以隨機森林判釋都市集水區保育地貌之空間分析研究

The Spatial analysis study using random forest to interpret conservation landforms in urban watershed area

林宸煒(Chen-Wei Lin)；萬絢(Shiuan Wan)；鄭育欣(Yu-Hsin Cheng)；劉建辰(Chien-Chen Liu)

貝葉斯優化 ； 紋理資訊 ； 隨機森林 ； Bayesian optimization ； Random forest ； Texture information

國土測繪與空間資訊

12卷2期（2024 / 07 / 01）

189 - 202

繁體中文;英文

在保育水源和確保永續發展的議題上，集水區水源保育範圍可能因為違規開發或自然災害導致的土砂災害等問題，常直接或間接影響水源水質或造成嚴重水庫淤積，對水資源的永續利用造成危害，目前集水區土地管理主要依賴人力巡查來監測是否有違規的情況，以維護集水區之生態保育。本研究探討如何透過資料探勘和影像辨識技術，搭配機器學習，快速計算和判斷森林的相關信息，本研究以高解析度影像，使用不同特徵後以隨機森林（Random Forest, RF）的方法進行分析，並以資料視覺化（Data Visualization）的方式呈現結果。透過監督學習方法，使用訓練樣本建立模型，再使用測試樣本進行預測並以貝葉斯優化（Bayesian Optimization, BO）進行隨機 森林的參數優化。最後，利用函式計算和繪圖套件，將結果輸出成主題圖（Thematic Map）和誤差矩陣圖（Confusion Matrix）資料視覺化的結果來分析和比對，本研究同時使用了如何使用參數選擇（Parameter Selection）、和紋理資訊（Texture Information）等方式提高模型的準確性。

The study of watershed conservation area for the sustainable development may be directly or indirectly affected by problems such as soil and debris flow disasters caused by illegal development or natural disasters. It often impacts the water quality and also causes severe reservoir sedimentation or produced by a threat to the sustainable utilization of water resources. In present, the land management in watersheds mainly relies on manual patrols to monitor for any illegal situations to maintain ecological conservation in the watersheds. Hence, this study explores how to use data mining and image recognition techniques, combined with machine learning, and rapidly calculate to determine the relevant information of environments. This study utilizes high-resolution images and analyzes them using different features with the Random Forest method, presenting the results through data visualization techniques. Applying a supervised learning approach, the training samples are used to build a model, and then using test samples for prediction with the Bayesian Optimization for modifying the Random Forest parameters for model tuning. Finally, utilizing function calculations and plotting libraries, the results are drawn as a thematic map and a series of confusion matrix for data visualization analysis with comparison. This study also employs methods such as parameter selection and texture information to improve model feasibility.