以高光譜影像技術建立蝴蝶蘭開花品質模型之研究

蝴蝶蘭優美，花色豐富且花期長，深受消費者喜愛，為台灣四大出口農產品之一。然而面對國際市場之競爭，台灣必須在栽培管理及儲運管理做更大的改善，以提升品質與國際競爭力。根據前人研究(Kubota and Yoneda, 1993；Wang, 1995)指出蝴蝶蘭抽梗率與澱粉、蔗糖和葡萄糖等碳水化合物之變化有關；其中在碳水化合物的變化上，蝴蝶蘭容易受儲運條件和栽培環境的影響。但目前檢測碳水化合物含量使用的化學方法，不但耗時，且須破壞樣本，無法進行全面性檢測。除了碳水化合物對於蝴蝶蘭造成影響外，亦有研究(蔡等人，2013；李與王，1997)指出大苗階段的蝴蝶蘭葉片幾何外觀與開花性狀表現有相關性。本研究希望應用光譜影像技術，利用其即時、非破壞性且具有空間資訊之優點，用以檢測大白花蝴蝶蘭( Phalaenopsis Sogo Yukidian ‘V3’ )葉片之碳水化合物值和其幾何性狀參數，並建立模式以預測開花品質。本研究分為三部份（三次試驗），試驗一為建立鮮葉高光譜及粉末光譜檢量模式，研究結果顯示鮮葉高光譜蔗糖檢量模式以(1,8,8,1)的檢量模式為最佳，其RSQ為0.545，1-VR為0.396，SEC為2.020，SECV為2.553﹔而鮮葉高光譜澱粉檢量模式以(1,6,6,1)為最佳，其RSQ達至0.823，1-VR為0.758，SEC為3.351，SECV為4.257。粉末光譜蔗糖檢量模式，以(2,8,8,1)的檢量模式為最佳，其RSQ為0.930，1-VR為0.863，SEC為1.193，SECV為1.669﹔粉末光譜澱粉檢量模式為(2,8,8,1)為最佳，其RSQ為0.957，1-VR為0.886，SEC為2.037，SECV為3.332。試驗二為建立鮮葉光纖光譜檢量模式，結果顯示鮮葉光纖光譜之蔗糖檢量模式以(1,8,8,1)的檢量模式為最佳，其RSQ為0.916，1-VR為0.860，SEC為1.219，SECV為1.204﹔而鮮葉光纖光譜之澱粉檢量模式為(2,6,6,1)為最佳，RSQ為0.863，1-VR為0.784，SEC為1.475，SECV為2.714。試驗三為以鮮葉光纖光譜及高光譜影像建立開花品質預測模型，其中蝴蝶蘭樣本量測鮮葉光纖光譜值和鮮葉高光譜光譜值後栽培成花，量測所得的高光譜值亦可用鮮葉光纖光譜值所推測而得之蔗糖值和澱粉值建立檢量線。結果顯示鮮葉高光譜蔗糖檢量模式以(1,8,8,1)的檢量模式為最佳，其RSQ為0.562，1-VR為0.498，SEC為1.291，SECV為1.351﹔而鮮葉高光譜澱粉檢量模式為(1,10,10,1)為最佳，其RSQ達至0.748，1-VR為0.707，SEC為2.335，SECV為3.094。建立開花品質預測模型中，本研究選用倒傳遞類神經網路作為開花品質預測模型。其中模型又依碳水化合物來自於不同試驗，分為模型I和模型II。模型I以試驗二中的粉末光譜檢量模式所推論得到兩個代表性內部品質參數(第2片葉蔗糖值、第2片葉澱粉值)和三個代表性外部品質參數(第2片葉面積、第2片葉面積、第3片葉周長)預測開花品質，結果顯示單梗模型的平均準確率比不分梗數模型的平均準確率高約6%左右，其中單梗模型中又以輸入層參數含有澱粉值得平均準確力為最高，均方根誤差為0.7295朵；預測能力在±1朵時其平均準確率，訓練組可達81.42%、驗證組為75%。模型II以試驗三中的所量測鮮葉高光譜值代入試驗一中的鮮葉高光譜蔗糖檢量模式得的蔗糖值和三個代表性外部品質參數(第2片葉蔗糖值、第2片葉面積、第2片葉面積、第3片葉周長)以單梗植株預測開花品質，均方根誤差為1.3180朵；預測能力在±1.4朵時其平均準確率，訓練組可達81.92%、驗證組為74.49%。上述之結果顯示值得作為蘭花產業的應用性。

