基于随机森林算法以及可见–近红外光谱的苹果糖度无损检测

Non-Destructive Detection of Apple Sugar Content Based on Random Forest Algorithm and Visible-Near Infrared Spectroscopy

摘要: 本文基于可见–近红外光谱分析技术结合随机森林算法实现不同产地的苹果糖度无损检测。研究通过漫反射采集系统收集三种不同产地苹果的光谱数据后经多种预处理办法比较，采用标准正态变换分别结合偏最小二乘、随机森林算法建立苹果糖度检测通用模型。结果显示该模型预测集相关系数(R_p²)和预测均方根误差(RMSEP)分别为0.89和0.44，相比偏最小二乘法检测模型相关系数(R_p²)和预测均方根误差(RMSEP)的0.85和0.47，均有提高。研究扩大了单一品种模型的预测范围，结合随机森林算法有效地提升模型的预测稳健性，对进一步实现水果品质无损检测具有良好的潜在意义。

Abstract: This paper is based on visible-near-infrared spectroscopy analysis technology combined with random forest algorithm to achieve non-destructive testing of apple sugar content in different producing areas. The study collects the spectral data of three apples from different origins through the diffuse reflectance collection system and compares them with a variety of preprocessing methods. The standard normal transformation is combined with partial least squares and random forest algorithms to establish a general model for apple sugar content detection. The results show that the correlation coefficient (R_p²) and root mean square error (RMSEP) of the prediction set of the model are 0.89 and 0.44, respectively. Compared with the partial least square method to detect the correlation coefficient (R_p²) and root mean square error (RMSEP) of the model 0.85 and 0.47, both improved. The research expanded the prediction range of the single-variety model, combined with the random forest algorithm to effectively improve the prediction robustness of the model, which has good potential significance for the further realization of non-destructive testing of fruit quality.