人工智慧技術於淹水災害分區及水位預報之研究

Flood hazard zoning and water-level forecasting using artificial intelligence techniques

王志煌(Jhih-Huang Wang)

指導教授 ： 林國峰

國立臺灣大學/工學院/土木工程學研究所/博士(2020年)

https://doi.org/10.6342/NTU202002605

摘要

洪水為世界上最具破壞力的自然災害之一，為了降低洪水所帶來的危害，發展淹水災害分區和水位預報模式在災害預警系統中是相當重要一環。人工智慧方法於模擬水文過程的潛力已經被許多研究所肯定。然而，過去文獻大部多使用人工智慧建立淹水潛勢模式，鮮少應用人工智慧方法並考慮周圍環境淹水潛勢以進行淹水災害分區評估。此外，過去傳統的人工智慧方法屬於淺層機器學習，例如，人工神經網絡、支援向量機和適應性網路模糊推論系統等，這些方法有著無法從原始數據資料當中提取有效特徵的問題。因此，本論文之目的為發展新型之淹水災害分區模式和水位預報模式來改良傳統模式之缺點。本論文內容將分成兩部份來展現所提出之模式優點。
於論文第一部份，本研究提出以結合隨機森林與自組織映射圖，建立一個新型淹水災害分區模式，以產生淹水災害風險分區地圖。所建立之模式包含兩個模組：淹水潛勢分析及淹水災害分區。首先，以隨機森林為基礎建構淹水潛勢分析模組，產生淹水潛勢地圖，接著依據淹水潛勢地圖中每個網格的淹水潛勢值，以自組織映射圖建構淹水災害分區模組，將地圖中每個淹水潛勢值進行分類，最後產生淹水災害風險分區地圖。另外，本研究也考慮了兩種不同自組織映射圖的輸入項，一種為只採用自身網格的潛勢值，另一種為採用自身及周圍網格的潛勢值。為了證明所建立之模式的改善效率，本研究與傳統用於淹水災害分區的自然斷點法進行比較。本研究以台灣宜蘭的蘭陽平原作為實際應用，以呈現所建立之模式的優越性。應用結果顯示，採用自身及周圍網格的潛勢值建立之淹水災害分區模式可改善災害風險劃分之準確度，且本研究建立之模式也比傳統模式更具有優勢及合理性。
於論文第二部份，本研究以擴展序列卷積神經網路為基礎，發展一個新型之水位預報模式，以改善時水位預報準確度。擴展序列卷積神經網路可有效地廣泛學習時間序列的歷史資料，且採用殘留層與捷徑連結方式使網路架構能夠更深層及強健，加速訓練之收斂速度。本研究以台灣東北部宜蘭河流域作為實際應用，以呈現所發展模式之優點，並將所發展之模式與分別以多層感知機及支援向量機為基礎之兩種傳統模式作比較。結果顯示本研究發展之模式優於傳統模式，且在較長的預報時間點能夠有效改善預報之表現。綜合以上所述，本研究建立之新型之淹水災害分區模式和水位預報模式對於災害預警系統有相當之助益。

關鍵字

淹水災害分區 ； 淹水潛勢 ； 水位預報 ； 人工智慧 ； 隨機森林 ； 自組織映射圖 ； 擴展序列卷積神經網路 ； 多層感知機 ； 支援向量機 ； 淹水預警系統

