Graph Constraint and Collaborative Representation Classifier Steered Discriminative Projection with Applications for the Early Identification of Cucumber Diseases

Yuhua Li, Fengjie Wang, Ye Sun, Yingxu WangSensors 2020 · DOI

TL;DR

This paper introduces a novel machine learning approach, CRC-DP, that uses hyperspectral imaging to identify cucumber diseases in their very early stages, even before visible symptoms appear. By combining collaborative representation with graph-based constraints, the method effectively reduces data dimensionality and improves classification accuracy, achieving near-perfect identification with rapid diagnosis times.

What Problem Does This Paper Address?

Accurate and timely identification of plant diseases is crucial for efficient agricultural production. Traditional methods, like visual inspection or lab-based pathology, are often slow, labor-intensive, and require expert knowledge, making them impractical for large-scale, real-time monitoring. While image processing has shown promise, its effectiveness is limited to diseases with obvious visual symptoms, which are typically in later stages of infection. The challenge lies in detecting diseases early, when symptoms are subtle or absent, to enable prompt intervention and prevent widespread crop loss. Hyperspectral imaging (HSI) offers a potential solution by capturing detailed spectral information, but the high dimensionality of HSI data and the subtle spectral changes in early-stage diseases make direct classification difficult.

Methods

The core of this paper is the CRC-DP (Collaborative Representation Classifier-Steered Discriminative Projection) method. It's a two-stage process:

  1. Offline Training Stage:
    • Discriminative Projection: The primary goal is to learn a linear projection matrix ($P$) that transforms high-dimensional hyperspectral data into a lower-dimensional space. This projection is designed to enhance the separation between spectral curves of different diseases while keeping curves from the same disease close together.
  2. Collaborative Representation Classifier (CRC) Integration: The training of this projection is steered by the decision rule of CRC. This means the projection is optimized to make it easier for CRC to classify samples in the reduced space.
  3. Reconstruction Fidelity Terms: Two terms are introduced:
  4. Same-class reconstruction residual: Encourages samples from the same disease class to be well-represented by other samples from that class.
  5. Different-class reconstruction residual: Encourages samples from different disease classes to be poorly represented by each other.
  6. Graph Constraints: To further improve separation and leverage spatial information, two graph constraints are employed:
  7. Same-class graph: Encourages samples belonging to the same class to be close in the low-dimensional space.
  8. Different-class graph: Encourages samples from different classes to be far apart.
  9. Objective Function: These fidelity terms and graph constraints are fused into a single objective function that is optimized to find the best projection matrix $P$. This optimization is performed iteratively.
  10. Online Identification Stage:
    • Once the projection matrix $P$ is trained, a new, unknown query sample is first projected into the low-dimensional space.
  11. Then, CRC is used to determine the class (disease type) of the projected sample by evaluating which training class provides the minimal reconstruction residual.

Datasets

The study utilizes hyperspectral data collected from:

  • Cucumber plants: Specifically, 'Lufeng' cucumber varieties were used.
  • Plants were inoculated with two common diseases: cucumber anthracnose and cucumber Corynespora cassiicola.
  • A control group of healthy cucumber plants was also included.
  • Data was collected over 12 days, starting about 24 hours after inoculation, focusing on leaves with less than 20% lesion area to capture early infection stages.
  • The hyperspectral imaging unit captured data in the visible range (391-1045 nm) with 256 spectral bands.
  • Each hyperspectral curve was vectorized and normalized, resulting in samples of 256 dimensions.
  • For experiments, training and testing sets were prepared with 1500 samples per class (500 for each of the three classes: anthracnose, Corynespora cassiicola, and normal).
  • Toy Dataset: A synthetic dataset was created to evaluate the dimensionality reduction capabilities of CRC-DP against other methods.
  • Wine Dataset (UCI): A standard dataset with 178 samples and 13 features was used for similar comparative evaluations.

