Temporal Spatial Vision Transformer[3]) along with other deep learning models in three different aspects. The utilized dataset consists of Himawari-8 satellite images and corresponding ground-based radar-detected rainfall distribution maps from ten days in June 2017, obtained from the Central Weather Bureau of Taiwan. Current rainfall prediction products typically require multiple data sources as inputs for the prediction models, including various meteorological satellite images, ground-based radar data, data collected from meteorological stations, and historical data. These products also rely on the expertise of meteorological professionals to select suitable rainfall prediction products and model inputs based on different weather conditions, resulting in a time-consuming and labor-intensive process. If satisfactory rainfall predictions can be achieved using only meteorological satellite spectral imagery, it would significantly save human resources. In this study, three experiments were conducted: (1) a generalized model, (2) a near real-time model with spatial proximity and similar time, and (3) a near real-time model with temporal proximity and adjacent space. The experimental results indicate that only the near real-time model with spatial proximity and similar time (Experiment 2) shows higher feasibility in rainfall prediction.
