A Reliable Methodology for Determining Seed Viability by Using Hyperspectral Data from Two Sides of Wheat Seeds

Tingting Zhang, Wensong Wei, Bin Zhao, Ranran Wang, Mingliu Li, Liming Yang, Jianhua Wang, Qun SunSensors 2018 · DOI

TL;DR

This paper presents a non-destructive method for determining wheat seed viability using hyperspectral imaging (HSI) of both sides of the seed. By combining spectral preprocessing (SNV), wavelength selection (SPA), and classification (PLS-DA), they achieved high accuracy in predicting viability, significantly improving germination rates compared to traditional methods.

What Problem Does This Paper Address?

Ensuring seed quality, particularly viability, is crucial for agricultural productivity and the seed industry. Traditional methods for assessing seed viability, such as tetrazolium staining, conductivity tests, and germination tests, are often labor-intensive, time-consuming, destructive, and prone to human error. This creates a need for rapid, non-destructive, and accurate methods to assess seed viability, especially for high-value crops like wheat. While hyperspectral imaging has shown promise in seed analysis, its application for differentiating viable from non-viable seeds, particularly considering different seed orientations, requires further investigation.

Methods

The researchers employed a hyperspectral imaging system operating in the visible and near-infrared (VIS/NIR) range (400-1000 nm) to capture spectral data from individual wheat seeds. Key methodological steps included:

  • Seed Preparation: Wheat seeds underwent an accelerated aging test to induce varying levels of viability, allowing for direct correlation between spectral data and germination status.
  • Hyperspectral Imaging: A custom HSI system was used to acquire spectral data from both the ventral groove and reverse sides of each seed.
  • Spectral Data Extraction: After image calibration and background removal, spectral data from the region of interest (ROI) for each seed was extracted and averaged.
  • Spectral Preprocessing: Several methods were explored to reduce noise and enhance spectral features, including Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), and Savitzky-Golay (SG) smoothing and derivatives.
  • Optimal Wavelength Selection: The Successive Projections Algorithm (SPA) was used to identify the most informative wavelengths, reducing data dimensionality and computational load.
  • Classification Models: Partial Least Squares Discriminant Analysis (PLS-DA) and Support Vector Machines (SVM) were employed to build classification models for distinguishing viable from non-viable seeds.
  • Dataset Construction: Four types of spectral datasets were created: ventral groove side, reverse side, mean of both sides, and a mixture dataset combining data from both sides.
  • Model Evaluation: Models were evaluated using overall accuracy, viability accuracy (recall), final germination percentage (precision), and F-measure.

Datasets

The study utilized a dataset of wheat seeds (cultivar Dunmaiwang) that were subjected to an accelerated aging test to create a range of viabilities. This resulted in a dataset where spectral data could be directly correlated with germination status. The paper does not introduce a new benchmark dataset but rather uses this prepared set for evaluating their methodology. The data was split into calibration (106 seeds) and prediction (54 seeds) sets.

Key Insights and Ideas

  • Dual-Sided Imaging is Crucial: The study highlights that considering spectral data from both sides of the wheat seed (ventral groove and reverse) provides richer information for viability assessment than using a single side alone.
  • Preprocessing Matters: Different spectral preprocessing techniques (SNV, MSC, SG) had varying impacts on model performance, emphasizing the need to select appropriate methods for specific datasets. SNV and SG were found to be more effective than MSC.
  • Wavelength Selection Enhances Efficiency: Using SPA to select a minimal set of optimal wavelengths significantly reduced the spectral data size (from 688 to as few as 16 bands) without compromising classification accuracy, making the method more practical and faster.
  • PLS-DA with SNV and SPA is a Winning Combination: The SNV-SPA-PLS-DA model, particularly when applied to a "mixture" dataset (combining data from both sides), demonstrated superior performance in classifying seed viability.
  • Improved Germination Rates: The developed methodology not only accurately predicts viability but also demonstrably increases the germination percentage of the seed lot, translating directly to practical agricultural benefits.

Key Results

The SNV-SPA-PLS-DA model applied to the "mixture" spectral dataset achieved the highest classification performance:

  • Overall Accuracy: >85.2%
  • Viability Accuracy (Recall): >89.5%
  • Final Germination Percentage (Precision): >89.5%
  • F-Measure: >89.5%

This model used only 16 selected wavelengths. The initial germination percentage of the seed lot was 70.4%, and after screening with this model, the final germination percentage increased to 89.5%. Models based on the reverse side spectral data also showed strong performance, with an SVM model achieving 88.9% overall accuracy and 97.4% viability accuracy.

Limitations and Future Work

  • Seed Orientation in Mixture Dataset: While the "mixture" dataset performed well, the paper notes that models performed better with an increasing proportion of reverse-side data in the prediction set. This suggests that the orientation of the seed during imaging can still influence results, and further optimization of how data from both sides is combined might be beneficial.
  • Scanner Cost and Speed: The paper mentions that acquiring average spectral data might require two cameras for simultaneous screening, potentially affecting speed and cost. Future work could focus on developing more integrated and cost-effective multispectral screening machines.
  • Generalizability: While the study focused on wheat seeds, further research could explore the applicability of this methodology to other seed types and varieties.
  • Integration with Other Data: Future work could investigate combining hyperspectral data with other non-destructive sensing techniques or machine learning approaches for even more robust seed quality assessment.

Summary updated

Abstract

This study investigated the possibility of using visible and near-infrared (VIS/NIR) hyperspectral imaging techniques to discriminate viable and non-viable wheat seeds. Both sides of individual seeds were subjected to hyperspectral imaging (400-1000 nm) to acquire reflectance spectral data. Four spectral datasets, including the ventral groove side, reverse side, mean (the mean of two sides' spectra of every seed), and mixture datasets (two sides' spectra of every seed), were used to construct the models. Classification models, partial least squares discriminant analysis (PLS-DA), and support vector machines (SVM), coupled with some pre-processing methods and successive projections algorithm (SPA), were built for the identification of viable and non-viable seeds. Our results showed that the standard normal variate (SNV)-SPA-PLS-DA model had high classification accuracy for whole seeds (>85.2%) and for viable seeds (>89.5%), and that the prediction set was based on a mixed spectral dataset by only using 16 wavebands. After screening with this model, the final germination of the seed lot could be higher than 89.5%. Here, we develop a reliable methodology for predicting the viability of wheat seeds, showing that the VIS/NIR hyperspectral imaging is an accurate technique for the classification of viable and non-viable wheat seeds in a non-destructive manner.
BibTeX
@inproceedings{zhang2018reliable,
  author = {Tingting Zhang and Wensong Wei and Bin Zhao and Ranran Wang and Mingliu Li and Liming Yang and Jianhua Wang and Qun Sun},
  title = {{A Reliable Methodology for Determining Seed Viability by Using Hyperspectral Data from Two Sides of Wheat Seeds}},
  booktitle = {Sensors},
  year = {2018},
  doi = {10.3390/s18030813},
}
Figure