關鍵字

高光譜影像；近紅外光；蝴蝶蘭；碳水化合物；外觀性狀；開花品質；主成分分析；倒傳遞類神經網路

並列摘要

Phalaenopsis is one of the most valuable potted floriculture crops in the world; and it is one of the most important exported flowers from Taiwan. Facing the competition from the international markets, the management of cultivation and storage treatments should be enhanced to improve the quality and the competition ability. According to previous studies, content of carbohydrate in leaves is a crucial indicator to evaluate the flowering quality of Phalaneopsis. However, the current chemical methods to measure carbohydrate content in different parts of plants and thin section specimen examination of lateral buds are time consuming, laborious and destructive. In addition to interior quality (carbohydrate content), previous study also showed spiking rate was positively correlated with leaf area of Phalaneopsis. Instead of correlating leaf area with spiking rate, canonical correlation could be used to identify significant correlations between leaf traits and flower traits. Furthermore, it also showed that the second and third leaf traits of tested varieties had the highest correlation with flowering traits. Therefore, the spectral analysis and imaging technology was adopted in this study to analyze the carbohydrate in the leaves of Phalaneopsis. Spectral imaging technology is a rapid, non-destructive inspection method, and the distribution of carbohydrates in each leaf can be measured, especially the second and third leaf. In order to predict the quality of flowers, we proposed to use above significant parameters with backpropagation artificial neural network (BANN) to build a classifier based on the interior and exterior parameters for predicting the flowering quality of Phalaneopsis. This study was devided into three parts (three experiments). The purpose of first experiment was to build the fresh leaf hyperspectral imaging calibration models (FLHI) and powder form near infrared calibration models (PFNI). The FLHI results showed that sucrose content model under math treatment of (1,8,8,1) gave the results of RSQ=0.545, 1-VR=0.396, SEC=2.020 and SECV=2.553; and starch content model under math treatment of (1,6,6,1) gave RSQ=0.823, 1-VR=0.758, SEC=3.351 and SECV=4.257. The PFNI sucrose content under math treatment of (2,8,8,1) gave the results of RSQ=0.930, 1-VR=0.863, SEC=1.193 and SECV=1.669 while PFNI starch content under math treatment of (2,8,8,1) concluded RSQ=0.957, 1-VR=0.886, SEC=2.073 and SECV=3.332. The second experiment was to build the fresh leaf fiber optic spectral calibration models (FLFOS). It showed that the results of FLFOS sucrose content model under math treatment of (1,8,8,1) were RSQ=0.916, 1-VR=0.860, SEC=1.219 and SECV=1.204 while FLFOS starch content model under math treatment of (2,6,6,1) were RSQ=0.863, 1-VR=0.784, SEC=1.475 and SECV=2.714. The third experiment was using fresh leaf fiber optic spectral and hyperspectral imaging system technique to predict the flowering quality. In this experiment, we cultivated all of samples until they had buds and flowers. Futhermore, we also used the hyperspectral measurements to predict the sucrose and starch content. The results showed that FLHI sucrose content model of the third experiment under math treatment of (1,8,8,1) were RSQ=0.562, 1-VR=0.498, SEC=1.291 and SECV=1.351. while FLHI starch content model under math treatment of (1,10,10,1) were RSQ=0.748, 1-VR=0.707, SEC=2.335 and SECV=3.094. Regarding the development of BANN prediction models for flowering quality, we devided into model I and model II based on the interior parameters. Model I used FLFOS sucrose and starch contents with the exterior parameterts to predict the flowering quality. The results showed the model I with the number of one spike, having the starch content and sucrose conttent, had the best predicted ability. The root mean square (RMS) error was XXX, and averaged accuracy rate is 81.42% when applied to the training set, while the test set is 75%. In model II, FLHI sucrose content and the exterior parameters were used to predict the flowering quality. The RMS error was XXX, and averaged accuracy rate is 81.92% when applied to the training set, while the test set is 74.49%. All of results proved that the hyperspectral imaging system and near infrared technique can be used as a rapid and non-destructive methods for the prediction of flowering quality in the orchid production.

並列關鍵字

Hyperspectral Imaging ； Near Infrared ； Phalaenopsis ； Carbohydrate ； Exterior Characteristics ； Flowering Quality ； Principal Component Analysis ； Artificial Neural Network