深度學習雨量降尺度與土地利用分類應用淹水風險評估

A Deep Learning Approach to Precipitation Downscaling and Land Cover Classification for Flood Risk Assessment

張益嘉(Yi-Chia Chang)

指導教授 ： 童慶斌

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(2018年)

https://doi.org/10.6342/NTU201803314

摘要

洪水災害在過去數十年間造成人類社會巨大損失，尤其對於亞洲開發中國家而言，損失更為重大。為了增加應對洪水災害的韌性與調適性，洪水的現況與未來評估研究與知識建構更顯重要。未來雨量模擬是未來洪水評估中極為重要的一項評估因子，然而，全球環流模式（General Circulation Models, GCMs）模擬未來雨量的空間解析網格過於大（低解析度），無法直接使用於盆地和小型流域的洪水模擬。因此，動力降尺度技術計算大氣和地表環境的物理因子交互作用，被應用於全球環流模式產出以提高解析度。然而，動力降尺度模型侷限於高計算消耗成本，對於開發中國家研究氣候變遷有不利影響。因此，發展動力降尺度的詮釋模型（metamodel）將解決高計算成本的問題並提高網格解析度。本研究發展創新的卷積神經網絡（Convolutional Neural Network, CNN）應用於緬甸勃固流域（Bago River Basin, Myanmar）雨量降尺度模型，使用資料為衛星雨量資料ERA-Interim。本卷積網絡將傳統類神經網絡擴增並產出高解析度12公里日雨量。卷積神經網絡產出訓練與驗證年份為1984-2000年與2070-2085年， 測試年份為2001-2015年與2086-2100年。訓練完成的卷積網絡具有易可得性與可修改性於其他研究區域之潛力，對於研究未來洪水與使用動力降尺度雨量資料有貢獻和幫助。卷積網絡產出結果與動力降尺度結果有高度相似性，並減低88%動力降尺度所需計算時間。未來使用降尺度雨量的研究將可以更快速有效率地使用卷積網絡產出未來雨量資料。除此之外，本研究第二部分亦使用卷積網絡於土地利用分類評估洪水區域之暴露與脆弱度。卷積網絡土地利用分類模型可達到約99%之分類準確度。本研究第三部分使用降尺度雨量資料產出洪水危害地圖，重現2011年緬甸勃固流域洪水災害歷史情境與氣候變遷情境下的洪水危害地圖。本研究主要貢獻於延伸深度學習於雨量降尺度與土地利用分類，並將結果結合洪水模擬，評估未來洪水風險。

關鍵字

雨量降尺度 ； 洪水危害地圖 ； 土地利用分類 ； 卷積網絡 ； 緬甸勃固流域

並列摘要

Over the past decade, flood issues have been detrimental to human society, especially to Asian developing countries. To increase the adaptation ability and resilience to water disasters, the better understanding of flood is essential. Future climate data are critical for conducting the hydrological assessment. However, due to the scarce scale of climate data generated from General Circulation Models (GCMs), the native-scale outputs of climate models could not be directly utilized for basin-scale hydrologic models. Regional Climate Model (RCM) is a dominant technique to dynamical downscaling from the global scale to regional scale with consideration of atmospheric physical processes. However, the computational expense of RCM is costly for developing countries. Therefore, the metamodel of RCM for the specific area is eager to be designed to support decision making. This research introduces novel development and application of Convolutional Neural Network (CNN) to precipitation downscaling in Bago River Basin, Myanmar using satellite precipitation data from ERA-Interim. The CNN augments conventional neural networks by including two conv layers, two max-pooling layers, and one final fully connected layer to deliver the outputs of high-resolution 12km daily precipitation. The evaluation was based on the simulation of CNN under current and future projected climate conditions. The training and validation years are 1984-2000 and 2070-2085. The testing years are 2001-2015 and 2086-2100. The CNN model can be easily accessed and modified to improve the application for another specific area, not only beneficial to Myanmar-related researchers but global researchers interested in conducting the hydrological assessment using downscaled climate data. The results of CNN for downscaled precipitation show the high similarity to outputs of RCM and reduction of the significant amount of required computational resources. By including the time of training and implementing deep learning model, CNN takes approximately 12% of the time needed for RCM dynamical downscaling. With the CNN, researchers related to future climate change impact assessment may be capable of utilizing the metamodel to rapidly and efficiently retrieve the results of future downscaled climate data. Furthermore, CNN land cover classification model is also vital to understand the exposure and sensitivity of the society to flood events. The accuracy of CNN land cover classification model in training samples achieved 99.12%. The RRI model simulated daily inundation area under future climate impacts in 2011. Importantly, this study extends the study of deep learning model for precipitation downscaling, land cover classification, and inundation simulation.

並列關鍵字

precipitation downscaling ； flood inundation maps ； land cover classification ； convolutional neural network ； Bago River Basin

參考文獻

Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., et al. 2016. “TensorFlow: A System for Large-Scale Machine Learning.” In 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16), 265–83.

Campozano, L., Tenelanda, D., Sanchez, E., Samaniego, E., and Feyen, J. 2016. “Comparison of Statistical Downscaling Methods for Monthly Total Precipitation: Case Study for the Paute River Basin in Southern Ecuador.” Advances in Meteorology 2016.

Dee, D. P., Uppala, S. M., Simmons, A. J., Berrisford, P., Poli, P., Kobayashi, S., Andrae, U., et al. 2011. “The ERA-Interim Reanalysis: Configuration and Performance of the Data Assimilation System.” Quarterly Journal of the Royal Meteorological Society 137 (656): 553–97.

Demir, I., Koperski, K., Lindenbaum, D., Pang, G., Huang, J., Basu, S., Hughes, F., Tuia, D., and Raskar, R. 2018. “DeepGlobe 2018: A Challenge to Parse the Earth through Satellite Images.” CoRR abs/1805.0.

Field, C. B., Barros, V., Stocker, T. F., Dahe, Q., Dokken, D. J., Ebi, K. L., Allen, S. K., et al. 2012. Managing the Risks of Extreme Events and Disasters to Advance Climate Change Adaptation. Cambridge Univ. Press.