RAPID NON-DESTRUCTIVE ASSESSMENT OF HASS AVOCADO MATURITY USING MACHINE LEARNING-ASSISTED DIFFUSE REFLECTANCE SPECTROSCOPY

Abstract

Determining the optimum harvest maturity is crucial for ensuring high avocado fruit quality after harvesting. Currently, maturity is assessed using invasive methods that involve dissecting the fruit to measure markers such as dry matter (DM), moisture content (MC), or mesocarp oil (CM). These methods are time-consuming and require laboratory facilities and trained personnel. Therefore, rapid, non-invasive methods are needed. This study established a fast method to assess avocado maturity without damaging the fruit, employing diffuse reflectance spectroscopy and machine learning algorithms. Spectral data from 200 Hass avocados were recorded over the 900-2500 nm wavelength range using a NIR Quest spectrometer. To establish a reference, dry matter content was determined through a conventional destructive method. The spectral data were converted to absorbance and preprocessed using first-order and second-order derivatives, Savitzky-Golay filtering, standard normal variate (SNV) transformations, and multiplicative scatter correction (MSC). Machine learning models, including Convolutional Neural Networks (CNN), Artificial Neural Networks (ANN), Support Vector Machines (SVM), and Random Forest (RF), were employed to predict dry matter content. In this study, a CNN model trained on raw spectral data for the full range 1000-2350nm achieved superior performance, with a coefficient of determination (R²) value of 0.94, a root mean square error of prediction (RMSEP) of 1.52, and a mean absolute error (MAE) of 1.17. The narrower 1000-1500 nm range performed comparatively well, with an R² value of 0.94, RMSEP of 1.58, and MAE of 1.19, indicating it is a strong candidate for accurate predictions with potential for cost-effective NIR instruments. This range offers good accuracy and potential for creating cost-effective NIR instruments. These findings indicate that targeted NIR spectroscopy, combined with advanced machine learning, can accurately predict avocado maturity without damaging the fruit. The developed approach provides a non-destructive, rapid, and objective alternative to traditional laboratory-based methods, reducing the need for destructive sampling and minimizing post-harvest losses. This contributes to improved supply chain efficiency, consistent fruit quality, and enhanced competitiveness in both local and international avocado markets.