Evaluation of Critical Agro-meteorological Parameters Affecting Wheat Quality on the Canadian Prairies using Partial Least Squares (PLS) Regression

Spring wheat is one of the major crops grown on the Canadian prairies, of which about 80% is exported. However, wheat quality fluctuation is a major issue for both local and international wheat customers with expectations of consistent quality to meet the needs of the baking industry. Besides genetics and management, the environment especially the weather, has a significant impact on wheat quality. The Canadian prairies experience a wide range of weather conditions within and between growing seasons. The objective of the current study was to identify agro-meteorological factors that impact spring wheat quality using partial least squares (PLS) regression. PLS regression is particularly suited when the matrix of predictors has more variables than observations and there is multicollinearity among the predictor variables (which is usually the case with weather variables). The study utilised detailed crop and weather data collected from 2003 through 2006 from a series of wheat trials conducted across Saskatchewan and Manitoba. Fifty three (53) agro-meteorological indices (predictor variables) categorised into (i) water supply, (ii) water demand, (iii) water balance and (iv) water use were derived from the weather data and used in the PLS regression. Wheat quality characteristics (response variables) including protein content, farinograph absorption, dough development time and loaf-volume for two varieties (AC Berrie and Superb) were analysed using standard methods. Wheat quality data from AC Berrie were used to develop the PLS models, which were in-turn used to predict wheat quality characteristics for Superb. A prior student t-test analysis had shown that there was no significant difference (p>0.05) among the quality characteristics of the two cultivars.

Results showed that three latent vectors (components) explained 83% of the variability in wheat protein content and 86% of the variability in the predictor variables; 80% of the variability in dough development time and 85% of the variability in the predictor variables; and 69% of variability in loaf-volume and 85% of the variability in the predictor variables. Meanwhile, four latent vectors (components) explained 79% of the variability in farinograph absorption and 89% of the variability in the predictor variables. When the developed models were used to predict quality parameters for the variety Superb, a student t-test showed that the overall average (mean) of the predicted values were not statistically different (p>0.05) from the overall average of the observed values for all the wheat quality parameters except for farinograph absorption. The predicted values correlated very well with the observed values with R² values of 0.96 (p<0.0001) for protein content, 0.75 (p<0.0001) for dough development time, 0.69 (p<0.001) for loaf-volume and 0.83 (p<0.0001) for farinograph absorption. The RMSE values were 0.32, 0.85, 59.15 and 0.71, respectively. The models will be further tested and validated using data that is currently being collected across the Canadian prairies in a project that started during the 2015 cropping season.