Mohammad Mokhlesur Rahman, Plant Pathology, Kansas State University, Manhattan, KS, Jared Crain, Department of Plant Pathology, Kansas State University, Manhattan, KS, Ravi Prakash Singh, Wheat Improvement & Rust Research Global Wheat Program, International Maize and Wheat Improvement Center (CIMMYT), Texcoco, Mexico and Jesse Poland, Department of Plant Pathology and Department of Agronomy, Kansas State University, Manhattan, KS
A primary selection target for wheat (Triticum aestivum) improvement is grain yield. Selection for yield is hampered by the expense, equipment and space needed for large field trials along with low heritability, environmental variance, and the time needed to obtain multi-year assessments. Traits such as spectral reflectance and canopy temperature are frequently correlated to grain yield and possible to measure many times throughout the growing season. While the collection of this proximal sensing data on breeding nurseries is becoming more routine, the analysis of multi-temporal, multi-trait data is challenging. The objective of this study was to monitor plant growth and estimate grain yield using temporal phenotypic hand-measurements in wheat yield trials under extreme heat stress that is common to BangladeshÕs growing environments. We analyzed normalized difference vegetation index (NDVI) and canopy temperature (CT) measurements collected throughout the growing season in 660 advanced breeding lines. The lines were part of the International Maize and Wheat Improvement CenterÕs Southeast Asia breeding program grown in Jamalpur, Bangladesh. To optimize utilization of the dataset, we explored several variable reduction and regularization techniques. We then modeled each one of these approaches using a cross-fold validation approach to predict grain yield. We calculated the correlations, using the observed and predicted grain yield values for univariate and multivariate models with accuracies of 0.50 and 0.71 respectively. The multivariate models provided higher prediction accuracies than the univariate models. We completed the data analysis using penalized models and examined their prediction accuracies. The ridge regression model resulted in prediction accuracy of 0.87. We used the ridge regression model for yield prediction and selected 53 lines superior to local check variety. Our results suggest that optimized prediction models can provide additional information to enhance the wheat breeding program in Bangladesh by leading to more rapid and accurate selections.