鄭克聲臺灣大學：生物環境系統工程學研究所鍾建文Chung, Chien-WenChien-WenChung2007-11-272018-06-292007-11-272018-06-292006http://ntur.lib.ntu.edu.tw//handle/246246/56070近年來，遙測技術被廣泛的運用在各個領域上，其中人類想要以最迅速及最正確的方式來瞭解地表土地利用之情形與土地變遷的判釋，但隨著土地利用型態的複雜化，欲區分的土地利用類別亦是增加。本研究以階段式分類法，探討各個階段使用不同之特徵空間進行分類，各階段採取之分類方法皆為貝氏分類法，第一階段以光譜特徵區分大類別，第二階段則加入紋理特徵輔助分類，將裸露地與農地於大類別中個別取出，進行小類別的分類，並計算小類別間的發散度，選擇較容易區分小類別之特徵。最後結果以貝氏分類並配合階段式分類法進行多類別的土地利用判釋，可達到更佳之分類正確率。 變遷偵測方面，將光譜特徵進行主成分轉換，並使用第一主成分作為判釋的特徵，其具有降低資訊量及多維度判釋變遷之優點。判釋方法是以條件機率的方式，建立判斷變遷像元的信心水準，以信心水準作為判釋變遷的依據，改變以往以經驗法則或試誤法訂定之變遷門檻值。最後與一般常用之變遷偵測方法互相比較，結果顯示，以本研究方法判釋變遷的地區，較能符合現地資料狀況。Remote sensing images and technologies have been widely applied to environmental monitoring, in particular landuse/landcover (LULC) classification and change detection. The accuracy of LULC classification depends on the spatial resolution of remote sensing images, features (spectral or textural) adopted for classification, the desired landcover classes, and also the classification method. In cases where complex landuses are present and detailed LULC classes are desired, it is often difficult to achieve high level of classification accuracy. In this study, a two-stage Bayesian classification approach was proposed to circumvent such difficulties. In the first stage, only spectral features were adopted for coarse classification (bare land, agriculture, water body, grassland, and forest). Then, textural features were considered to conduct within-class classification in the second stage. The bare land class was divided into bare soil and built-up and the agriculture class was divided into orchard, vegetation garden, and tea plantation. Application of the proposed approach in the Chi-Jia-Wuan Creek watershed in central Taiwan for 2004 and 2005 yields about 98% overall accuracy in the first stage and 86% and 89% overall accuracies in the second stage. For LULC change detection, a hypothesis-test-based multispectral algorithm was developed. The whole study area was classified into three major classes - forest, water body and bare land using multi-date (2004 and 2005) and multispectral images. Such coarse classification can achieve high level of classification accuracies. No-change pixels of individual classes were then identified and used as the basis for establishing 95% confidence intervals for LULC change detection. The first principal component of the original multispectral features of 2004 (PCX1) and the first principal component of the multe-date multispectral differences (PCΔX) were used to construct bivariate normal distributions for the three major LULC classes. Then, given the value of PCX1, the conditional probability distribution of PCΔX can be spefied. Therefore, under the null hypothesis of no change, the 95% confidence intervals of individual LULC classes can be established. Using a set of validation data, the proposed change detection algorithm is shown to be capable of achieving high accuracies.中文摘要 i 英文摘要 ii 目錄 iv 圖目錄 vi 表目錄 viii 第一章 緒論 1 1-1 前言 1 1-2 研究內容與架構 2 第二章 文獻回顧 4 2-1影像分類 4 2-2 紋理特徵與特徵選取 7 2-3 變遷偵測 8 第三章 理論介紹 12 3-1遙測基本原理 12 3-1.1 簡介 12 3-1.2 遙測衛星資料特性 17 3-2 影像幾何校正 19 3-3 影像分類 22 3-3.1 紋理特徵 22 3-3.2 特徵選取 24 3-3.3 分類演算法 27 3-3.4 分類正確性評估 30 第四章 研究地區與資料 32 4-1 研究地區 32 4-2 研究資料 33 4-2.1 衛星遙測影像 33 4-2.2 五千分之一相片基本圖 35 4-2.3 紋理特徵資料 36 4-3 地面控制點 36 4-4 訓練樣區 38 第五章 研究內容 40 5-1 研究流程 40 5-2 幾何校正 43 5-2.1 控制點選取 43 5-2.2 初期轉換 44 5-2.3 半變異元分析與模式選取 45 5-2.4 模式驗證與推估誤差評估 48 5-2.5 重新取樣 49 5-3 訓練樣區分析 50 5-3.1 一階段樣區分析 50 5-3.2 二階段樣區分析 52 5-4 特徵選取 56 5-5 階段式分類 59 5-5.1 一階段分類 59 5-5.2 二階段分類 60 5-6 變遷偵測 62 5-6.1 大氣校正 62 5-6.2 條件機率之建立 63 5-6.3 變遷像元之決定 64 第六章 結果與討論 66 6-1 幾何校正 66 6-2 影像分類 69 6-2.1 一階段分類結果 69 6-2.2 二階段分類結果 72 6-2.3 分類結果討論 76 6-3 變遷偵測結果 78 6-3.1 條件機率之建立 78 6-3.2 變遷像元決定 82 6-3.3 變遷偵測結果驗證 85 6-3.4 評估與比較 89 第七章 結論與建議 96 參考文獻 98 附錄 101 附錄A 紋理特徵影像 101 附錄B 大氣校正前後之像元組體圖 1041646371 bytesapplication/pdfen-US階段式貝氏分類法紋理特徵發散度變遷偵測第一主成分條件機率Two-Stage Bayesian ClassificationTextural FeatureChange DetectionPrincipal ComponentConditional Probability[SDGs]SDG15thesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/56070/1/ntu-95-R93622048-1.pdf