Key Insights and Ideas

  • Early Detection is Key: The paper strongly emphasizes that identifying diseases in their nascent stages, before visible symptoms, is critical for effective crop management.
  • HSI for Subtle Changes: Hyperspectral imaging is well-suited for this task because it can detect subtle biochemical and biophysical changes in plant tissues that precede macroscopic symptoms.
  • Dimensionality Reduction is Crucial: The high dimensionality of HSI data poses a significant challenge. The CRC-DP method's discriminative projection is designed not just to reduce dimensions but to do so in a way that enhances class separability.
  • CRC as a Steering Mechanism: Integrating the CRC decision rule into the projection learning process is a novel way to ensure that the learned low-dimensional space is optimized for the subsequent classification task, rather than being a generic dimensionality reduction.
  • Graph Constraints Add Power: The use of same-class and different-class graph constraints effectively leverages the label and spatial distribution information, which is often overlooked in purely spectral-based methods, leading to better separation.
  • Robustness to Dimension: The method demonstrates high accuracy even with very low-dimensional representations (e.g., m=2), suggesting it can be efficient without sacrificing performance.

Key Results

  • High Identification Accuracy: The CRC-DP method achieved a maximum identification accuracy of 98.2% for the three-class classification of cucumber diseases.
  • Superior Performance: It consistently outperformed other compared methods (KNN, RF, NB, DA, SVM) in terms of classification accuracy.
  • Fast Diagnosis: The average online identification time was remarkably short, at 0.65 ms, making it suitable for real-time applications.
  • Effective Dimensionality Reduction: Experiments on toy and wine datasets showed that CRC-DP provided better separation of classes in the reduced-dimensional space compared to PCA, NPE, LPP, and SPP.
  • Benefit of Graph Constraints: Figure 8a demonstrates that the inclusion of graph constraints significantly improves identification accuracy, especially for smaller dimensions.

Limitations and Future Work

  • Data Scarcity: The paper acknowledges that collecting sufficient agricultural data can be costly and time-consuming, which can affect performance, particularly when the number of training samples per class is very small (less than 150 in their experiments).
  • Parameter Sensitivity: While the paper sets some parameters (e.g., $\epsilon=0.05$), the influence of other parameters like the number of neighbors in KNN or decision trees in RF was explored, and the authors suggest that the number of training samples per class is a critical factor.
  • Future Work: The authors suggest that the method has promising potential for practical production due to its high accuracy and speed. Further research could explore optimizing parameter selection and investigating its applicability to a wider range of plant diseases and crop types. The robustness to small sample sizes could also be a focus for future improvements.

Summary updated

Abstract

Accurate, rapid and non-destructive disease identification in the early stage of infection is essential to ensure the safe and efficient production of greenhouse cucumbers. Nevertheless, the effectiveness of most existing methods relies on the disease already exhibiting obvious symptoms in the middle to late stages of infection. Therefore, this paper presents an early identification method for cucumber diseases based on the techniques of hyperspectral imaging and machine learning, which consists of two procedures. First, reconstruction fidelity terms and graph constraints are constructed based on the decision criterion of the collaborative representation classifier and the desired spatial distribution of spectral curves (391 to 1044 nm) respectively. The former constrains the same-class and different-class reconstruction residuals while the latter constrains the weighted distances between spectral curves. They are further fused to steer the design of an offline algorithm. The algorithm aims to train a linear discriminative projection to transform the original spectral curves into a low dimensional space, where the projected spectral curves of different diseases own better separation trends. Then, the collaborative representation classifier is utilized to achieve online early diagnosis. Five experiments were performed on the hyperspectral data collected in the early infection stage of cucumber anthracnose and <i>Corynespora cassiicola</i> diseases. Experimental results demonstrated that the proposed method was feasible and effective, providing a maximal identification accuracy of 98.2% and an average online identification time of 0.65 ms. The proposed method has a promising future in practical production due to its high diagnostic accuracy and short diagnosis time.
BibTeX
@inproceedings{li2020graph,
  author = {Yuhua Li and Fengjie Wang and Ye Sun and Yingxu Wang},
  title = {{Graph Constraint and Collaborative Representation Classifier Steered Discriminative Projection with Applications for the Early Identification of Cucumber Diseases}},
  booktitle = {Sensors},
  year = {2020},
  doi = {10.3390/s20041217},
}
Figure