深度學習與資料擴增於山崩監測預測之可行性評估;Development of Deep learning and GAN applied to landslide prediction

NCUIR > Engineering College > Graduate Institute of Civil Engineering > Electronic Thesis & Dissertation > Item 987654321/89062

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/89062

Title:	深度學習與資料擴增於山崩監測預測之可行性評估;Development of Deep learning and GAN applied to landslide prediction
Authors:	邱鈺智;Qiu, Yu-Zhi
Contributors:	土木工程學系
Keywords:	深度學習;山崩預測;資料擴增;生成對抗網路;支持向量機;長短期記憶神經網路;門閥循環單元網路;Deep Learning;Landslide Prediction;Data Augmentation;GAN;SVM;LSTM;GRU
Date:	2022-09-19
Issue Date:	2022-10-04 10:50:02 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著氣候的劇烈變遷，山崩破壞也越加嚴重，因此如果能預測山崩就能減少人員傷亡以及財產損失，而現今山崩預測研究中又以深度學習（Deep Learning）的成長最為顯著，但利用深度學習預測山崩上有一困難點為山崩監測資料較為缺乏，如地下水位、邊坡位移量獲取相對於其他水文或氣象資料更難以取得。而在深度學習模型的訓練過程中，是需要有大量的資料才能有良好的預測結果，故本研究同步提出使用生成對抗網路（Generative Adversarial Networks, GAN），以2007-2008年廬山監測資料為範例進行資料擴增，並建置支持向量機（Support Vector Machine, SVM）、長短期記憶網路（Long Short-Term Memory Network, LSTM）和門閥限循環單元網路（Gated Recurrent Unit Network, GRU）三種預測模型，除探討利用原始資料進行訓練預測外，本研究並利用資料擴增後之資料集探討預測結果。實驗顯示使用原始資料且預測成效較差的水位(孔位B01、B04、B07)，其GRU預測R-Square值分別為 0.690、0.347、0.759，再使用GAN資料擴增後，B01、B04、B07 GRU預測R-Square值提升至0.943、0.901、0.760。最後使用文獻兩組山崩實際案例進行比較可得知，在資料擴增後預測山崩位移皆有更佳預測結果。本研究所提出深度學習應用可應用對於缺乏山崩資料而想使用深度學習預測之參考依據。;With the dramatic climate changes, landslides damage has become more serious. Therefore, if landslides can be successfully predicted, casualties and property damage can be reduced. For various landslide predictions, the growth of Deep Learning is the most significant. However, a major difficulty in landslide prediction is obtaining landslide monitoring data, such as groundwater level and slope displacement. They are more challenging to obtain than other hydrological or meteorological data. In the deep learning model training process, a large amount of data is necessary for good prediction results. Therefore, this study proposes to combine Generative Adversarial Networks, using data augmentation with 2007-2008 Lushan monitoring data as an example, and build three prediction models: Support Vector Machine (SVM), Long Short-Term Memory(LSTM), and Gated Recurrent Unit (GRU), to compare whether the augmented data has more predictive results than using the original data. Experiment results show that GRU with original data can predict water levels and R-Square values of B01, B04, and B07 are 0.690, 0.347, and 0.759. Using GAN data augmentation for water levels has the predicted R-Square values of B01, B04, and B07 of 0.943, 0.901, and 0.760. The latter one has significantly improved. This study further evaluates deep learning applications that can be applied in other cases. Consequently, the proposed GRU with the GAN method is a feasible approach for landslide prediction.
Appears in Collections:	[Graduate Institute of Civil Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	112	View/Open

社群 sharing

Loading...