| 摘要: | 傳統的雨量預測通常仰賴地面氣象儀器，如雨量計和氣象雷達，然而這些設備的部署與維護成本高昂，且在地理覆蓋上存在局限性。相較之下，若能使用氣象衛星，如向日葵8號提供了全面且連續的日本、東亞、西太平洋地區觀測數據，作為雨量預測數據，則能大幅減少雨量預測上的成本。
本研究的目標是探索僅使用向日葵8號氣象衛星的遙測數據，結合卷積長短期記憶網路(Convolutional LSTM，後稱CLSTM)和TransUNet架構，進行台灣地區的雨量預測，評估其準確性並衡量其實用性。
本研究流程大致分為三步驟，數據收集與預處理、模型構建、模型訓練與評估。將向日葵8號氣象衛星影像數據進行歸一化的預處理，並分割成大量的圖像塊。設計一個卷積長短期記憶網路(CLSTM)模型捕捉衛星影像數據中的時間依賴性，結合TransUNet架構，提取和強化數據的層次特徵以提高雨量預測的準確度。訓練上使用交叉驗證技術(Cross-Validation)和交叉熵損失函數(Cross-Entropy Loss)進行權重的調整與優化，測試上則是以準確度(Accuracy)和系統預測的物件區域與真實物件區域的交集除上這兩個區域的聯集(Intersection over Union)指標作爲評估方法。根據實驗結果顯示，結合CLSTM和TransUNet的模型能夠準確的預測台灣區域的雨量級別，並在極端的降雨事件也有良好的表現。
本研究證明了僅使用向日葵8號衛星影像數據，結合CLSTM和TransUNet架構進行台灣地區雨量級別預測的可行性。此方法不僅減少了對昂貴的地面儀器的依賴，也提升了雨量級別預測的空間覆蓋範圍。未來研究可以更進一步優化模型，縮短訓練模型上所需花費的計算成本與時間，並探索其在其他區域和不同氣候條件下的應用潛力。;Conventional methods of precipitation forecast usually relies on ground-based meteorological instruments, such as rain gauges and weather radars. However, these devices are too expensive to deploy and maintain, and have limitations in terms of geographic coverage. In contrast, the use of meteorological satellites, such as Himawari 8, which provides comprehensive and continuous observations of Japan, East Asia, and the Western Pacific as rainfall prediction data, can significantly reduce the cost of rainfall prediction.
The objective of this study is to explore the accuracy and utility of rainfall prediction in Taiwan using only the remote sensing data from Himawari-8 in combining the Convolutional LSTM (later called CLSTM) and TransUNet architectures.
The process of this study is divided into three steps: data collection and preprocessing, model construction, and model training and evaluation. The image data of the Himawari-8 meteorological satellite were normalized and preprocessed and divided into a large number of patches. A CLSTM model was designed to capture the time dependence in patches of satellite, and combined with the TransUNet architecture, the hierarchical features of the data were extracted and enhanced to improve the accuracy of rainfall prediction. Cross-Validation and Cross-Entropy Loss are used for training to adjust and optimize the weights. The evaluation method is based on the Accuracy and the Intersection over Union. Based on the experimental results, the model combining CLSTM and TransUNet is able to accurately predict the rainfall levels in the Taiwan region, and also performs well in extreme rainfall events.
This study demonstrates the feasibility of combining the CLSTM and TransUNet architecture for rainfall level prediction in Taiwan region using only Himawari-8 satellite image data. This method not only reduces the reliance on expensive ground-based instruments, but also improves the temporal coverage of rainfall level prediction. Future research can further optimize the model, reduce the computational cost and time spent on training the model, and explore its potential application in other regions and under different climatic conditions. |
