利用機器視覺技術檢測水稻徒長病

Abstract

本研究旨在以量化病徵的方式，觀察受徒長病感染的水稻植株之病徵，並提出利用機器視覺檢測植物病害的方法，來辨別出水稻徒長病。水稻徒長病是臺灣水稻古老病害之一，主要由病原真菌Fusarium fujikuroi所引起，水稻受感染後不易結穗，甚至在插秧前植株便已凋萎枯死，嚴重影響了水稻的收成。主要的傳播途徑是藉由帶菌的種子，播種後會成為二次感染源，在土壤中的感染源可感染插秧後的健康植株，因此必須在水稻生長早期進行檢測。然而，罹病植株其病徵變化複雜，不同的品種其變化也有所不同，所以本研究在建立檢測模型前，先對不同的水稻品種，作病徵觀察，以定義的量化特徵觀察染病與健康植株間的差異，這可以避免肉眼觀察的主觀性，且能以較精確的數值記錄；之後，以這些特徵，建立分類染病與健康植株之檢測模型。本研究所使用的水稻植株，皆在經過接種F. fujikuroi過程後種植3週，然後將健康與接種的植株取像，並得到量化的形狀與顏色特徵。在病徵觀察中，量化具不同感病性的品種之形狀特徵，從觀察的結果顯示，對於不同的水稻品種，接種的植株其植株寬、植株高寬比、第二莖節長及葉間夾角，與健康植株間的差異有較大的變化；而在徒長病檢測中，使用支持向量機，對臺南11號與豐錦二品種，進行分辨植株的健康狀況，也使用了基因演算法選出重要的特徵及最佳的分類器訓練參數，結果顯示平均準確度為87.92%，而染病植株平均正確預測度為91.77%。

This work aims to observe symptoms of Bakanae disease using quantified traits and propose an approach to detect Bakanae disease using machine vision. Bakanae disease, or “foolish seedling,” is a seed-borne disease of rice (Oryza sativa L.). Infected plants can yield empty panicles or perish, resulting in a loss of grain yield. The disease occurs most frequently when contaminated seeds are used. Once the seeds are contaminated, the pathogen Fusarium fujikuroi spreads in the field, and provides an initial site for secondary infection. Therefore, early detection of Bakanae disease is essential for disease control. Also, the symptoms of the disease are complex, and can vary with cultivars. Hence, morphological symptoms in rice of different cultivars were first observed and analyzed before developing models for distinguishing infected and healthy seedlings. In this study, rice seeds were inoculated with a conidial suspension of F. fujikuroi, and then cultivated in an incubator for 3 weeks. The inoculated and control seedlings were photographed for quantifying morphological and color traits. In symptom observation, the morphological traits of eight cultivars of different disease severity (DS) levels were quantified. The results indicated that large variation between the healthy and inoculated seedlings were observed in seedling width, aspect ratio, second-internode length and leaf angle. In disease detection, support vector machine (SVM) classifiers were developed for distinguishing the inoculated and healthy seedlings for the cultivars, Tainan 11 and Toyonishiki. A genetic algorithm was used for selecting essential traits and optimal model parameters for the SVM classifiers. The proposed approach distinguished inoculated and healthy seedlings with an accuracy of 87.92% and a negative predictive value of 91.77%.