Scholarlink

<p>Chlorophyll content is an important indicator for evaluating the physiological health of vegetables. Accurate and non-destructive detection techniques are of great significance for monitoring vegetable growth and improving quality and yield. This study proposes a non-destructive detection method for chlorophyll content in zucchini leaves based on machine learning and color features: 400 zucchini leaf images were collected, color features were extracted, and their correlations with SPAD and other features were analyzed. Combined with principal component analysis (PCA) for dimensionality reduction, four prediction models—Linear Regression (LR), Support Vector Regression (SVR), Gaussian Process Regression (GPR), and Bayesian Neural Network (BNN) prediction models and a tenfold cross-validation system were established to systematically analyze the results. Finally, the model performance was further validated using Friedman tests, Nemenyi post-hoc tests, and Bootstrap sampling. The experimental results showed that PCA processing significantly improved the performance of the four models, with PCA-BNN performing the best: in tenfold cross-validation, the validation set R<sup>2</sup> reached 0.9759 ± 0.0096, RMSE was 0.6213 ± 0.1307, and RPD was 6.8914 ± 1.6549. Statistical tests showed that the average rank of PCA-BNN in the Friedman and Nemenyi tests was the smallest at 1.3, significantly outperforming all other models except PCA-GPR; Bootstrap sampling validation further confirmed the reliability of the PCA-BNN model, with the validation set R<sup>2</sup> confidence interval ranging from 0.9578, 0.9903, RMSE from 0.4251, 0.7921, and RPD from 4.8829, 8.8931, with all interval ranges smaller than those of PCA-GPR, indicating superior stability. Additionally, directly inputting the original RGB image into the MobileNetV2 model yielded unsatisfactory results, confirming the necessity of manual feature extraction and dimensionality reduction. This study provides new insights for non-destructive chlorophyll detection in zucchini leaves and offers a technical approach for data feature processing and model construction in agricultural precision monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research on Non-destructive Detection of Chlorophyll in Zucchini Leaves Based on Machine Learning and Color Features

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Xinyue Gou,
  • Yuemei Sun,
  • Xin Han,
  • Yubin Lan,
  • Huizheng Wang,
  • Shuo Zhao,
  • Jingbo Bai,
  • Maochang Song

摘要

Chlorophyll content is an important indicator for evaluating the physiological health of vegetables. Accurate and non-destructive detection techniques are of great significance for monitoring vegetable growth and improving quality and yield. This study proposes a non-destructive detection method for chlorophyll content in zucchini leaves based on machine learning and color features: 400 zucchini leaf images were collected, color features were extracted, and their correlations with SPAD and other features were analyzed. Combined with principal component analysis (PCA) for dimensionality reduction, four prediction models—Linear Regression (LR), Support Vector Regression (SVR), Gaussian Process Regression (GPR), and Bayesian Neural Network (BNN) prediction models and a tenfold cross-validation system were established to systematically analyze the results. Finally, the model performance was further validated using Friedman tests, Nemenyi post-hoc tests, and Bootstrap sampling. The experimental results showed that PCA processing significantly improved the performance of the four models, with PCA-BNN performing the best: in tenfold cross-validation, the validation set R2 reached 0.9759 ± 0.0096, RMSE was 0.6213 ± 0.1307, and RPD was 6.8914 ± 1.6549. Statistical tests showed that the average rank of PCA-BNN in the Friedman and Nemenyi tests was the smallest at 1.3, significantly outperforming all other models except PCA-GPR; Bootstrap sampling validation further confirmed the reliability of the PCA-BNN model, with the validation set R2 confidence interval ranging from 0.9578, 0.9903, RMSE from 0.4251, 0.7921, and RPD from 4.8829, 8.8931, with all interval ranges smaller than those of PCA-GPR, indicating superior stability. Additionally, directly inputting the original RGB image into the MobileNetV2 model yielded unsatisfactory results, confirming the necessity of manual feature extraction and dimensionality reduction. This study provides new insights for non-destructive chlorophyll detection in zucchini leaves and offers a technical approach for data feature processing and model construction in agricultural precision monitoring.