類神經網路於降雨引致深層崩塌影響之初步研究

台灣山區的地質條件普遍不佳，每逢颱風侵襲都會導致大範圍的土石流，近年因極端氣候發生的頻率增加，使得以往較少發生的深層崩塌浮上檯面，自小林村事件後，此類議題更加受到社會關注，相關的防災、預警研究也加速發展，力求減少人民生命財產的損失。本研究以南投縣的廬山北坡為場域進行監測資料的分析，將各項監測儀器的資料整理成離散的雨場形式，並找出雨場對應的地下水位變化及地中變位，同時以累積雨量、最大時雨量、降雨延時等三項數據為輸入變數，地下水位變化或地中變位為對應輸出值的結構下，以試誤法建構一三層的倒傳遞類神經網路架構。考量本研究的樣本數導致的類神經網路驗證問題，除普遍使用的均方根誤差外，另以預測區間為輔助方法減少結果受到隨機選取的影響，同時維持廣義解釋能力。最後將類神經網路的結果與傳統的線性迴歸方法比較後，可以看出以邊坡問題而言，類神經網路能夠得到比線性迴歸更加準確的結果，在預警系統的應用上也有更好的可靠度。

關鍵字

降雨；深層崩塌；倒傳遞類神經網路；預測位移

並列摘要

The geological conditions in the mountain areas of Taiwan are generally bad. Every year when typhoon strikes, it will lead to a wide-range debris flow. In recent years, the frequency of extreme weather has increased. Which caused the deep-seated landslide that rarely occurred in the past come to the fore. Since the Siaolin Village incident, such issues have gotten more attention by the society, relevant disaster prevention technique and early-warning research have been accelerated to reduce the loss of resident's lives and properties. This study uses the northern slope of Lushan, which located in Nantou County as the research area to analyze the monitoring data. First, organizing various monitoring instruments data into discrete rain field form, and find the groundwater level fluctuation also the ground displacements corresponding to the rain field. Secondly build a three-layer back propagation neural network structure by trial and error, which have cumulative rainfall, maximum rainfall, and rainfall duration as input values and groundwater level fluctuation or ground displacement as output value. Under the consideration of the neural network verification problem caused by the small sample size, besides the commonly used root mean square error, the ‘prediction interval’ is used as an auxiliary method to reduce the effect of randomly selected training samples to maintain the general interpretation ability of neural network models. Finally, by comparing the results of the neural network with the traditional linear regression method, we can discover that the neural network can obtain more accurate prediction than the linear regression in terms of the slope stability problem, and it could have a better reliability in the application of the early-warning system.