蝴蝶蘭開花品質之光譜預測模式

Abstract

蝴蝶蘭是台灣重要的外銷花卉，每年為台灣的花卉出口提供不少產值，然而面對荷蘭與中國大陸等國家之競爭，如何提升產業競爭力成為一重要課題。由於蝴蝶蘭之開花品質與其大苗葉片之碳水化合物及外觀特徵有相關性，若能於蝴蝶蘭催花前便藉由葉片內部碳水化合物含量與外觀特徵得知其未來開花品質，勢必能增進台灣蝴蝶蘭產業之實力。本研究以大白花蝴蝶蘭'V3'品種為實驗對象，使用手持式光譜儀（DLP NIRscan Nano）以近紅外光譜技術建立蝴蝶蘭葉片內部成分檢量模型，再以此模型之結果建立蝴蝶蘭開花品質預測模型；本研究也使用高光譜系統(自行研發)，以光譜影像技術建立蝴蝶蘭葉片內部成分檢量模型，並以此模型之結果搭配光譜影像擷取之葉片外部特徵進行蝴蝶蘭開花品質模型之建立，最後比較兩種設備所建立之模型其優缺點與應用性。 本研究是以MPLSR對蝴蝶蘭葉片中六種內部成分含量進行模型建立，分別為葡萄糖、果糖、蔗糖、總可溶性糖、澱粉、總碳水化合物。使用手持式光譜儀所建立之模型，以預測蔗糖含量之結果最佳，其RSQ值可達0.73，預測總可溶性醣與總碳水化合物含量之結果次之，但其RSQ值皆高於0.55，預測葡萄糖、果糖及澱粉之結果則稍差，其RSQ值皆低於0.43。至於使用高光譜系統對葉片內部成分所建立之預測模型，蔗糖含量之預測結果其RSQ值可達0.81，總可溶性醣、總碳水化合物含量之預測結果其RSQ值也都高於0.69，雖然澱粉含量之預測結果RSQ值為0.75，但誤差稍大，而葡萄糖與果糖含量之預測結果RSQ雖然只有0.47和0.54，但SEC皆小於10.3 mg/g。比較兩種設備對蝴蝶蘭葉片各內部成分所建立之預測模型，高光譜系統之預測結果皆優於手持式光譜儀。 本研究以PLSDA、SVM以及ANN三種方法對蝴蝶蘭的開花品質進行建模分析，其中PLSDA與SVM用於蝴蝶蘭高品質與中低品質兩種級別之分類模式建立，ANN則是用於蝴蝶蘭開花朵數之預測。比較兩種設備所建立之開花品質模型，高光譜系統以PLSDA與SVM兩種方法建立之模型皆較手持式光譜儀為佳，其驗證組預測正確率皆超過六成；而使用ANN建模之結果，手持式光譜儀預測總開花朵數之誤差為0.87朵，高光譜系統預測總開花朵數之誤差為0.85朵，雖然高光譜系統之預測能力稍佳，但兩種設備在預測開花朵數之誤差皆小於1朵，具備不錯的預測能力。 本研究結果顯示，以光譜技術預測蝴蝶蘭內部成分含量與開花品質是可行的。而高光譜系統雖然性能優於手持式光譜儀，但其價格高且不若手持式光譜儀便利，未來若要將光譜技術應用於產業界，可朝多光譜系統或是優化手持式光譜儀預測模型之方向進行。

Phalaenopsis is an important exported flowers in Taiwan and it provides a big output value every year. Facing the competetion from Netherlands and China, however, how to enhance the competitiveness of the flower industry becomes an important issue. Because the flowering quality of Phalaenopsis was reported to be correlated to its leaves’ carbohydrates content and external traits, our Phalaenopsis industry can be strengthened if we can predict the flowewing quality by using these internal and external traits. Phalaenopsis Sogo Yukidian 'V3' was used as the experimental samples in this research, and this study used near infrared technique with hand-held spectrometer (DLP NIRscan Nano) to build calibration models of interior contents in Phalaenopsis leaves. Then, we used the results from the models to predict the flowering quality. We also used hyperspectral imaging system (Self-developed system in this study) technique to build another interior content calibration model and combined the external traits of leaves from hyperspectral imaging to predict flowering quality. Finally, we compared the models’ advantages and applications between these two devices. This study used MPLSR model to predict six interior contents in Phalaenopsis leaves, incuding glucrose, fructose, sucrose, total soluble sugar, starch, and total carbohydrates. The results from hand-held spectrometer showed that sucrose content model with RSQ = 0.73 was the best; total soluble sugar content model and total carbohydrates content model were secondly, and both of them gave the results of RSQ above 0.55; the results of glucrose model, fructose model and starch model were slightly worse, all of them gave the result of RSQ below 0.43. As for the results from hyperspectral imaging system, sucrose content model gave RSQ = 0.81; and total soluble sugar content model and total carbohydrates content model gave the results of RSQ above 0.69; although starch content model gave RSQ = 0.75, the predict error was a little high; glucrose model and fructose model gave the result of RSQ = 0.47 and RSQ = 0.54, but SEC were both below 10.3 mg/g. Comparing with the interior contents models between these two devices, the results from hyperspectral imaging system were all better than those from hand-held spectrometer. This study used three kinds of methods to predict the flowering quality including PLSDA, SVM and ANN. Among them, PLSDA and SVM were used to discriminate high and medium-low quality levels of Phalaenopsis; and ANN was used to predict the number of Phalaenopsis flowers. Comparing the results of PLSDA model and SVM model between two devices, those from hyperspectral imaging system were better than those of hand-held spectrometer, which prediction accuracy rate of validation group were higher than 60%. Comparing the results of ANN models, the error that using hand-held spectrometer to predict the number of flowers was 0.87, while the error that using hyperspectral imaging system was 0.85. Although the prediction ability from hyperspectral imaging system was a little bit better, the errors from these two devices were less than 1 flower, and it indicated that both devices had good results. The research proved that using spectral technique to predict the interior contents in Phalaenopsis leaves or flowering quality were both feasible. Although the performance of hyperspectral imaging system is better than that of hand-held spectrometer, it was expensive and not as convenient as hand-held spectrometer. We can adopt some other ways like using multispectral imaging system or optimize the performance of hand-held spectrometer if we want to apply spectral technique to the Phalaenopsis industry.