Predicción de las propiedades fisicoquímicas del aguaymanto (Physalis peruviana) usando imágenes hiperespectrales y machine learning
No Thumbnail Available
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad Nacional de Trujillo
Abstract
El aguaymanto (Physalis peruviana) es un alimento oriundo de los Andes de nuestro país. Debido a su rápido crecimiento en la producción están buscando tecnologías que puedan controlar la calidad interna con un alto grado de confiabilidad, rapidez y que no sean destructivas. En este trabajo se propuso utilizar un sistema de imágenes hiperespectrales de infrarrojo cercano (NIR-HSI) y quimiometría para predecir el contenido de vitamina C, sólidos solubles (°Brix) % acidez titulable y la textura de aguaymantos. Se recolectaron muestras de diferentes regiones locales en Perú (Cajamarca, Chachapoyas, Huaraz y Otuzco), luego se tomaron imágenes con una cámara NIR-HSI en un rango de 953 a 1703 nm en 3 diferentes orientaciones (zona del cáliz, ápice y ecuatorial). Se entrenaron y validaron modelos quimiométricos predictivos: Regresión de Mínimos Cuadrados Parciales (PLSR) y la Regresión de Máquina de Vector de Soporte (SVMR), usando el rango de longitudes de onda completa y seleccionadas para predecir las propiedades fisicoquímicas. Los métodos usados para seleccionar variables importantes fueron: recursiva PLS (rPLS), intervalo PLS (iPLS), algoritmo genético (GA) y covarianza entre variables (Covsel). Los modelos de regresión basados en la longitud de onda completa mostraron a la orientación de la zona del cáliz (2) como la que obtuvo mejores indicadores resultando para PLSR un R_P^2 entre 0.43 - 0.76 y valores de RPD entre 1.54 - 2.25, mientras que para SVMR presentaron un R_P^2entre 0.37 - 0.82 y valores de RPD entre 1.54 - 2.30.Resultando la vitamina C como la propiedad que presenta los resultados más altos, con un R_p^2 entre 0.66 - 0.78 y valores de RPD entre 1.97 - 2.30 para SVMR. Por otro lado, iPLS fue superior para predecir Vitamina C y firmeza tanto para PLSR como para SVMR. El método GA fue superior para los sólidos solubles (°Brix) y rPLS para %acidez titulable. En general los resultados muestran una dificultad para la estimación de las propiedades tanto por PLSR como por SVMR. Esto puede deberse a la superposición de la señal de las moléculas de agua y la capa cerosa en todo el espectro. Finalmente, los trabajos futuros deben probarse en algoritmos no lineales y aprendizaje profundo para mejorar la predicción de propiedades.
ABSTRACT Golden gooseberry (Physalis peruviana) is a food native to the Andes of our country. Due to its rapid growth in production are looking for technologies that can control the internal quality with a high degree of reliability, speed and non-destructive. In this paper it was proposed to use a system of near infrared hyperspectral imaging (NIR-HSI) and chemometrics to predict the vitamin C content, soluble solids (°Brix) % titratable acidity and texture of golden gooseberry. Samples were collected from different local regions in Peru (Cajamarca, Chachapoyas, Huaraz and Otuzco), then images were taken with a NIR-HSI camera in a range from 953 to 1703 nm in 3 different orientations (calyx, apex and equatorial zone). Predictive chemometric models were trained and validated: Partial Least Squares Regression (PLSR) and Support Vector Machine Regression (SVMR), using the full and selected wavelength range to predict physicochemical properties. The methods used to select important variables were: recursive PLS (rPLS), interval PLS (iPLS), genetic algorithm (GA) and covariance between variables (Covsel). The regression models based on the full wavelength showed the calyx zone orientation (2) as the one that obtained the best indicators, resulting for PLSR an R_p^2 between 0.43 - 0.76 and RPD values between 1.54 - 2. 25, while for SVMR they presented an R_p^2 between 0.37 - 0.82 and RPD values between 1.54 - 2.30. Vitamin C was the property that presented the highest results, with an R_p^2 between 0.66 - 0.78 and RPD values between 1.97 - 2.30 for SVMR. On the other hand, iPLS was superior for predicting Vitamin C and firmness for both PLSR and SVMR. The GA method was superior for soluble solids (°Brix) and rPLS for titratable % acidity. In general the results show a difficulty in estimating properties by both PLSR and SVMR. This may be due to the overlapping of the signal from water molecules and the waxy layer over the whole spectrum. Finally, future research should be tested on nonlinear algorithms and deep learning to improve property prediction.
ABSTRACT Golden gooseberry (Physalis peruviana) is a food native to the Andes of our country. Due to its rapid growth in production are looking for technologies that can control the internal quality with a high degree of reliability, speed and non-destructive. In this paper it was proposed to use a system of near infrared hyperspectral imaging (NIR-HSI) and chemometrics to predict the vitamin C content, soluble solids (°Brix) % titratable acidity and texture of golden gooseberry. Samples were collected from different local regions in Peru (Cajamarca, Chachapoyas, Huaraz and Otuzco), then images were taken with a NIR-HSI camera in a range from 953 to 1703 nm in 3 different orientations (calyx, apex and equatorial zone). Predictive chemometric models were trained and validated: Partial Least Squares Regression (PLSR) and Support Vector Machine Regression (SVMR), using the full and selected wavelength range to predict physicochemical properties. The methods used to select important variables were: recursive PLS (rPLS), interval PLS (iPLS), genetic algorithm (GA) and covariance between variables (Covsel). The regression models based on the full wavelength showed the calyx zone orientation (2) as the one that obtained the best indicators, resulting for PLSR an R_p^2 between 0.43 - 0.76 and RPD values between 1.54 - 2. 25, while for SVMR they presented an R_p^2 between 0.37 - 0.82 and RPD values between 1.54 - 2.30. Vitamin C was the property that presented the highest results, with an R_p^2 between 0.66 - 0.78 and RPD values between 1.97 - 2.30 for SVMR. On the other hand, iPLS was superior for predicting Vitamin C and firmness for both PLSR and SVMR. The GA method was superior for soluble solids (°Brix) and rPLS for titratable % acidity. In general the results show a difficulty in estimating properties by both PLSR and SVMR. This may be due to the overlapping of the signal from water molecules and the waxy layer over the whole spectrum. Finally, future research should be tested on nonlinear algorithms and deep learning to improve property prediction.
Description
Keywords
Machine learning, Selección de variables, Agroberries, Control de calidad