應用高光譜影像法偵測草莓炭疽病之發生

Abstract

草莓在台灣是一種重要的經濟作物，年產值可高達 16 億新台幣，屬於一種宿根性作物，可存活 2~3 年，然其在台灣病蟲害眾多，尤以炭疽病為害甚劇，造成台灣草莓種植需年年更新母株。本試驗針對草莓育苗期間面臨主要病害-炭疽病，利用高光譜影像偵測系統，找出可辨識炭疽病的特徵波段，針對草莓葉部炭疽病進行快速且非破壞性的早期偵測，以協助炭疽病的診斷。最終期望達到自動化偵測草莓炭疽病，且在肉眼未見病徵前可利用光譜影像診斷出炭疽病潛伏感染，以協助草莓育苗業者由源頭控管母株及種苗的健康程度，降低炭疽病在苗期危害的嚴重度，將有利於草莓健康種苗供應鏈的建立。本文採用利用逐步迴歸分析進行簡易降維，再利用簡單迴歸分析評估草莓炭疽病的病害程度，將模型分為發病害與發病前潛伏感染兩階段，發病期與潛伏感染期間分別以發病比例與接種病原菌後天數作為發病程度的依據，在評估模型的 R2、RMSEC 值後，其模型平均的 R2 和 RMSEC 分別為 R2 = 0.79、0.91; RMSEC = 0.11、0.53，另外計算偵測極限作為辨識健康與生病樣本的門檻值進行驗證，驗證結果中發病模型的健康樣本辨識率為 87%，而發病樣本的辨識率為 72%，而評估潛伏感染期間的模型在驗證結果中，健康樣本的辨識率為 72%，潛伏感染樣本的辨識率為 71%，兩模型中被辨識為發病或潛伏感染的樣本可再經由檢量模型進一步評估其發病的程度，此研究結果證實高光譜影像分析配合逐步迴歸分析對於草莓炭疽病的定量與田間防治是具有潛力的工具之一。

Strawberry is an economically important herbaceous plant and the annual output value of strawberry industry is up to 1.6 billion NT dollars in Taiwan. Strawberry can subsist two to three years in field; however, there are many diseases during its cultivation, especially, anthracnose, which is the most devastating fungal disease that threatens in strawberry and causes farmers to replant new plants every year. Hyperspectral imaging has the potential to extract integrated spatial and spectral information related to the plant's functional dynamics regarding both structure and physiology. The objective of this research is to detect strawberry anthracnose with the hyperspectral imaging system, which is expected to find out the characteristic bands of anthracnose identification and detect disease as early as possible. Finally, we hope it can provide novel and non-destructive tools to diagnosis plant disease severity more practically and efficiently. Therefore, strawberry farmers could profit greatly from this technology. In this study, we use a variety of methods to carry out simple dimensionality reduction. We hope to find the characteristic bands that are significant for distinguishing anthracnose. After evaluating the R2 and RMSEC values of various models, we will infer that a better model can be established by stepwise regression analysis of the whole band from 470-950 nm. The model is divided into two stages: disease and latent infection. The average R2 and RMSEC of the two models are R2 = 0.79, 0.91, respectively and RMSEC = 0.11, 0.53, respectively. And then using limited of detection value as a threshold for distinguishing between healthy and sick samples. The diseased model provided the detection of healthy samples with accuracy of 87% and symptomatic samples with accuracy of 72%, while the model for evaluating latent infection provided the detection of healthy samples with accuracy of 72% and symptomatic samples with accuracy of 71%. The samples classified as infected or latent infections in the two models were further evaluated the disease severity by the model. This study confirms that stepwise regression analysis can be one of the classifier tools in hyperspectral imaging analysis for quantitative analysis and field control of strawberry anthracnose.