並列摘要

Floods are among the most harmful natural catastrophes in the world, often resulting in loss of human lives and properties. To mitigate flood damage, the development of flood hazard zoning and water-level forecasting models has been played an essential role in disaster warning systems. Previous studies have shown the potential of artificial intelligence (AI) for modeling hydrological processes. However, in previous flood hazard zoning studies, there is no literature on the use of AI for performing flood hazard zoning assessments and consideration of potential flooding in surrounding environment. Moreover, in the past, traditional AI methods belong to shallow machine learning such as artificial neural network (ANN), support vector machine (SVM), and adaptive network-based fuzzy inference systems. The application of these methods is not sufficient to extract stable recognition features because they can only process natural data in the original format. In this thesis, novel approaches are established to construct flood hazard zoning and water-level forecasting models. Two parts are conducted herein to demonstrate the superiority of the proposed models.
In the first part of the thesis, a new type of flood hazard zoning model that uses integrated random forest (RF) and self-organizing map (SOM) methods is proposed. The model has two steps. The first is the creation of a module for flood susceptibility analysis to yield flood susceptibility values using the RF method. The second is the classification of flood susceptibility values according to the results of flood susceptibility analysis to obtain flood hazard zones with the use of a flood hazard zoning module based on the SOM network. Moreover, two different inputs for SOM are considered: (i) only the flood susceptibility value of a self-pixel is used as input, and (ii) the flood susceptibility values of the self-pixel and surrounding pixels are used as input. To examine the efficiency of the proposed model for flood hazard zoning, this study compares it with the existing model that is based on the natural break (NB) method. The proposed model is applied to the Lanyang Plain in Yilan County, Taiwan to demonstrate its advantages. The results indicate that the proposed model with flood susceptibility values from the self-pixel and surrounding pixels do improve assessment performance. The proposed model also performs better than the existing model, and it can provide optimal flood hazard zoning maps.
In the second part of the thesis, a novel water-level forecasting model based on dilated causal convolution neural network (DCCNN) is proposed to obtain water-level forecasts with a lead time of 1- to 6-h, because a DCCNN model can efficiently exploit a broad range of history. Residual and skip connections are also applied throughout the network to enable the training of deeper networks and to accelerate convergence. To demonstrate the superiority of the proposed forecasting technique, it is applied to a dataset of 16 typhoon events that occurred during 2012–2017 in the Yilan River basin in Taiwan. To examine the efficiency of the improved methodology, this study also compares the proposed model with two existing models that are based on multilayer perceptron (MLP) and SVM. The results indicate that the DCCNN-based model is superior to both SVM and MLP models, especially in terms of modeling peak water levels. Much of the performance improvement in the proposed model is due to its ability to provide water-level forecasts with a long lead time. In conclusion, the proposed modeling technique is expected to be particularly useful in supporting disaster warning systems.

並列關鍵字

Flood hazard zoning ； Flood susceptibility ； Water-level forecasting ； Artificial intelligence ； Random forest ； Self-organizing map ； Dilated causal convolutional neural network ； Artificial neural network ； Support vector machine ； Flood-warning systems

參考文獻

Al-Abadi, A.M., 2018. Mapping flood susceptibility in an arid region of southern Iraq using ensemble machine learning classifiers: a comparative study. Arabian Journal Geosciences 11, 218. https://doi.org/10.1007/s12517-018-3584-5.

Al-Abadi, A.M., Shahid, S., Al-Ali, A.K., 2016. A GIS-based integration of catastrophe theory and analytical hierarchy process for mapping flood susceptibility: a case study of Teeb area, Southern Iraq. Environmental Earth Sciences 75, 687. https://doi.org/10.1007/s12665-016-5523-7.

Alipanahi, B., Delong, A., Weirauch, M.T., Frey, B.J., 2015. Predicting the sequence specificities of DNA-and RNA-binding proteins by deep learning. Nature Biotechnology 33(8), 831–838. https://doi.org/10.1038/nbt.3300.

Alfieri, L., Salamon, P., Bianchi, A., Neal, J., Bates, P., Feyen, L., 2014. Advances in pan‐European flood hazard mapping. Hydrological Processes 28(13), 4067–4077. https://doi.org/10.1002/hyp.9947.

ASCE Task Committee on Application of Artificial Neural Networks in Hydrology. 2000a. Artificial Neural Networks in hydrology, I: Preliminary Concepts. Journal of Hydrologic Engineering, ASCE 5(2), 115–123. https://doi.org/10.1061/(ASCE)1084-0699(2000)5:2(115).