基於LSTM結合降雨影像強度分類以進行雨量預測之研究;Rainfall Prediction via Integration of LSTM and Rainfall Image Intensity Classification

NCUIR > College of Electrical Engineering & Computer Science > Executive Master of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/98170

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98170

Title:	基於LSTM結合降雨影像強度分類以進行雨量預測之研究;Rainfall Prediction via Integration of LSTM and Rainfall Image Intensity Classification
Authors:	黃信愷;Huang, Hsin-Kai
Contributors:	資訊工程學系在職專班
Keywords:	雨量預測;異質資料整合;深度學習;LSTM;影像分類;Rainfall Prediction;Heterogeneous Data Integration;Deep Learning;LSTM;Image Classification
Date:	2025-06-25
Issue Date:	2025-10-17 12:27:07 (UTC+8)
Publisher:	國立中央大學
Abstract:	近年極端氣候加劇，臺灣因強降雨導致的水災日益頻繁，使防汛行動備受政府重視，其中「防汛熱點警戒」被視為重要的決策依據，其以雨量站累積雨量作為標準。為提升應變效率，現階段多採用中央氣象署提供之定量降水預報掌握未來累積雨量，可在警戒達標前提前預警爭取應變時間。然而定量降水預報主要以區域尺度呈現，未細化至個別雨量站，導致警戒標準難以直接套用，決策者仍須依經驗整合各項情資進行判斷。此外，現有雨量預測模型多半僅採用單一或同屬性資料進行推論，資料面向較為侷限，使預測能力受限。本研究模擬決策者整合異質資訊的思考方式，提出一個利用深度學習結合多源資料的雨量預測方法，基於LSTM輸入雨量站歷史降雨數據，捕捉降雨時序關聯後得出初估預測值，並結合鄰近CCTV之降雨影像進行強度分類以取得現況雨勢大小的資訊，分別得出階段性結果後再以校正權重模型綜合資訊得出最終預測值。實驗結果顯示所提出的方法在整體預測準確度上優於單一情資模型，平均MSE降低50~70.8%、MAE降低25.6~35.8%、R²提升25.4~45%，在極端強降雨(40mm以上)的情況下預測表現亦顯著改善，平均RMSE降低65.1~74.5%，證明異質情資整合對於短時累積雨量的預測具有正向發展，並展現自動化提前預警應用上的潛力。 ;In recent years, extreme rainfall events have caused increasingly frequent flood disasters in Taiwan, emphasizing the importance of accurate rainfall prediction for flood hotspot alerts. These alerts, based on cumulative rainfall at gauge stations, are a key reference for emergency response. To support early warning, regional-scale quantitative precipitation forecasts (QPF) from the Central Weather Administration are commonly used. However, the lack of station-level resolution limits direct application, requiring decision-makers to integrate information manually. Existing models also often rely on single-source data, constraining predictive performance. This study proposes a deep learning approach that mimics decision-making by integrating heterogeneous data. A Long Short-Term Memory (LSTM) model processes historical rainfall data to capture temporal patterns, while nearby closed-circuit television (CCTV) images are classified to assess current rainfall intensity. These intermediate outputs are then combined using a calibrated weighting model to produce the final prediction. Experiments show the proposed method significantly outperforms single-source models, reducing mean squared error (MSE) by 50–70.8%, mean absolute error (MAE) by 25.6–35.8%, and increasing R² by 25.4–45%. For extreme events (over 40 mm), root-mean-square error (RMSE) drops by 65.1–74.5%. These results demonstrate the potential of heterogeneous data integration for short-term rainfall forecasting and automated flood hotspot early warnings.
Appears in Collections:	[Executive Master of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	3	View/Open

社群 sharing

Loading...