土石流災害數值預測模式之分析與研究

Abstract

台灣自歷經921地震後，山區經常發生大規模土石崩塌，尤其每當颱風季節或豪雨來臨，便引發土石流災害，因此土石流災害發生之預測與相關決策系統建置，已成為國家防災計畫中關鍵議題之一。有鑑於此，本論文利用數值回歸分析法提出可凖確預測土石流發生之數學模式，並利用代理人機制與個人行動通訊網路建置土石流防災即時決策支援系統。本研究樣本為南投縣境內181條土石流潛勢溪流，並由土石流發生學理與歷史資料分析提出八項土石流災害發生因子，且藉由全球衛星定位系統(GPS)、地理資訊系統(GIS)與遙感探測(RS)資訊技術找出土石流災害因子參數值進行回歸分析與驗証。本研究所提出之土石流災害預測模式建立在數值回歸預測模式上，為採用多元線性回歸、多變量不安定指數法與倒傳遞類神經網路等三種統計分析法進行回歸分析與比較，並依據模式驗証歷史資料以建立精確及客觀之最佳土石流預測模式。即時防災決策支援系統之設計則利用GIS系統連結行動式手持式設備(Handheld Devices)，並建置環境認知代理人提供雨量、地形地貌等客制化(Customized)資訊，有效實現土石流災害發生之即時責訊決策與傳送，以達到精確的災害預測與救災指引功能。

The effective disaster prediction is based on the correct debris-flow decision model and the real-time information communications between the disaster area and the rescue-control center. In this paper, we proposed and designed a Real-time Mobile Debris Flow Disaster Forecast system (RM(DF)^2), which is composed of the mobile clients, the application servers, and the decision support server based on the wireless/mobile and Internet communications. Mobile clients use handheld devices, e.g., PDA combining a cellular phone, to transmit and receive multimedia debris-flow information via the GSM/GPRS network. The application server, which is composed of a Virtual-Reality manager and seven intelligent agents, provides the debris-flow VR emulation and the customized information with mobile users. The customized operations could effectively reduce the bandwidth consumption of the mobile network and release the computing load of handheld devices. We proposed three effective debris-flow prediction models and the inference engine in the decision support server. The proposed prediction models are based on the linear regression, the multivariate analysis, and the back-propagation network schemes. To have a practical simulation environment, the database of the decision support server is the pre-analyzed 181 potential debris-flows in Taiwan. According to the simulation results, the prediction model of adopting the back-propagation network scheme achieves the effective debris-flow prediction with high degrees of accuracy. We also defined eight prediction factors of debris-flows, which can be extracted using GPS, GIS, and Satellite Remote Sensing (RS) techniques, as the parameters of prediction models. The implementation results of the RM(DF)^2 system reveal that the proposed prediction models and system architecture are feasible and could achieve effective prediction and presentation of debris flows.