|ZHU, QIBING - Jiangnan University|
Submitted to: ASABE Annual International Meeting
Publication Type: Proceedings
Publication Acceptance Date: 7/11/2013
Publication Date: 7/21/2013
Citation: Mendoza, F., Lu, R., Zhu, Q. 2013. Assessing the sensitivity and robustness of prediction models for apple firmness using spectral scattering technique. In: Proceedings of the American Society of Agricultural and Biological Engineers Annual International Meeting, July 21-24, 2013, Kansas City, Missouri. Paper No. 13-1648152.
Interpretive Summary: Spectral scattering technique provides a means for quantifying light scattering features in food products, which can then be used to assess quality attributes like firmness of apples. The technique, however, relies on the calibration model relating scattering features to standard firmness measurements. The performance of the calibration model can be influenced by such factors as size of calibration samples (or variability of the dataset), data processing method, and harvest season. This research was conducted to quantitatively assess the effect of these factors on the calibration model for predicting the firmness of apples. More than 6,500 ‘Jonagold’, ‘Golden Delicious’, and ‘Delicious’ apples from two harvest seasons (2009 and 2010) were evaluated for firmness using a prototype laboratory spectral scattering system, followed by the standard destructive penetration measurement. A statistical method, called the mixed factorial design, was used to evaluate the effect of these factors and their interactions on the calibration models for firmness prediction. Results showed that the complexity or size of the calibration models, as measured by number of variables, increased with the dataset size. The two data processing methods (i.e., mean reflectance and continuous wavelet transform decomposition) did not have a significant effect on the firmness prediction accuracy. Overall, 400 samples were found to be adequate for establishing a calibration model for firmness prediction using spectral scattering technique. These findings can be used as a guide for selecting samples and data processing method in the development of a spectral scattering-based calibration model to predict fruit firmness.
Technical Abstract: Spectral scattering is useful for nondestructive sensing of fruit firmness. Prediction models, however, are typically built using multivariate statistical methods such as partial least squares regression (PLSR), whose performance generally depends on the characteristics of the data. The aim of this research was to evaluate the influence of range of variability for Magness-Taylor firmness data (i.e., number of samples at 100%, 80% and 60% of the total variability), preprocessing method [mean reflectance and continuous wavelet transform (CWT) decomposition], and harvest season (2009 and 2010) on the performance and robustness of the calibration models for predicting the firmness of 'Jonagold', ‘Golden Delicious', and 'Delicious' apples. A 3×2**2 mixed factorial experimental design with six replicates per run was used for assessing PLSR models for the spectral scattering data. The same prediction set of apple samples for each replicate was tested for the models. The main effects and interactions for the three variables, and their polynomial models were calculated based on the number of latent variables needed for the model building and the standard error of prediction (SEP). Overall results showed that models using large datasets needed a larger number of latent variables and produced smaller SEP values than those using smaller sample datasets for all apple cultivars. Models preprocessed by mean reflectance method resulted in a smaller number of variables than the models preprocessed by CWT. The results also demonstrated that increasing the number of samples used in the calibration set generally resulted in decreases in the SEP. It is thus recommended that number of calibration samples should be determined according to the prediction accuracy required. This study found that 400 apple samples were appropriate to establish calibration models for firmness with lower prediction errors.