ABSTRACT

Water deficit is the most limiting factor for wheat production, so wheat-breeding programmes are currently focused on developing high-performance genotypes under such conditions. Carbon isotope discrimination (∆¹³C) in grains is a trait closely related to yield and stress tolerance. However, conventional measurement of ∆¹³C is expensive, limiting its widespread use for genotype selection in breeding programmes. Predicting ∆¹³C through remote sensing could be useful for large-scale phenotyping. A set of 384 cultivars and advanced lines of spring bread wheat (Triticum aestivum L.) was grown under contrasting water conditions during two seasons. Grain yield (GY) and the ∆¹³C of grains were obtained at the end of both seasons, and canopy reflectance measurements were taken at anthesis and grain filling. Hyperspectral canopy reflectance was used to estimate GY and ∆¹³C through Multilinear Regression Analysis (MRL) considering wavelength selection using a Genetic Algorithm (GA), spectral reflectance indices (SRIs), Partial Least Square Regression (PLSR), Support Vector Regression (SVR), Random Forest (RF) and Artificial Neural Networks (ANN). The best models of both GY and ∆¹³C explained 78% and 60% of data variability, respectively. Additionally, the MRL models showed higher prediction rates than SRIs and similar or slightly lower rates, in most cases, than multivariate regression models, but required only 4–9 wavelengths instead of the full hyperspectral data used to develop the regression models. The use of canopy spectral reflectance data and MRL models to predict GY and Δ¹³C via GA for selection of the reflectance wavelengths could be a practical tool for genotype selection in wheat breeding systems.

摘要

缺水是小麦生产的最大限制因素，因此，小麦育种计划目前的重点是在这种条件下开发高性能基因型。谷物中的碳同位素歧视（∆ ¹³ C）是与产量和胁迫耐性密切相关的性状。然而，Δ的常规测量¹³ C是昂贵，限制了在育种计划其基因型选择广泛使用。通过遥感预测∆ ¹³ C对于大规模表型分析可能很有用。在两个季节中，在不同的水分条件下，种植了384个品种和高级品系的春面包小麦（Triticum aestivum L.）。谷物产量（GY）和∆ ¹³在两个季节结束时都获得了谷物的碳，在花期和灌浆期进行了冠层反射率测量。考虑到使用遗传算法（GA），光谱反射率指数（SRI），偏最小二乘回归（PLSR），支持向量回归（SVR）的波长选择，使用高光谱冠层反射率通过多线性回归分析（MRL）来估计GY和∆ ¹³ C ），随机森林（RF）和人工神经网络（ANN）。GY和∆ ¹³的最佳模型C分别解释了数据变异性的78％和60％。此外，在大多数情况下，MRL模型显示的预测率高于SRI，而与多变量回归模型相比，则显示出相似或略低的预测率，但仅需要4–9个波长，而不需要用于开发回归模型的完整高光谱数据。使用冠层的光谱反射率数据和MRL模型来预测GY和Δ ¹³经由GA下的反射率波长的选择可能是在小麦育种系统基因型选择的实用工具。