# Research on non-destructive detection model of tomato fruit quality based on electrical properties and machine learning algorithms

## TL;DR

This paper introduces a new non-destructive method using machine learning to assess tomato quality based on electrical properties, offering faster and more accurate results than traditional methods.

## Contribution

The novel LSTMAE–XGBoost model combines autoencoder, LSTM, and XGBoost for improved non-destructive tomato quality prediction.

## Key findings

- The LSTMAE–XGBoost model achieved high prediction accuracy for four tomato quality indicators.
- The model outperformed traditional methods by 7.8% to 14.3% in prediction accuracy.
- It can simultaneously predict all four internal quality indicators, improving efficiency.

## Abstract

To break through the limitations of traditional destructive detection methods, achieve rapid, non-destructive, and accurate detection of internal tomato quality, and provide more efficient technical means for agricultural product quality assessment, this study proposes a novel predictive method that integrates a Long Short-Term Memory Autoencoder (LSTMAE) and XGBoost (LSTMAE–XGBoost). This method combines the feature extraction capabilities of the autoencoder, the sequence data processing abilities of LSTM, and the high-precision prediction capabilities of XGBoost. Within the frequency range of 0.1–1000 kHz, electrical parameters such as parallel equivalent capacitance, parallel equivalent resistance, and quality factor—among nine electrical parameters—were obtained from 300 tomato samples using an electrical parameter analyzer. Additionally, four indicators—vitamin C, soluble sugar, soluble protein, and titratable acidity—were obtained through physicochemical analysis of the tomatoes. Based on the electrical parameters and internal physicochemical indicator data of the tomatoes, a non-destructive detection model for tomato internal quality indicators was constructed. Experimental results demonstrate that the LSTMAE–XGBoost model exhibits superior adaptability. In the test set, the coefficients of determination for vitamin C, soluble sugar, soluble protein, and titratable acidity were 0.805, 0.945, 0.838, and 0.845, respectively. Compared to traditional machine learning models, this model offers better prediction accuracy. It improves upon the traditional Pearson correlation coefficient (PCC) feature extraction method by 14.3%, 13.1%, 7.8%, and 9.5%, respectively. Furthermore, LSTMAE–XGBoost can simultaneously predict all four indicators, enhancing the model’s efficiency. Therefore, LSTMAE–XGBoost can be utilized as an effective ensemble model for non-destructive detection of tomato internal quality indicators, which holds significant importance for fruit quality non destructive detection in the horticultural industry.