Measurement of Early Disease Blueberries Based on Vis/NIR Hyperspectral Imaging System

Yuping Huang, Dezhen Wang, Ying Liu, Haiyan Zhou, Ye SunSensors 2020 · DOI

TL;DR

This paper presents a hyperspectral imaging system capable of detecting early-stage fungal diseases in blueberries, which are often invisible to the naked eye. By analyzing spectral data in the 685-1000 nm range and employing autoscale preprocessing, the system achieved near-perfect classification accuracy, paving the way for improved quality control in the blueberry industry.

What Problem Does This Paper Address?

Blueberries are highly susceptible to fungal infections during post-harvest handling and storage. Early detection of these diseases is crucial to prevent spoilage and cross-contamination, but it's challenging because the visual signs of infection are often subtle or hidden beneath the opaque skin in the early stages. Traditional methods like RGB imaging are insufficient for detecting these hidden defects. While hyperspectral imaging has shown promise for other fruits, its application for early disease detection in blueberries, particularly at the microscopic level, requires further investigation. This paper aims to bridge this gap by developing and validating a Vis/NIR hyperspectral imaging system for this specific purpose.

Methods

The researchers developed a hyperspectral imaging system operating in the visible and near-infrared (Vis/NIR) range (400-1000 nm). They collected hyperspectral images of both healthy and early-stage diseased blueberries. To identify the most informative spectral regions, they performed spectral correlation analysis, which revealed that the 685-1000 nm range was most effective for distinguishing between healthy and diseased samples.

For data analysis, they employed Partial Least Squares Discriminant Analysis (PLSDA), a chemometric technique well-suited for spectral data. To optimize model performance, they explored different preprocessing methods, including "no preprocessing," "autoscale," and "log(1/R)." They also investigated the impact of selecting an "effective spectral range" (685-1000 nm) versus using the full spectral range. Scanning Electron Microscopy (SEM) was used to visually confirm microstructural changes in the blueberry tissues due to fungal infection, providing a biological basis for the spectral differences observed.

Datasets

The study utilized a dataset of 400 blueberry samples, comprising 200 healthy and 200 early-stage diseased blueberries. These samples were obtained through a controlled storage process designed to induce natural decay. The dataset was split into training and testing sets (50% each) for model development and validation. No new benchmark datasets were introduced; rather, the focus was on analyzing the collected blueberry samples.

Key Insights and Ideas

  • Microscopic Damage Precedes Visible Signs: SEM images revealed that fungal infections cause significant damage to the cellular structure (cell membrane destruction) even when no visible mold is present on the surface. This microscopic damage is key to why hyperspectral imaging, which probes internal chemical and structural properties, can detect early disease.
  • NIR Region is Crucial for Disease Detection: The spectral correlation analysis highlighted that wavelengths beyond 685 nm, particularly in the NIR region, showed higher correlations with disease presence. This is attributed to the interaction of NIR light with chemical constituents and structural changes related to fungal infection, such as altered water content and degradation of nutrients.
  • Effective Wavelength Selection is Key: Simply using the full spectral range (400-1000 nm) led to less optimal results. Narrowing down to the effective spectral range (685-1000 nm) significantly improved classification accuracy, demonstrating the importance of removing irrelevant or noisy spectral information.
  • Autoscale Preprocessing Enhances Discrimination: Among the tested preprocessing methods, "autoscale" proved most effective, especially when combined with the selected spectral range. This method likely helps to emphasize subtle spectral differences between healthy and diseased samples by standardizing the variance of each spectral variable.

Key Results

The study achieved very high classification accuracies for detecting early-stage disease in blueberries:

  • Optimal Model: Using the effective spectral range (685-1000 nm) with the "autoscale" preprocessing method, the PLSDA model achieved:
  • Calibration Set: 100% accuracy for healthy blueberries and 100% for early disease blueberries.
  • Test Set: 100% accuracy for healthy blueberries and 99% for early disease blueberries.
  • Comparison: These results were significantly better than using the full spectral range or other preprocessing methods. For instance, the full spectral range with autoscale preprocessing achieved 91.0% accuracy on the test set, while the selected range with autoscale reached 99.5%.
  • Early Disease Detection: The system was able to distinguish early-stage diseased blueberries from healthy ones with remarkable precision, even when visual inspection would fail.

Limitations and Future Work

  • Limitations:
    • The study focused on one specific blueberry variety ('Wanqi'). The performance of the system might vary with different varieties due to differences in skin properties and internal composition.
  • The "early disease" stage was defined by a specific storage duration. Further research could explore a wider range of disease progression stages.
  • While SEM confirmed microstructural changes, a deeper understanding of the specific biochemical changes driving the spectral differences could further refine the models.
  • Future Work:
    • Investigate the system's performance on other blueberry varieties and under different storage conditions.
  • Explore more advanced feature selection and classification algorithms to potentially further improve accuracy and robustness.
  • Develop a portable or industrial-scale hyperspectral imaging system for real-time quality control applications in the blueberry industry.

Summary updated

BibTeX
@article{huang2020measurement,
  author = {Yuping Huang and Dezhen Wang and Ying Liu and Haiyan Zhou and Ye Sun},
  title = {{Measurement of Early Disease Blueberries Based on Vis/NIR Hyperspectral Imaging System}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20205783},
}
Figure