2019 ASHS Annual Conference

In an apple breeding program, development of cultivars with desirable eating quality and postharvest characteristics is of paramount importance. To optimize fruit quality and harvest time of seedling selections, fruits are destructively evaluated during the season. This presents a challenge when young seedling trees do not bear sufficient fruits for destructive sampling. A non-destructive alternative would allow prediction of fruit quality indices regardless of fruit count, and increase selection efficiency. In recent years, near infrared (NIR; λ = 750 – 2500 nm) spectroscopy has garnered interest in the food and agribusiness industries, enabling facile analysis of parameters associated with traditional destructive assessments. Previous reports of NIR use in fruit crops were based mainly on one to two cultivars, instead of the large genetic variability in a typical breeding program. The objective of this study was to evaluate the accuracy of NIR prediction for postharvest traits of 20 selections for routine application in the Washington State University apple breeding program (WABP).

In 2015 and 2016, fruits from 20 advanced seedling selections at three orchard sites were harvested and stored for two months at 2 °C. After storage, fruits were evaluated using destructive analytical tests (dry matter concentration [DMC], soluble solids concentration [SSC], titratable acidity [TA] and firmness) as well as an NIR instrument (Felix F-750). NIR measurements and the corresponding trait values were used to build partial least squares regression models. Subsequently, regression coefficients in a training model were used to predict the trait values of another set (i.e., validation). Deviations in the validation set between the predicted values and the instrumental values were reported as prediction accuracy metrics, such as ratio of prediction to deviation (RPD) and correlation (ρ).

Wide variation in values of the four postharvest traits was observed among the 20 advanced selections across three locations in both years. Although training models failed to build for TA and firmness, reasonably robust training models (R² > 0.80) were developed for DMC and SSC in both years. Validation of DMC exhibited high prediction accuracy (RPD > 2.5; ρ > 0.90), while validation of SSC exhibited moderate accuracy (RPD ≈ 1.5; ρ ≈ 0.70). Current work is underway to: (1) validate within- and between-years sets, (2) assess the accuracy of longitudinal forecasting, and (3) optimize a training model with robust prediction accuracy based on all 20 selections for use in the breeding program.