運用空間分析技術與Logit迴歸法建立崩塌地發生推估機率模式之研究－以雪霸國家公園為例

Logistic Regression Model for Predicting the Landslide Potential Probability based on Spatial Information Database-Case Study on Shei-Pa National Park

周天穎(Tien-Yin Chou) ； 雷祖強(Tsu-Chiang Lei) ； 李易儒(Yi-Ju Lee) ； 萬絢(Shiuan Wan) ； 林冠廷(Kuan-Ting Lin)

《水保技術》 7卷3期 (2012/07) Pp. 126-137

https://doi.org/10.30040/JSWCT.201207.0001

摘要

在台灣目前大部分的研究多集中在土石流潛勢溪流「危害度」部分，也就是透過模擬方式找出土石流在不同重現年期下的發生機率，但對於上游集水區土地利用因子的改變情形，特別是崩塌地時空消長情形與下游土石流的再發生機率，欠缺相互結合的機制，導致崩塌風險問題的評估結果產生了不確定性。因此本研究判釋數場颱風事件前後的影像變異情形產生出崩塌地變遷資訊，另外配合DEM資料以計算坡度門檻值方式，以改善影像混淆問題，最後整合崩塌地、降雨及地文資訊，建立上游集水區土利用因子資料庫。再以Logit模型取出時空變化之崩塌風險機率值後，建立所謂的「崩塌易損性機率模型」，其結果顯示影響崩塌地發生機率之因子包括：植生種類、土壤性質、距河道遠近及光譜指標等，而模式之整體判釋準確度可達76%。

關鍵字

遙感探測 ； 地理資訊系統 ； 崩塌地 ； Logit模式

並列摘要

In Taiwan, most studies are focused on the hazards analysis of debris flow potential streams recently. For instance, a simulation model to estimate of debris flow probability value at different rainfall return periods is designed to evaluate hazard levels of each watershed. However, few studies discussed the relations on the land use changes on up-streams with regards to the landslide hazards on down-streams for model evaluation. More specifically, the evaluation procedure needs an integrated mechanism to link the hazard values with the vulnerability values for reducing uncertainty problems at landslide risk model. Therefore, the collected data of study site into the landslide data warehousing, especially the induced factors (ex: rainfall and typhoons events), the potential factors (ex: DEM, geological factors, sensitive site factors...) and the SPOT satellite image data (1994, 1999, 2002, 2004) of prior and post typhoons events for landslide problems. To improve the easy confuse categories of rivers and landslide on the image classification produce, this study calculated the threshold of slope value from DEM data, in which this process can improve the classification results better. (That is, overall accuracy from 89% to 90%; Kappa Value from 80% to 85%). In addition, this study uses Logistic regression statistic technique to discuss the high relationship of landslide and environmental factors at vulnerability analysis to produce a landslide potential risk map. The results showed that the import factors of this model included: types of plant, quality of soil, distance from river and spectrum indices. The Logistic probability model has 76% of prediction accuracy on landslide region in this study site.

並列關鍵字

Remote Sensing ； Geographic Information System ； Landslide ； Logistic Regression Analysis