Study on the Determination of Flavor Value of Rice Based on Grid Iterative Search Swarm Optimization Support Vector Machine Model and Hyperspectral Imaging

Han Yang, Fuheng Qu, Yong Yang, Xiaofeng Li, Ping Wang, Sike Guo, Lu WangSensors 2024 · DOI

TL;DR

This paper introduces a novel non-destructive method for determining rice flavor value using hyperspectral imaging combined with a Grid Iterative Search Particle Swarm Optimization Support Vector Machine (GISPSO-SVM) model. This approach achieves a high prediction accuracy of 96%, outperforming traditional methods and offering a faster, more efficient way to assess rice quality.

What Problem Does This Paper Address?

Assessing the flavor value of rice, a crucial aspect of its quality, traditionally relies on sensory evaluation or chemical analysis. These methods are often time-consuming, labor-intensive, subjective, and can even damage the rice samples, depleting valuable resources. The need for a rapid, efficient, and non-destructive technique to accurately determine rice flavor is therefore significant, especially for agricultural advancements and quality control in the food industry.

Methods

The core of this research lies in the development and application of a new machine learning model: the Grid Iterative Search Particle Swarm Optimization Support Vector Machine (GISPSO-SVM).

  • Hyperspectral Imaging: This technique captures detailed spectral information from rice grains across a wide range of wavelengths (400-1000 nm). This allows for the detection of subtle variations in the rice's chemical composition that correlate with flavor.
  • Image Preprocessing: Standard techniques like Multiplicative Scatter Correction (MSC), smoothing, and Standard Normal Variate (SNV) were applied to the hyperspectral data to reduce noise and improve signal quality.
  • Feature Selection: To manage the high dimensionality of hyperspectral data and improve model efficiency, two methods were explored:
  • Competitive Adaptive Reweighted Sampling (CARS): This algorithm iteratively selects characteristic spectral bands that are most informative for prediction.
  • Principal Component Analysis (PCA): This method reduces dimensionality by identifying the principal components that capture the most variance in the data.
  • Optimization Algorithm (GISPSO-SVM):
    • Support Vector Machine (SVM): Used as the primary regression model for predicting flavor values.
  • Particle Swarm Optimization (PSO): A metaheuristic optimization algorithm used to tune SVM hyperparameters (like 'c' and 'g').
  • Grid Iterative Search (GIS): This novel addition to PSO addresses its tendency to get stuck in local optima. GIS performs an initial broad grid search to identify a promising parameter range, then iteratively narrows this range with reduced step sizes. This guided search allows PSO to more effectively find the global optimum, combining the efficiency of PSO with the thoroughness of grid search.

Datasets

The study utilized hyperspectral images of 40 different rice varieties grown in the National Modern Agricultural Industrial Park of Zhalite Banner, Inner Mongolia Autonomous Region. For each variety, approximately 70 seeds were collected, resulting in a substantial dataset for training and evaluation. The data was split into a 7:3 ratio for training and testing.

Key Insights and Ideas

  • Hyperspectral Imaging as a Flavor Proxy: The study demonstrates that spectral reflectance patterns captured by hyperspectral imaging can serve as a reliable proxy for rice flavor, moving beyond traditional sensory or chemical methods.
  • GISPSO-SVM's Superiority: The integration of Grid Iterative Search (GIS) into the PSO algorithm significantly enhances its ability to find optimal SVM hyperparameters, overcoming the limitations of standard PSO (e.g., local optima) and grid search (e.g., computational cost).
  • Feature Selection Matters: Both CARS and PCA proved effective in reducing data dimensionality and improving model performance compared to using the full spectral bands. The choice of feature selection method, in conjunction with GISPSO-SVM, impacts the final accuracy.
  • Wavelength Significance: The research highlights that specific wavelength ranges, particularly above 450 nm, exhibit unique reflectance patterns that are strongly correlated with rice flavor differences.

Key Results

  • The proposed GISPSO-SVM model achieved a 96% accuracy in predicting rice flavor values when combined with PCA (50 principal components).
  • This accuracy (96%) was higher than that achieved by the CARS feature selection method combined with GISPSO-SVM (93%).
  • The GISPSO-SVM model also outperformed standard SVM and PSO models across various feature selection scenarios, consistently showing lower RMSE and higher accuracy. For instance, with PCA50, GISPSO-SVM achieved an RMSE of 0.3019 and 96% accuracy, compared to PSO-SVM's 0.3123 RMSE and 93% accuracy.
  • The study also noted that the flavor value of rice is influenced by its starch, protein, and moisture content, which are reflected in the spectral data.

Limitations and Future Work

  • Parameter Tuning: While GISPSO-SVM is an improvement, the paper notes that the effectiveness of the algorithm can still be influenced by the initial parameter range settings.
  • Feature Selection Refinement: The authors suggest that further refinement of feature selection algorithms could enhance the model's performance.
  • Integration with UAVs: A promising future direction is the integration of these models with unmanned aerial vehicles (UAVs) for real-time monitoring of rice flavor characteristics in agricultural landscapes.
  • Multispectral Studies: Further exploration of the GISPSO algorithm is recommended for multispectral studies, especially those involving large datasets.
  • Innovative Particle Update Mechanisms: The paper suggests that developing novel particle update strategies within the PSO framework could further improve search efficiency.

Summary updated

BibTeX
@article{yang2024study,
  author = {Han Yang and Fuheng Qu and Yong Yang and Xiaofeng Li and Ping Wang and Sike Guo and Lu Wang},
  title = {{Study on the Determination of Flavor Value of Rice Based on Grid Iterative Search Swarm Optimization Support Vector Machine Model and Hyperspectral Imaging}},
  journal = {Sensors},
  year = {2024},
  doi = {10.3390/s24144635},
}
Figure