During its growth, winter wheat (Triticum aestivum L.) can be impacted by multiple stresses involving fungal diseases that are responsible for high yield losses. Enhancing the breeding and the identification of resistant cultivars could be achieved by collecting automated and reliable information at the plant level. This study aims to estimate the severity of stripe rust (SR), brown rust (BR) and septoria tritici blotch (STB) in natural conditions and to highlight wavebands of interest, based on images acquired through a multispectral camera embedded on a ground-based platform. The severity of the three diseases has been assessed visually in an agronomic trial involving five wheat cultivars with or without fungicide treatment. An acquisition system using multispectral imagery covering the visible and near-infrared range has been set up at the canopy level. Based on spectral and textural features, estimations of area under disease progress curve (AUDPC) were performed by means of artificial neural networks (ANN) and partial least squares regression (PLSR). Supervised classification was also implemented by means of ANN. The ANN performed better at estimating disease severity with R2 of 0.72, 0.57 and 0.65 for STB, SR and BR respectively. Discrimination in two classes below or above 100 AUDPC reached an accuracy of 81% ( κ = 0.60) for STB. This study, which combined the effect of date, cultivar and multiple disease infections, managed to highlight a few wavebands for each disease and took a step further in the development of a machine vision-based approach for the characterisation of fungal diseases in natural conditions.

中文翻译：

---

利用多光谱图像的反射率和纹理特征对冬小麦中的小麦锈病、条锈病和褐锈病进行现场近端传感

在其生长过程中，冬小麦（Triticum aestivum L.）会受到多种胁迫的影响，这些胁迫涉及导致高产损失的真菌病害。通过在植物层面收集自动化和可靠的信息，可以加强抗性品种的育种和鉴定。本研究旨在估计自然条件下条锈病 (SR)、褐锈病 (BR) 和小麦锈斑病 (STB) 的严重程度，并根据通过嵌入地面的多光谱相机获取的图像突出显示感兴趣的波段。平台。这三种病害的严重程度已在一项农艺试验中进行了视觉评估，该试验涉及五个小麦品种，有或没有杀菌剂处理。已在冠层级别建立了使用覆盖可见光和近红外范围的多光谱图像的采集系统。基于光谱和纹理特征，通过人工神经网络 (ANN) 和偏最小二乘回归 (PLSR) 进行疾病进展曲线下面积 (AUDPC) 的估计。监督分类也是通过人工神经网络实现的。ANN 在估计疾病严重程度方面表现更好，STB、SR 和 BR 的 R2 分别为 0.72、0.57 和 0.65。对 STB 低于或高于 100 AUDPC 的两个类别的区分达到了 81% (κ = 0.60) 的准确度。这项研究结合了日期、栽培品种和多种疾病感染的影响，