# New Analytical Strategies for Quality Control and Classification of Apple Juices Using Digital Image Processing (DIP) Combined with Machine Learning (ML)

## TL;DR

This paper introduces a low-cost, smartphone-based method using image analysis and machine learning to detect apple juice authenticity and quality.

## Contribution

The study is the first to combine smartphone imaging with machine learning for classifying and quantifying apple juice authenticity.

## Key findings

- Classification models achieved 95.9% accuracy in testing for apple juice types.
- Predictive models estimated juice concentration with 93.1–93.5% accuracy in testing.
- The approach is a rapid, nondestructive alternative for quality control.

## Abstract

Apple juice is widely consumed for its pleasant sensory
attributes
and nutritional value. However, due to its high commercial value,
this beverage, particularly whole juice, is susceptible to fraud and
adulteration. This underscores the need for rapid, noninvasive analytical
strategies to ensure product authenticity and quality. This study
reports, for the first time, the application of smartphone-based image
analysis combined with machine learning as a low-cost and nondestructive
approach for both classifying apple juice types and predicting the
actual juice content in apple-based beverages. Images of nine whole
juice (WJ), four reconstituted juice (RE), and five nectar (NE) samples
were analyzed to develop models capable of discriminating beverage
categories and estimating apple juice concentration. Classification
models generated using k-nearest neighbors (kNN) and extreme gradient boosting (XGBoost) algorithms
achieved good global accuracies of 95.9% (99.1% and 100.0% in training
and 95.9% in testing), with cross-validation accuracies of 98.7% and
95.5%, respectively. Predictive models constructed from calibration
curves (5–100% apple juice) combined with commercial samples
yielded good estimates for coefficients of determination (R
2 = 93.1–93.5% in testing, 97.7–96.2%
in training, and 91.9–92.4% in cross-validation) and root-mean-square
errors (RMSE = 8.0–8.2%) for models generated by XGBoost and
CatBoost, respectively. These findings demonstrate the scientific
novelty and practical feasibility of integrating smartphone imaging
with machine learning for the quantitative analysis of apple juice.
The proposed approach represents a rapid, accurate, and cost-effective
alternative for industrial quality control and regulatory inspection.