Predicting bruise susceptibility in apples using Vis/SWNIR technique combined with ensemble learning

Abstract

Bruise susceptibility in fruits is an important indicator in evaluating risk factors for bruising caused by external factors. Prediction of the bruising susceptibility of fruit can provide useful information for proper postharvest handling and storage operations. In this study, visible and shortwave near-infrared (Vis/SWNIR) technique was used to develop nondestructive method for predicting the bruise susceptibility of apples. Vis/SWNIR spectra covering 400-1100 nm were collected for 300 ‘Golden Delicious’ apples over a time period of three weeks after harvest. A pendulum-like device was used to simulate impact bruise at three impact energy levels of 1.11 J, 0.66 J and 0.33 J. Bruise volumes were estimated from the digital images of the bruised apples by using the bruise thickness model. Three prediction models, i.e. partial least squares model (PLS), partial least squares model combined with successful projection algorithm (SPA-PLS), and selective ensemble learning based on feature selection (SELFS), for bruise susceptibility were developed for each impact energy level as well as for the pooled data. Compared with PLS and SPA-PLS model, SELFS gave the better prediction results for bruise susceptibility, with the correlation coefficient of Rp=0.800-0.886 for the prediction set, the root-mean-square error of 38.7-62.1 mm3/J for the prediction set (RMSEP), and the residual predictive deviation (RPD) of 1.78-2.14 for three impact energy level. For three impact energy levels, the RMSEP and RPD value obtained by SELFS model improved by 14.8%-20.0% and 15.0%-24.5% compared to PLS model, and 11.4%-21.2% and 11.5%-27.1% compared to SPA-PLS model, respectively. The SELFS model achieved relatively lower prediction accuracies for the pooled data, with the Rp values of 0.731, RMSEP of 85.46 mm3/J, and RPD of 1.46, which were also better than that of PLS model and SPA-PLS model. This research demonstrated that Vis/SWNIR technique combined with ensemble learning is promising technique for rapid assessment of bruise susceptibility of fruit, which would be useful for postharvest handling of fruit.