以圖像辨識深度學習進行即時淹水預測

Real-time Flood Inundation Forecasting Based on Image Recognition Deep Learning

李毅家(Yi-Chia Lee)

指導教授 ： 張倉榮

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202104223

摘要

水利署現行的淹水預警系統是以雨量站的即時觀測資訊提供淹水預警，但淹水警戒範圍多以村里為單位。若要獲得更詳細的淹水位置，可利用二維淹水模式，但模擬時間較長，難以在災害期間給予即時淹水資訊。為了解決此問題，本研究提出以圖像辨識之深度學習進行淹水預測之相關研究。
本研究首先使用時間-空間序率模擬所衍伸出的500場颱風事件為降雨資料庫，並以NTU-CAFIM進行淹水模擬，產生淹水資料庫，且資料庫將分為訓練組285場與測試組215場。接下來，分別考慮直接以真實樣本訓練並調整預測模型權重的ConvLSTM模型，與以條件式生成對抗網路進行訓練並調整權重的cWGAN模型，比較此兩個圖像辨識深度學習模型於淹水預測的能力。此兩個模型的輸入項為訓練組之降雨分佈圖、淹水深度分佈圖與其他淹水影響因子，輸出項為未來3小時淹水預測。模型預測之結果將以NTU-CAFIM模擬結果視為基準，透過評估淹水深度的均方誤差，進行ConvLSTM模型與cWGAN模型的預測能力比較，挑選出較適用於淹水預測的圖像辨識之深度學習模型後，再演示於測試組之颱風事件下該模型於事件的開頭與事件的中途開始預測等2種情況下的表現。
本研究以國立臺灣大學及周邊區域作為研究區域，結果顯示：(1)比較兩個模型之預測能力，ConvLSTM模型預測結果較cWGAN模型佳，ConvLSTM模型於測試組下，其預測未來1～3小時淹水深度分佈的均方誤差皆可低於3.00E-05；(2)應用ConvLSTM模型於測試組的1場颱風事件，無論是由事件的開頭開始進行直接預測與多步階預測，或是由事件的中途開始進行多步階預測，其預測結果之淹水範圍評估指標表現於大部分的時刻之預兆得分、敏感度與精確度可以維持在0.9以上，而均方誤差的表現在大部分時刻可低於1.00E-05。(3) ConvLSTM模型於每個時刻預測所花費的時間平均約為1.05秒，顯示本研究發展的淹水預測模型在淹水預警上可以提供即時且準確的預測。

關鍵字

淹水預測 ； 圖像辨識 ； 深度學習 ； 捲積長短期記憶神經網路 ； 條件式生成對抗網路

並列摘要

Currently, the warning system of flooding disaster established by Water Resources Agency raises warning regions in village level based on total rainfall amount. Two-dimensional inundation models can provide more accurate predictions of flooding areas despite the computational times of them are still too long for real-time flooding warning. To overcome this limitation, this study proposes a real-time flood inundation forecasting model based on image recognition deep learning.
In the research, two image recognition deep learning models are adopted. One is the ConvLSTM model trained directly with real samples; the other is the cWGAN model trained with the conditional Generative Adversarial Network. The spatial-temporal inputs of the two models are rainfall distribution maps, flooding inundation distribution maps and a flooding impact factor. The outputs of the two models are 1 to 3 hours ahead flooding inundation distribution maps. The database for training and validating the two models comprises of rainfall data built by the spatial-temporal stochastic simulation and flooding data by the NTU-CAFIM, and it is divided into the training group (285 events) and the test group (215 events). To evaluate the accuracy of the two models, mean square error (MSE) value is utilized to measure the difference in inundation depth between the two models and the NTU-CAFIM. One of the two models determined as the better model by MSE values and is latter used to show the hourly forecasting ability in a selected rainfall event in the test group.
The National Taiwan University and its surrounding area are adopted as the research area. The results of the model comparison indicate that the ConvLSTM model outperforms the cWGAN model in terms of 1 to 3 hours ahead forecasting. In the display part, three scenarios, i.e., the forecasting starts from the beginning of the events with conventional prediction or multi-step prediction, and from the midway of the events with multi-step prediction are also investigated. The MSE values (less than 1.00E-05 at most of the time) are all small enough so that the accuracy of the model on the hourly forecasting in the three scenarios is satisfying. Furthermore, the flooding areas of the ConvLSTM model have good agreement with those of the NTU-CAFIM in the three scenarios. As to the efficiency of the model, the average calculation time of the ConvLSTM model is about 1.05 seconds. In summary, it is proved that the proposed flood inundation forecasting model can provide instant and accurate prediction.

並列關鍵字

Flood forecasting ； Image recognition ； Deep learning ； ConvLSTM ； cWGAN

參考文獻

1. Arjovsky, M., Chintala, S., Bottou, L. (2017). Wasserstein generative dversarial networks. arXiv preprint arXiv:1701.07875.

2. Chang, L.-C., Shen, H.-Y., Chang, F.-J. (2014). Regional flood inundation owcast using hybrid SOM and dynamic neural networks. Journal of hydrology, 519, 476-489.

3. Chang, L.-C., Shen, H.-Y., Wang, Y.-F., Huang, J.-Y., Lin, Y.-T. (2010). Clustering-based hybrid inundation model for forecasting flood inundation depths. Journal of hydrology, 385(1-4), 257-268.

4. Chang, M.-J., Chang, H.-K., Chen, Y.-C., Lin, G.-F., Chen, P.-A., Lai, J.-S., Tan, Y.-C. (2018). A support vector machine forecasting model for typhoon flood inundation mapping and early flood warning systems. Water, 10(12), 1734.

5. Dottori, F., Todini, E. (2010). A 2D flood inundation model based on cellular automata approach. XVIII international conference on water resources.