Sparse regression for selecting fluorescence wavelengths for accurate prediction of food properties

• Sparse regression models were applied on fluorescent spectra for predicting different quality parameters.
• The models were tested for predicting the count of viable bacteria on the surface of porcine meat.
• The models were tested for predicting freshness of frozen fish.
• The performance of the model was compared with the traditional PLS and MLR models.
• The proposed sparse models provide efficient variable selection.

This paper tested various regression models (PLS, Ridge, Lasso, and sparse group Lasso) to select the appropriate fluorescence wavelengths/variables in excitation–emission matrices (EEMs) to improve the prediction of food identities. A framework using sparse models (the Lasso and sparse group Lasso) was proposed and compared with the conventional models. These sparse regression techniques can simultaneously achieve the ideal design of the estimator and select the most effective feature-related wavelengths. The experimental results showed that the proposed framework provided high prediction accuracy in selecting variables for accurate prediction of fish freshness and meat safety. Specifically, in case of predicting fish freshness, the sparse group Lasso regression had a determination coefficient R2 of 0.790 with 493 EEM variables while the standard PLS regression had R2 of 0.748 using all 1054 EEM variables.