應用高光譜影像技術於青江菜黑斑病與草莓葉部炭疽病之偵測與分析

Abstract

病害侵入植物後，會影響植物成長與外表，在農產品部分則會降低產量，進而造成經濟損失。評估應用高光譜影像技術辨識青江菜黑斑病與草莓葉部炭疽病的潛力是本研究的主旨。高光譜影像科技是一項結合在可見光與近紅外光的數位影像與光譜資訊之非破壞性量測技術，透過高光譜影像可以獲取植物葉部許多生理與化學特性等，諸如葉綠素與水分的變化。本研究以逐步判別分析法 (Stepwise Discriminant Analysis, SDA) 與彈性網路 (Elastic Net, EN) 兩種特徵選取演算法找出具有判別植物病害的重要波長，並且將這些重要的波長以線性判別分析法建立分類器，判別植物是否健康。後續應用假彩色於高光譜影像上，透過顏色的不同表示植物葉部感染病害的位置與感染的範圍。 本研究透過反射率、反射率之一階導數與二階導數，比較不同波長數量組合的模型之成效。SDA在光譜反射率中選取12波長組合的模型，辨識青江菜是否有黑斑病的正確率為98.3%。同樣由SDA在光譜反射率選取4波長組合的模型辨識青江菜的健康、發病前與發病三類別，其正確率達77.4%。辨識草莓葉部是否有炭疽病，將反射率進行一階導數計算後，由SDA挑選2波長組合的模型，辨識正確率達98.5%。而在辨識草莓葉部的健康、發病前與發病的部分，SDA挑選出反射率中16個波長組合的模型為最理想，其正確率可達89.3%。 實驗與分析結果說明應用機器學習建立少量波長組合的模型，可以進行植物病害的辨識。未來可以透過多光譜影像技術來建立一套自動非破壞性植物病害偵測系統，以利於監控植物病害。

The growth and appearance of plant are affected and destroyed when infected by pathogens. Plant diseases not only decrease the production rate, but also cause economic losses. The thesis aims to evaluate the potential of strawberry foliar Anthracnose disease and Bok Choy black spot disease identification by using hyperspectral images. Hyperspectral imaging is a non-destructive measurement technique that combines digital imaging and spectroscopy in visible and near infrared wavelength regions. The technique can provide physical and chemical information of leaves simultaneously, such as the change of chlorophyll and water in leaves. In this study, two feature selection algorithms, stepwise discriminant analysis (SDA) and elastic net (EN), were applied to define the significant wavelengths for discriminating plant diseases, and then employed the defined to establish a linear discriminant analysis classifier. Moreover, pseudo color image were also used to represent the infected locations and regions on the leaf. In this approach, the reflectivity of the selected wavelengths, and the first and second derivative of the reflectivity were employed as the features. The recognition performance of the features was compared. The accuracy of classifying healthy Bok Choy leaves and infected leaves is 98.3%, and the classification model was the built with 12 wavelengths selected by SDA. To discriminate healthy, incubation and symptomatic of Bok Choy leaves, the accuracy is 77.4% with 4 selected wavelengths. For strawberry, the accuracy of the classifying healthy and infected leaves reaches to 98.5%, and the classification model was established with 2 wavelengths selected by SDA according to the first derivative. For classifying three different Anthracnose infection status (healthy, incubation and symptomatic), the accuracy is 89.3%, and 16 wavelengths defined by SDA were applied to build the model. The experimental results imply that the use of machine learning algorithm is able to discriminate plant disease by few and significant wavelengths. Furthermore, the proposed method and procedure can be applied to establish an automatic, non-destructive plant disease detection system using multispectral imaging for monitoring plant disease.