类神经网路与SPOT卫星影像分类之研究.docVIP

類神經網路與SPOT衛星影像分類之研究 Study in Neural Network and SPOT Imageries Classification 研究生：指導教授： 【摘要】 遙感探測可應用於大尺度自然資源之研究、規劃及經營管理，而遙測衛星影像分類之準確度為遙測影像應用之基礎。欲提昇遙測影像分類準確度，除改善衛星感測器之各項解析力外，執行適當的分類演算法則，亦為其關鍵。類神經網路系統應用於遙測影像分類，在許多的研究中，已被證實優於傳統統計分類法則。本研究以涵蓋墾丁國家公園陸域範圍之恆春半島為研究區，選擇SPOT HRV LEVEL 10 影像為分類材料，取其SPOT影像及加入綠度指數（NDVI）之影像，分別檢測最短距離法（Minimum Distance to Means Classifier）、最大概似法 Maximum Likelihood Classifier 及倒傳遞類神經網路（Back-Propagation Neural Network）分類法則，對高密度森林、中密度森林、低密度森林、草生地、水域、農耕地、建成地－河床－裸岩、沙灘、開墾後裸露地九種分類別之分類準確度及其漏授、誤授準確度（Omission and Commission Accuracy）。結果以類神經網路分類法對SPOT影像之分類準確度81.2%，KAPPA值0.774 為最高，而後依序為：加入綠度指數影像後倒傳遞類神經網路分類法準確度80.3%，KAPPA值0.764；加入綠度指數影像後最大概似法準確度77.7% ，KAPPA值0.732；最大概似法對SPOT影像準確度76.8%，KAPPA值0.722；加入綠度指數影像後最短距離法準確度68.2%，KAPPA值0.617；最短距離法對SPOT影像準確度66.7%，KAPPA值0.604。而各種分類法則，對不同之地被類別之準確度，卻各有其擅長之處。本研究取各分類別準確度較高者為優先，依序疊合出分類準確度最高之地被類別分佈圖，其總體準確度達82.3%，總體KAPPA值0.787。 關鍵字：恆春半島 倒傳遞類神經網路 最大概似法 最短距離法 綠度指數 【Abstract】 Remote sensing technology can be applied to a large-scale nature resource research, planning and management. The precision in classifying the remote sensing satellite imageries is a foundation of the application of the imageries. However, to improve the precision, it must first improve the resolutions of the sensor and execute a proper while classifying imageries. The use of neural network has been proven to provide a higher precision in the classification of satellite imageries than the classical statistical methods. It is a fact that different classification methods can get different degrees of precision on different types of ground covers. However, the degrees of precision may vary with the selective incorporations of proper synthesized imageries into different image classification methods. The area around the Kenting National Park in the Hengchuen Peninsula, Pingtung County, was selected as a test site for this study. The SPOT HRV multi-spectral imagery（original imagery） and superimposing the original imagery with the NDVI imagery was used as the test mater