台灣本島因地形陡峭,河川源短流急,河川上游集水區遇暴雨時,土質鬆軟處本即易於出現崩塌,致河道沖蝕益形嚴重,致集水區之水源涵蓄能力降低,而沖蝕之土石進入水庫後又造成淤積,降低了現有蓄水設施供水及調節能力。因此,如何能加以判釋與監測崩塌地發生並有效的進行災害管理為當務之急。 本研究的目的乃是利用遙測及地理資訊系統技術進行崩塌地之自動化判釋。首先根據衛星影性之光譜特性,對研究區(石門水庫上游)進行植生指標的分析,嘗試將裸露地與植生地區分出來;再進一步針對裸露地進行影像分類,將河川地與崩塌地分離,針對崩塌地的部分,主要利用判別分析法在統計上之特性,加強各地徵之差異度以提高判釋精度;最後並與農林航測所航測影像,以人工圈選崩塌地當作地真資料做成果的比對,實驗成果顯示可達85%左右的正確率。 由判釋所得之崩塌區域進而研究各影響因子對其崩塌潛勢之影響,利用類神經網路所具有之非線性及平行處理能力,來處理各項參數對邊坡崩塌之影響,由此判釋邊坡是否可能發生崩塌。研究結果顯示,經由網路之訓練與測試,證明類神經網路對於研究區的崩壞與否之正確判釋率達92%。 本研究將藉由衛星影像進行崩塌地判釋與其災因分析,了解石門水庫集水區現況與自然環境、人文開發的關連性,以提供土石災害之決策與管理,採取必要的預防措施。
Because of the precipitous terrain and short rapid rivers in Taiwan, the landslides usually happen after heavy rains and the resulting floods. Landslides are natural hazards that often cause series property damages and even life losses. This makes the landslide monitoring and mitigation techniques an important study issue for the related professional disciplines in Taiwan. As the developing of remote sensing technology, both the spatial and spectral resolution of satellite images become more mature for objection identification and detection. In this research, the landslide areas are detected using the pre- and post-images by the decision trees classifier. In the first stage of the decision tree, the thresholding method according to the variety of vegetation index (VI) is used to separate the areas into vegetated and exposed areas. In the second stage, In the third stage, using the property in statistics ofdiscriminant analysis to strengthen the difference of land features before doing image classification. In the third stage, a supervised classifier is used to distinguish the natural streams from the exposed areas to obtain the areas of landslides. Finally, the result will check with ground truth data which made manually in aerial photos. Using the ability of artificial neural network(ANN) to predict each factor effect in landslide. After training and testing data by ANN, results of this study indicate that the ANN method can predict with an accuracy up to 93%. Therefore, it is concluded that the ANN method is a good way to predict the hazard before occurring. The research applies satellite image and hazard analysis to realize the environment in Shihmen reservoir watershed. The detected landslide information could be utilized as a reference for the mitigation of the future landslide.