中文摘要 雲是影響地球天氣系統的重要因素之一，但其型態變幻萬千，因此在對雲進行研究之前，須先針對雲參數進行探討，包括總雲量、雲速、雲層厚度及本文所要探討的雲高。 雲高的辨識，除了使用目測、探空氣球及光達觀測等傳統方法外，目前亦可應用衛星觀測影像配合遙測理論來進行雲分類之應用研究。由於衛星觀測資料可有效率地提供大範圍的訊息，故本文將使用新一代之MODIS( Moderate-resolution Imaging Spectroradiometer)衛星影像資料來進行雲高之分類。MODIS觀測資料的波長範圍從可見光、近紅外、中紅外到熱紅外線，共有36個頻道可供分析應用。 本文採用17個在雲辨識上比較常使用之頻道，並利用倒傳遞類神經網路(BackPropagation Networks；BPN)來進行分類。由於雲高之分類結果在驗證上具有較高困難度，因此本文將利用遙測理論所獲得雲頂溫度及雲高等結果，作為分類模式輸出目標值以及辨識結果驗證之參考。最後，本文以多筆資料來進行結果之測試比對，顯示此法具頗高之準確度及可行性。 Abstract Cloud is one of the very important factors influencing the weather system of the earth. Since the cloud type in attitude fluctuates multifariously, discussions on the cloud parameter have to be conducted firstly in this thesis. These include “total cloud amount”, “cloud speed”, “cloud thickness”, and “cloud high”. The classification of cloud high in traditional methods is performed by the estimating of human eyes, radiosunde, and liadar detection …etc. In addition, we can apply satellite images to collocate with remote sensing theory to implement the cloud classifies. Since satellites can offer the information on a large scale to observe the materials efficiently, a new generation's MODIS (Moderate-resolution Imaging Spectroradiometer) is used in this thesis to obtain satellite image data to conduct the “cloud high” classification. The observation wavelength range of MODIST is from visible, near IR, middle IR to thermal IR, which has 36 channels suitable for analyzing and accomplishing the application. In this thesis, 17 channels that are frequently used in cloud distinguishing are adopted, and the “Back Propagation Networks” (BPN) is employed for classification purpose. Since it is difficult to evaluate the results of cloud high classification, the cloud top temperature and cloud high results generated from remote sensing theory are used to obtain the goal value of classification model outputs to distinguish the results for verification and consulting purposes. In our experiments, image data of different seasons were collected to test and compare the performance. Experimental results demonstrate that the proposed method is feasible and achieves very high accuracy rate in cloud classification
