多時遙測光學與雷達資料於水稻田辨釋之研究

Abstract

本研究以多時段遙測光學影像及雷達微波資料，採用逐像元式及區塊式分類判釋水稻田。研究之目的在探討可以得到最佳分類成果的分類方式及不同水稻生長期所攝的影像組合。其中區塊式分類利用耕地坵塊資訊，輔以水稻不同生長期光譜反射及雷達回波強度相異的特性；藉由水稻生長光譜知識輔助水稻田分佈判釋。 逐像元式分類分別以單一時期與結合多時段光學影像資料判釋水稻田。單一時期影像利用SPOT三個波段資料，及再加入正規化差異植生指數與綠度指數影像，利用非監督式分類法判釋水稻田。多時段影像以各時期SPOT第二、三波段影像組合後進行水稻田判釋。相同的資料亦引進先分類，再導入區塊界限的分類修飾法判釋水稻田。 區塊式分類以坵塊為單元判釋水稻田；採用資料分別為SPOT影像的正規化差異植生指數、綠度指數、亮度指數及前兩種指數差值影像、RADARSAT資料的雷達背向散射係數及其差值。判釋分類結果以誤差矩陣（Error Matrix）表示，同時進行精度評估及滿意度分析。 區塊式分類統計結果，選擇符合分類精度90%以上，kappa 及Tau分析達0.85的差值影像，先去除確定不是水稻田坵塊後分類，以討論分類精度是否提昇。另外亦選擇兩個測試區，利用相同的判釋準則，驗証及探討區塊分類法推廣到廣大範圍的可行性。 對於光學影像，本研究獲得下列結果:多時段影像組合的最佳分類精度為93.06%，kappa 及Tau指標均為0.89；單一時期影像最佳分類精度為90.41%，kappa 及Tau指標均為0.85。證實多時段影像組合比單一時期影像有較好的分類結果。-結合水稻插秧期與最茂盛期的影像可達到最佳分類結果，加入其它時期影像的分類精度與滿意度分析並沒有顯著性差異(Z0.05測試均小於1.75)。R區塊式分類較逐像元式分類約提高1.0~1.5%的分類精度，kappa 及Tau分析亦提昇0.02(Z0.05測試均大於1.99)。 對於雷達資料，本研究獲得結果為:單一時期RADARSAT資料的水稻辨識率約為69.0%，兩個時期回波強度差值的分類精度為71.13%， 及Tau分析分別為0.54與0.39；此結果遠較光學影像分類結果差。-無法順利取得多時段光學影像判釋水稻田時，單一時期光學影像仍比RADARSAT資料有較好的水稻田辨識率。

Multi-temporal optical and radar data were used in this study to evaluate various classification approaches and search for an optimal combination of images taken in various stages of crop growth. Both conventional pixel-based and parcel-based classification approaches were adopted. For the parcel-based classification, attribute information for cadastral parcels and optical spectral characteristics and radar backscattering coefficient of various growing stages are all taken into account for achieving a better understanding of the rice paddy fields. For the pixel-by-pixel classification, NDVI (Normalized Difference Vegetation Index) and GI (Greenness Index) are derived for the identification of paddy fields in addition to the three SPOT spectral bands. In the multi-temporal processes, image combinations of SPOT band 2 and band 3 are used to analyze rice paddy by unsupervised classification method. Besides, these are also adopted in a modified classification method by first using a pixel-based classification and then introducing the parcel boundaries for grouping the classified results. For the parcel-based classification, indices such as NDVI, GI, BI (Brightness Index), and difference of NDVI and GI derived from SPOT images and backscattering coefficient, and difference of backscattering coefficient of RADARSAT are applied to identify the rice fields. Results are further characterized by error matrix, kappa , and τ. For evaluating the improvement of classification accuracy, the parcels of non-paddy are mask-out before setting a threshold to classify the rice fields for the difference images with a classification accuracy better than 90%, kappa , and τmore than 0.85. Images from two test areas are analyzed in a similar approach for verifying the possibilities of adopting common criteria. The results of the processing of optical images include: (1) An overall accuracy of 93.06% with kappa and Tau of 0.89 can be achieved by using an optimal combination of multi-temporal images; whereas, by using single time of images, a maximum overall accuracy of 90.41% with kappa and Tau of 0.85. This result shows that multi-temporal images give better classification accuracy than single time images. (2) It is shown that the best results were obtained by combining transplant and lush rice stages images. Little improvement of accuracy and indications of and Tau was gained by using additional images from other rice stages that the Z-test is not significant (Z0.05<1.75). (3) An improvement of classification accuracy of 1.0~1.5% and kappa and Tau of 0.02 was obtained by applying parcel-based classification as compared to pixel-based classification (Z0.05>1.99). The results of the processing of radar data include: (1) A classification accuracy of 69.0% was obtained by single time Radarsat images; an accuracy of 71.13% with kappa and Tau of 0.54 and 0.39 by difference images taken in two times. (2) In case where multi-temporal images are not available, optical images of single time would give a better classification result than Radarsat data.