基于高光谱成像和机器学习的苹果可溶性固形物含量的无损检测

Non-Destructive Prediction of the Soluble Solids Content of Apples Using Hyperspectral Imaging and Machine Learning

摘要:

为实现苹果可溶性固形物（Soluble Solids Content, SSC）含量的快速无损检测，本文通过高光谱成像系统，采集160个苹果在960 ~ 2500 nm波长范围内的288个高光谱数据。经过多元散射校正（Multi-Scatter Calibration, MSC）、最大-最小标准化（Maximum Minimum Standardization, MMS）、均值中心化（Mean Centering, MC）以及标准正态变换（Standard Normal Variation, SNV）光谱预处理后，构建偏最小二乘回归（Partial Least Square Regression, PLSR）模型。然后选用竞争性自适应重加权算法（Competitive Adaptive Reweighted Sampling, CARS）、连续投影算法（Successive Projections Algorithm, SPA）和基于灰狼优化算法（Grey Wolf Optimization algorithm, GWO）、基于鲸鱼优化算法（Whale Optimization Algorithm, WOA）、多策略融合的斑马优化算法（Multi-Strategy Fusion Zebra Optimization Algorithm, MFZOA）提取特征波长，构建PLSR、极限学习机（Extreme Learning Machine, ELM）和反向传播神经网络（Back Propagation Neural Network, BPNN）苹果SSC预测模型。预测结果表明，SNV-MFZOA-PLSR模型预测性能最优，测试集的均方根误差为0.1869，决定系数为0.9252，均高于其它模型。研究表明高光谱成像技术结合机器学习可实现对苹果SSC含量的快速无损检测，为水果品质检测模型提供了基础理论参考。

Abstract:

In order to achieve rapid and non-destructive detection of soluble solids content (SSC) in apples, hyperspectral imaging was employed to acquire 288 spectral data from 160 apple samples across the 960 ~ 2500 nm wavelength range. The spectra were preprocessed using four methods: Multi-Scatter Calibration (MSC), Maximum–Minimum Standardization (MMS), Mean Centering (MC), and Standard Normal Variate (SNV). A Partial Least Squares Regression (PLSR) model was then established based on the full-spectrum data. Feature wavelengths were selected using several wavelength selection algorithms, including Competitive Adaptive Reweighted Sampling (CARS), Successive Projections Algorithm (SPA), Whale Optimization Algorithm (WOA), and a Multi-Strategy Fusion Zebra Optimization Algorithm (MFZOA) derived from the Grey Wolf Optimization (GWO) approach. These selected wavelengths were subsequently used to construct PLSR, Extreme Learning Machine (ELM), and Back Propagation Neural Network (BPNN) models for SSC prediction. The prediction results showed that the SNV–MFZOA–PLSR model exhibited the best predictive performance, achieving a root mean square error of prediction (RMSEP) of 0.1869 and a coefficient of determination (Rp2) of 0.9252 on the test set, outperforming other models. The findings suggest that hyperspectral imaging combined with optimized machine learning algorithms provides an effective approach for rapid and accurate non-destructive detection of apple SSC, offering valuable insights for intelligent fruit quality assessment and postharvest evaluation.