First and second authors: Department of Plant Pathology, Kansas State University, Manhattan 66506; and third and fourth authors: Department of Plant Pathology, The Ohio State University, Ohio Agricultural Research and Development Center, Wooster 44691.
Phytopathology. 2019 Jan;109(1):96-110. doi: 10.1094/PHYTO-11-17-0386-R. Epub 2018 Dec 3.
In past efforts, input weather variables for Fusarium head blight (FHB) prediction models in the United States were identified after following some version of the window-pane algorithm, which discretizes a continuous weather time series into fixed-length windows before searching for summary variables associated with FHB risk. Functional data analysis, on the other hand, reconstructs the assumed continuous process (represented by a series of recorded weather data) by using smoothing functions, and is an alternative way of working with time series data with respect to FHB risk. Our objective was to functionally model weather-based time series data linked to 865 observations of FHB (covering 16 states and 31 years in total), classified as epidemics (FHB disease index ≥ 10%) and nonepidemics (FHB disease index < 10%). Altogether, 94 different time series variables were modeled by penalized cubic B-splines for the smoothing function, from 120 days pre-anthesis to 20 days post-anthesis. Functional mean curves, standard deviations, and first derivatives were plotted for FHB epidemics relative to nonepidemics. Function-on-scalar regressions assessed the temporal trends of the magnitude and significance of the mean difference between functionally represented weather time series associated with FHB epidemics and nonepidemics. The mean functional weather-variable curve for epidemics started to deviate, in general, from that for nonepidemics as early as 40 days pre-anthesis for several weather variables. The greatest deviations were often near anthesis, the period of maximum susceptibility of wheat to FHB-causing fungi. The most consistent separations between the mean functional curves were seen with the daily averages of moisture-related variables (such as average relative humidity) and with variables summarizing the daily variation in temperature (as opposed to the daily mean). Functional data analysis was useful for extending our knowledge of relationships between weather variables and FHB epidemics.
在过去的研究中，美国的镰刀菌顶腐病（FHB）预测模型的输入气象变量是按照窗格算法的某个版本确定的，该算法将连续的气象时间序列离散化为固定长度的窗口，然后搜索与 FHB 风险相关的总结变量。另一方面，功能数据分析通过平滑函数重建假设的连续过程（由一系列记录的气象数据表示），是一种处理与 FHB 风险相关的时间序列数据的替代方法。我们的目标是对与 865 个 FHB 观测（涵盖 16 个州和 31 年）相关的基于天气的时间序列数据进行功能建模，这些观测分为流行（FHB 疾病指数≥10%）和非流行（FHB 疾病指数<10%）两类。总共对 94 个不同的时间序列变量进行了建模，使用惩罚三次 B 样条作为平滑函数，从开花前 120 天到开花后 20 天。为流行和非流行绘制了功能均值曲线、标准差和一阶导数。功能标量回归评估了与 FHB 流行和非流行相关的功能表示天气时间序列的幅度和均值差异的时间趋势。流行的功能平均天气变量曲线通常从开花前 40 天开始与非流行的曲线开始偏离，对于几个天气变量而言。最大的偏离通常发生在开花期附近，这是小麦对镰刀菌真菌最易感染的时期。在功能平均曲线之间最一致的分离是在与水分相关的变量（如平均相对湿度）的日平均值以及总结温度日变化的变量（与日平均值相反）中看到的。功能数据分析有助于扩展我们对天气变量与 FHB 流行之间关系的认识。
