無人機多光譜影像分析技術改善水稻白葉枯病之抗病育種

Abstract

本研究利用無人機多光譜影像分析技術評估水稻白葉枯病之抗病程度，希望提供具客觀性的田間快篩分級系統，節省人力、加速國內抗病育種的進展。實驗病圃位於彰化大村鄉的台中農業改良場，材料為193個水稻試驗品系與5個栽培品種，共種植兩重複。採用剪葉法進行病菌接種。無人機影像資料包含108年一、二期作。使用主效應分析定位稻穗位置，並採用K-均值群聚演算法 (K-means clustering)，將影像分成原始影像、分離土壤後的植株影像和分離土壤及稻穗的葉部影像，最後從影像中萃取42個植生指數 (Vegetation indices) 進行分析。迴歸分析顯示單品種、多品種的罹病率與植生指數具有高相關性，broad-band chlorophyll vegetation index (CVI)與plant senescence reflectance index (PSRI) 指數具有潛力建造罹病程度預測模型，但由於測試資料呈現不平均的分布，造成模型預測的低準確率。因此將五個罹病程度劃分成「具抗性」與「不具抗性」兩個組別，並使用k-最近鄰分類演算法 (k-nearest neighbor, KNN) 建立新模型。儘管模型的預測結果仍受到不平均罹病等級資料的影響，KNN模型仍展示出某些植生指數具有潛力分辨具抗性與不具抗性的水稻。若後續研究採用罹病等級較平均的資料建模，模型的預測結果將更具說服力。結果也顯示不同影像分割的資料並不會影響模型的表現，因此建議後續研究省略影像分割的步驟。本研究的初步結果提供建議給其他欲利用無人機影像建立罹病分類模型的研究者，關於材料準備、影像處理、模型建立的相關意見。相信未來若持續收集影像資料，配合嚴謹的田間調查，確保各罹病等級的資料平均散佈於田中，必可突破不同水稻品種光譜差異的限制，建立預測模型客觀篩選抗病品系的水稻。

The purpose of this study is to develop an analytical system using unmanned aerial vehicle (UAV) multispectral imagery. The system aims for quickly differentiating resistant and non-resistant varieties to bacterial leaf blight of rice (Oryza sativa L.). The experimental site was located in the Taichung District Agricultural Research and Extension Station. 193 breeding lines and 5 cultivars with two replicates were tested. Rice was inoculated by the leaf-clipping method. The aerial image data were taken at the first and second crop seasons in 2019. The principal component analysis was used to locate the panicle in the image, and K-means clustering was used to perform the image segmentation. Image data was divided into non-segmented data, vegetation data, and leaf data. Forty-two vegetation indices (VIs) were extracted from the images. Regression analysis between the VIs and disease severity of the single and multiple cultivars had high R2adjusted. However, multi-cultivar regression models showed low prediction accuracy for the five resistance levels of rice. Thus the five resistance levels were reclassified into the “resistant” group and “susceptible” group, and the k-nearest neighbor (KNN) algorithm was adopted to test new models. The KNN models based on VIs showed the potential to differentiate the two resistant groups, but further studies should be tested with a more balanced dataset. According to the results, image segmentation may be unnecessary to build the disease classification models. In this study, the results can give recommendations on preparing materials, image processing, and model selection for the relevant studies. With more data collected in the future, the barrier of differentiation between different levels of disease from multi-varieties can be overcome. Eventually, breeders can apply the UAV multispectral imaging system to select resistant varieties efficiently in the field.