Discriminate Rice Cultivars Using Canopy Reflectance Spectra

本研究旨在量測不同水稻品 種，即臺農67 號(TNG 67)、臺8 號 (TK 8)及臺農 71 號 (TNG 71)，在生育期間及期作間之近地面高解析植被反射光譜 (350-2400 nm)，以篩檢合適光譜特徵，建立辨識水稻品種之光譜遙測模式。田間試驗係於2002年在臺中縣霧峰鄉行政院農業委員會農業試驗所農場進行，計有一、二期稻作。將植被光譜涵蓋之所有窄波段先經由逐步判別分析 (stepwise discriminant analysis， SDA)篩 檢出若干條合適的光譜特徵窄波段，續由主成分轉換 (principle components transformation，PCT)構成多項主成分，再由判別分析 (discriminant analysis，DA)將各主成分建立辨識水稻品種之光譜遙測模式。結果顯示，TNG 67、TK 8 及 TNG 71 等三水稻品種的辨識準確度與所選擇特徵窄波段數之 間呈現飽和指數型曲線模型 (exponential rise to max)，一、二期稻作皆然。整體而言，在選擇相同窄波段數的條件下，一期稻作的辨識準確度高於二期作者;例如，一、二期稻作在選擇7條特徵窄波段時，全體準確度分別達到87%及82%。當選擇的特徵窄波段數介於7-10條時，辨識度緩慢提升，惟當窄波段數目超過10條以上時，所增加的辨識能力將十分有限。欲達到95%以上的全體準確度 (overall accuracy) ，則特徵窄波段數至少需要達到15條。又由檢視的三品種間樣品族群空間距離分佈結果，發現一期稻作的 TK 8 與 TNG 71 之植被反射光譜特徵表現較相近，而 TNG 67 與此二品種的差異較大;二期稻作則以TK 8與TNG 67 之植被反射光譜特徵表現較類似，而TNG 71 則與此二品種的差異較大。本研究據此提供了栽培於一、二期稻作之 TNG 67、TK 8 與 TNG 71 等三品種利用植被光譜進行品種辨識的方法，並初步瞭解此三品種間光譜特徵之差異。 Ground-based remotely sensed high-resolution canopy reflectance spectra (350-2400 nm) were obtained from field experiments to study spectral differences among rice cultivars during the growing periods and between the cropping seasons of 2002. The suitable spectral characteristics were also selected to build the spectral models for cultivar discrimination. In the first step, 3-11 narrow bands were selected progressively from reflectance spectra as the spectral characteristics by the stepwise discriminant analysis. Secondly, through the principal components transformation processes the selected narrow bands were composed to various components, which were used to establish the spectral models by the discriminant analysis. It indicated that the accuracy in discriminating cultivars TNG67, TK 8, and TNG71 was related to numbers of selected narrow bands, expressing as an exponential rise to maximum function in both cropping seasons. Generally the accuracy in cultivar discrimination was higher in the first crop than in the second crop under same numbers of the selected narrow bands. For example, when 7 narrow bands were selected, overall accuracy was 87% in first crop and was 82% in second crop. The percent overall accuracy increased slowly when 7 to 10 bands were selected. To reach 95% of overall accuracy, at least 15 bands should be selected. It further showed that the distance distribution among the sampling populations of three cultivars was different between cropping seasons. The distance between TK 8 and TNG71 was smaller than other pairs in the first crop while the distance between TK 8 and TNG67 was smaller than others in the second crop. As a result, this study provides information about differences in spectral characteristics among cultivars TNG67, TK 8 and TNG71 as well as an appropriate analysis method for discriminating these three cultivars using hyperspectral data of both cropping seasons