基於物件之隨機森林法於 UAV 航照影像之土地利用分類及變遷偵測

An Object-based Random Forest Approach for Landuse Classification and Change Detection Using UAV Imagery

王嘉陽(Jia-Yang Wang)

指導教授 ： 徐百輝

共同指導教授 ： 趙鍵哲(Jen-Jer Jaw)

國立臺灣大學/工學院/土木工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202102589

摘要

隨著都市的發展，土地利用方式越來越複雜，土地利用違規以及後續違規調查 及取締的問題亦日漸增加，針對土地利用變遷偵測的方法及技術因此受到重視。同 時，由於感測技術的進步，利用無人飛行載具所獲得高解析度影像已經相當便利。 與過去空間解析度較低的衛星影像相比，無人飛行載具的影像提供更多的空間細 節，有助於獲得都市區中較複雜的地表資訊。此外，隨著影像解析度愈來越高，過 去以像元為基礎的影像分析方法計算時間也大幅增加，且很容易產生椒鹽現象，因 此以物件為基礎的影像分析方法成為熱門的解決方式，其不僅能夠解決椒鹽現象 的問題，亦能提供更多分析特徵以及可進行多尺度分析等優勢。目前已有許多結合 影像物件及決策樹偵測進行土地變遷偵測之相關研究，其雖可獲得一定程度的準 確率，然而制定決策樹的規則過程複雜且耗時。除此之外，每當更換研究區域時， 就必須重新制定規則，尤其浪費大量時間及人力。近年來，機器學習演算法廣泛應 用於遙感探測領域，其優點就是能夠藉由訓練資料建立一套自動化的預測模型。
因此，本研究目的在於使用無人飛行載具拍攝高解析度影像資料，針對都會區
進行土地利用變遷偵測，利用以物件為基礎的變遷偵測概念，並利用隨機森林演算
法以提升變遷偵測之效率，期望結合影像物件及機器學習演算法，以建立一套適用
於台灣都會區變遷偵測的機器學習模型。本研究模型整體準確度為 0.94，Kappa 值
為 0.93，將訓練模型應用在不同測試區域時，整體準確度皆能達到 0.85 以上，Kappa
值也能達到 0.8 以上。另外，針對隨機森林模型的投票機制提供閥值設定的依據，
確保模型成果的正確性。最後，探討不同特徵對於各型態判釋的重要程度，提供後
續研究建立更快速、更準確的偵測模型。

關鍵字

以物件為基礎影像分析 ； 變遷偵測 ； 隨機森林

並列摘要

With the growth of economy and urban development, the types of landuse become more and more complex. Moreover, some problems such as landuse planning and illegal construction are emerging one after another. However, current change detection methods waste too much human resources and time, besides, it needs professional knowledge to deal with these kinds of problems. Therefore, we are looking for an automatic change detection method. With the advance of remote sensing technology, the image resolution become higher and higher. Unmanned Aerial Vehicle (UAV) and modern satellites can provide very high resolution (VHR) images. VHR images can provide more detail information. Pixel-based Change Detection (PBCD) no longer to be a appreciate method which has salt-and-pepper effect and poor performance on land-use change detection. Therefore, many studies prefer to use Object-based Change Detection (OBCD) which is based on the concept of Object-based Image Analysis (OBIA). OBCD reduces small spurious changes, provides more features, such as shape, area, size and provides spatial multiscale analysis. In this study, using very high resolution image and digital surface model data and extract four types of features, including shape, spectral, texture and elevation features from each object. We use Random Forests (RF) algorithm combining OBIA for urban change detection. Random Forests is a kind of ensemble machine learning algorithm, which is a model that makes predictions based on a number of different models. Combining OBIA and Machine Learning to establish an automatic model for urban change detection, the overall accuracy of model is 0.94 and kappa coefficient is 0.93. The overall accuracy of testing area is 0.85 and kappa coefficient is 0.80 in average. Therefore, Discussing the threshold setting of RF voting and providing important features as a reference for future research.

並列關鍵字

Object-based Image Analysis ； Change Detection ； Random Forests

參考文獻

Belgiu, M. and L. Drăguţ , 2016. Random forest in remote sensing: A review of applications and future directions. ISPRS Journal of Photogrammetry and Remote Sensing, 114: 24-31.

Blaschke, T., S. Lang, and G. Hay, (Eds.). , 2008. Object-based image analysis: spatial concepts for knowledge-driven remote sensing applications. Springer Science Business Media.

Breiman, L., 2001. Random forests. Machine learning, 45(1): 5-32.

Butt, A., R. Shabbir, S. S. Ahmad N. Aziz, 2015. Land use change mapping and analysis using Remote Sensing and GIS: A case study of Simly watershed, Islamabad, Pakistan. The Egyptian Journal of Remote Sensing and Space Science,

18(2), 251-259.