航j則及遙測學刊第九卷第一期第 13-26 頁民國 93 年 3 月
Journal of Photogrammetry and Remote Sensing Volume 9,No.1,March 2004,pp. 13-26
不同影像分類方法應用於水稻辨識之探討
吾吾國鑫 l 劉治中 1 徐律城 2
摘要
13
本研究結合遙測多時影像與耕地資料，利用 CDNDVI 指數區間判釋法@貝氏機率法@統計機率法@
物件導向模糊分類等方法辨識水稻垣塊分布。後三種分類方法完成對應垣塊被辨識為水稻的機率後，再 轉換為硬式二分類別，以判釋該垣塊是否為水稻田，並比對航照判釋資料評估分類精度。研究結果顯示 以兩帽的， 000 像片基本圖為測試範圍，評估二分類別辨識水稻的全體精度與 k 指標，以貝氏機率分類法 最佳(96.27% 與 0.92)' NDVI 指數區間判釋法次之 (92.22% 與 0.83) ，統計機率分類法再次之 (9 1. 89%及 0.82) ,
物件導向模糊分類法則為 90.97%與 0.81 。前兩種分類方法推廣到苗栗、台南與屏東地區的水稻田判釋分
類，貝氏機率分類的全體精度與 k 指標平均較 NDVI 指數區間判釋法高 1-4%及 0.04-0.15
0所以利用貝氏
機率分類除了可以得到恆塊被辨識為水稻的機率外，轉換為二分類別亦可得到較佳的分類結果。
周鍵詞:木菇、主丘塊辨識、逼到、影像兮類、正規化差異植生指數
1.前言
稻米為台灣地區重要的糧食作物;因此，快 速精準的統計稻作種植面積與估算產量為當務之 急的工作;特別是進入國際貿易組織(WTO)後的逐 年開放稻米進口、乾早年配合的水稻田休耕或轉 作等措施，均會影響稻米產量;所以提前獲取水 稻相關資訊為目前農政單位重要的工作之一。
結合遙測多時 (multi-temporal) 影像與耕地資 料判釋水稻分布的前期研究，如陳益凰、曾義星 (1 999)曾結合水稻生長知識，採用差分影像分類法
(Differenced ImageClassification)、結合水稻生長的 時間剖面匹配法(TemporalProfileMatching)及波峰
l 工研院能資所研究員 2 工研院能資所副研究員
偵測法(Peak Detection)判釋水稻田;實驗成果以差 分影像分類法可以得到較佳的全體分類精度。本研 究先期工作亦曾結合多時段影像的 NDVI 植生指 數、農地垣塊及水稻全生育期(full-growth-period) 光譜反射資訊，以NDVI 指數區間判釋法辨識彰化 全區水稻分布;並證實此種區塊式分類法的分類精
收到日期:氏國 92 年 02、丹也 修改日期:民國 92 年 08 月 22 日 接受日期:民國 92 年 08 月 24 日
26 Journal of Photogrammetry and Remote Sensing, Volume 9, No.1, March 2004
The Evaluation of Image Classification Methods for Rice Paddy Interpretation
Kuo-Hsin HSIAO
IChi-Chung LAU
IWei-Chen HSU
2ABSTRACT
Multi-temporal imageries and cadastral GIS datasets were combined to interpret the rice paddy distributions, by applying the following means: (1) NDVI intervals approach on basis of multi-temporal SPOT images, (2) the Bayesian classifier based on spectral reflectance curve measured from different growth stages, (3) the statistic probability classifier, (4) the object oriented fuzzy classifier. Results of last three classification approaches were coupled with a prediction of probability. A selected threshold was applied to the results to obtain a dichotomy category. These results were compared with visual interpreting data from aerial photos for assessing accuracy. It
"
shows that the overall accuracies and
k
index for the four means are 92.22% & 0.83, 96.27% &0.92,91.89% & 0.82 and 90.97% & 0.81, respectively. Similar approach is applied to a contrast test site for the rice paddy in the first season. The results of the above-mentioned classifiers show overall
"
accuracies and
k
index of 94.27% & 0.83, 95.43% & 0.87 and 92.75% & 0.79 that deducted from the object oriented fuzzy classifier. For the rice paddy in the second season, results of the"
above-mentioned three classifiers show overall accuracies and
k
index of 93.99% & 0.79, 95.33%& 0.84 and 93.47% & 0.77. There are the same classified results that applied to Tainan and Pindong County for rice paddy interpretation. It is concluded that the Bayesian classifier not only can generate probability of each class, but also give better classified results after thresholding process.
Key Words: Rice Paddy Interpreted, Remote Sensing, Image Classification, Normalized Difference Vegetation Index (NDVI)
1Researcher, Energy and Resource Laboratories, Industrial Technology Research Institute 1Researcher, Energy and Resource Laboratories, Industrial Technology Research Institute 2Assistant researcher, Energy and Resource Laboratories, Industrial Technology Research Institute.
