Differentiation of Apple Varieties and Investigation of Organic Status Using Portable Visible Range Reflectance Spectroscopy

Jordan Vincent, Hui Wang, Omar Nibouche, Paul MaguireSensors 2018 · OpenReview

TL;DR

This paper explores the use of affordable, portable visible range spectroscopy combined with machine learning (PLS-DA) to authenticate apples. The system successfully differentiates apple varieties with high accuracy but struggles with reliably determining organic status, highlighting the need for further research in this area.

What Problem Does This Paper Address?

Food fraud, where products are mislabeled or tampered with, is a growing global concern. Current food authentication methods often rely on expensive, bulky laboratory equipment, limiting their accessibility throughout the supply chain, including at the consumer level. This creates a gap where more portable and affordable solutions are needed to ensure food authenticity. Specifically, this study investigates whether a low-cost, portable visible range spectrometer can differentiate between apple varieties and, more challenging, their organic status, addressing the need for accessible food authentication tools.

Methods

The core methodology involves using a portable OceanOptics SPARK visible range spectrometer (380-700 nm) to collect spectral data from apples. This data is then preprocessed using computationally inexpensive methods: zero correction, two-norm (Euclidean) normalization, and Savitzky-Golay smoothing. The primary analysis technique employed is Partial Least Squares Discriminant Analysis (PLS-DA), a well-established chemometric method for spectral data classification. The authors employ two validation strategies: 10-fold cross-validation within a single session ('single session experiment') and using data from one session as training and a separate session with different ambient conditions as testing ('inter-session experiment'). This inter-session approach is crucial for assessing the system's robustness in real-world, less controlled environments. Variable Importance in Projection (VIP) scores are used to identify which spectral regions are most influential in the models.

Datasets

The study utilized a dataset of 132 apples, comprising three varieties: Gala, Braeburn, and Pink Lady. For each variety, half the apples were organic, and the other half were non-organic. The apples were sourced from at least two countries of origin for each class to ensure geographic diversity. This dataset was collected specifically for this study and serves as the benchmark for evaluating the proposed method.

Key Insights and Ideas

  • Portability and Cost are Key: The choice of a low-cost, consumer-portable spectrometer is a significant factor, aiming to democratize food authentication beyond specialized labs.
  • Visible Range Potential: While near-infrared spectroscopy has seen more attention for food authentication, this study demonstrates that the visible range, often overlooked for this purpose, holds valuable information for differentiating apple types.
  • Variety Differentiation is Achievable: The system shows strong performance in distinguishing between different apple varieties, suggesting that subtle spectral differences are detectable and classifiable.
  • Organic Status is a Harder Nut to Crack: Differentiating organic from non-organic apples proved significantly more challenging. This suggests that the spectral features indicative of organic status are either less pronounced in the visible range or require more sophisticated analysis than what was applied here.
  • Ambient Conditions Matter: The inter-session experiment highlights the system's sensitivity to environmental changes, but also its ability to maintain reasonable accuracy, indicating potential for real-world application with appropriate calibration.
  • Spectral Regions of Interest: VIP scores pinpoint specific wavelengths (e.g., around 673 nm and 615 nm for apple type) that are most discriminative, offering avenues for targeted future research.

Key Results

  • Apple Type Differentiation: The system achieved a peak accuracy of 94% in the inter-session experiment for differentiating apple types, even with mild ambient interference.
  • Organic Status Differentiation: The best accuracy achieved for distinguishing organic from non-organic apples was 66% in the inter-session experiment, a modest improvement over the baseline.
  • RGB vs. Spectral Data: Converting spectral data to RGB values and using camera spectral sensitivity functions resulted in significantly lower classification accuracy (best at 63%) compared to using the raw spectral data. This strongly supports the superiority of full spectral information over simple color information for this task.

Limitations and Future Work

  • Organic Status Performance: The primary limitation is the relatively low accuracy in differentiating organic from non-organic apples. The authors suggest this could be due to insufficient information in the visible spectrum or the need for more complex preprocessing techniques.
  • Ambient Interference: While the inter-session test showed some robustness, the accuracy drop for organic status differentiation indicates that ambient conditions can still impact performance.
  • Future Work: The authors propose investigating alternative preprocessing and classification algorithms to improve performance, particularly for organic status. Further exploration of the identified spectral regions of importance (e.g., around 961 nm, 673 nm, and 615 nm) is also recommended to understand the underlying chemical constituents responsible for these spectral features. Expanding the dataset to include more apple varieties and a wider range of organic/non-organic samples could also enhance the models.

Summary updated

BibTeX
@inproceedings{vincent2018differentiation,
  author = {Jordan Vincent and Hui Wang and Omar Nibouche and Paul Maguire},
  title = {{Differentiation of Apple Varieties and Investigation of Organic Status Using Portable Visible Range Reflectance Spectroscopy}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=-PZBBSTzw},
}
Figure