Evaluation of Stem Rust Disease in Wheat Fields by Drone Hyperspectral Imaging

Jaafar Abdulridha, An Min, Matthew N. Rouse, Shahryar Kianian, Volkan Isler, Ce Yang

TL;DR

This paper presents a drone-based hyperspectral imaging system for accurately detecting and classifying wheat stem rust disease severity. By leveraging machine learning algorithms and spectral vegetation indices, the system achieves high classification accuracies, even for low disease levels, offering a promising tool for precision agriculture and plant breeding.

What Problem Does This Paper Address?

Wheat is a staple crop globally, and diseases like stem rust pose a significant threat to production, potentially causing up to 70% yield loss. Early and accurate detection of stem rust is crucial for timely management decisions, such as fungicide application, and for plant breeders to select resistant varieties. Traditional methods of disease assessment rely on manual scouting, which is labor-intensive, time-consuming, and prone to human error, especially in large fields or for detecting subtle early-stage symptoms. The challenge lies in developing a cost-effective, efficient, and accurate method to monitor disease severity across vast agricultural areas, particularly at its initial stages.

Methods

The researchers employed a drone equipped with a hyperspectral camera to capture spectral data from wheat fields. This hyperspectral data, which contains detailed spectral information across a wide range of wavelengths, was then processed using various machine learning techniques.

Key methodological aspects include:

  • Hyperspectral Imaging: A Resonon Pika II hyperspectral camera mounted on a DJI Matrice 600 Pro UAV was used to acquire spectral data. This camera captures reflectance across the visible and near-infrared spectrum (400-900 nm) with high spectral resolution.
  • Machine Learning Classifiers: Several classification algorithms were tested to differentiate between healthy and diseased wheat plants, and to classify different levels of disease severity. These included:
  • Quadratic Discriminant Analysis (QDA)
    • Random Forest Classifier (RFC)
    • Support Vector Machine (SVM)
    • Decision Tree Classification (DTC)
  • Feature Selection: To identify the most informative spectral bands and reduce data dimensionality, feature selection techniques were applied. This helps in pinpointing the wavelengths most sensitive to stem rust.
  • Spectral Vegetation Indices (SVIs): Fourteen different SVIs were calculated and evaluated. These indices combine spectral bands to highlight specific plant biophysical properties, and their ability to discriminate disease severity was assessed.
  • Ground Truth Data: The spectral data was validated against ground-truth measurements of disease severity, categorized into four classes: healthy (class 0), mildly diseased (class 1), moderately diseased (class 2), and severely diseased (class 3).

Datasets

The study utilized spectral data collected from a field experimental design at the University of Minnesota Rosemount Research Outreach Center. This experiment involved 6 blocks, each containing 150 plots representing unique wheat lines. Three blocks were inoculated with the stem rust pathogen, while the remaining three were treated with fungicide to serve as healthy controls. The spectral data was collected from these plots, with a total of 699-784 spectral data samples used for training and testing the classification models across the different classes.

Key Insights and Ideas

  • Early Detection is Feasible: The study demonstrates that hyperspectral imaging, combined with machine learning, can effectively detect stem rust even at very low severity levels (mildly diseased, class 1), which are often difficult to discern with the naked eye. This is a significant advancement for proactive disease management.
  • RFC Outperforms Other Classifiers: The Random Forest Classifier (RFC) consistently achieved the highest classification accuracies across both multiclass and binary classifications, highlighting its robustness for this application.
  • Specific Wavelengths are Key: Feature selection identified specific wavelengths (e.g., around 817.5 nm, 612.2 nm, 760 nm) that are particularly sensitive to stem rust, suggesting potential for developing optimized multispectral sensors.
  • SVIs Offer Valuable Information: Certain spectral vegetation indices, particularly GNDVI, RARSc, and PRI, showed strong performance in discriminating disease severity, indicating their utility as indicators of plant stress.
  • Potential for New Sensors: The ability to identify critical wavelengths and effective SVIs opens the door for designing new, potentially less expensive, multispectral sensors tailored for stem rust detection.

Key Results

  • Multiclass Classification: The RFC achieved the highest overall classification accuracy of 85% for multiclass classification (healthy, mild, moderate, severe). For individual classes, RFC achieved up to 97% accuracy for class 2 (moderately diseased).
  • Binary Classification: When classifying between diseased and non-diseased plants, RFC achieved an accuracy of 88%.
  • Best Wavelengths: For binary classification, the best six bands identified were at approximately 817.5, 612.2, 760, 725.1, 497.2, and 706.6 nm.
  • Top Vegetation Indices: GNDVI and RARSc were identified as the most significant SVIs for distinguishing healthy plants from diseased ones.

Limitations and Future Work

  • Challenging Early Detection: While the study shows promise for early detection, the authors acknowledge that accurate and real-time detection of wheat rust infection remains challenging before the crop shows clearly associated symptoms.
  • Distinguishing Diseases: The study focused on stem rust. Future work could explore distinguishing wheat stem rust from other biotic and abiotic stresses, which can have similar spectral signatures.
  • Sensor Development: The authors suggest that the findings could lead to the development of new, inexpensive multispectral sensors for diagnosing wheat stem rust.
  • Field Application: Further research is needed to validate these techniques in large-scale field applications and to integrate them into existing precision agriculture workflows.

Summary updated

BibTeX
@inproceedings{abdulridha2023evaluation,
  author = {Jaafar Abdulridha and An Min and Matthew N. Rouse and Shahryar Kianian and Volkan Isler and Ce Yang},
  title = {{Evaluation of Stem Rust Disease in Wheat Fields by Drone Hyperspectral Imaging}},
  booktitle = {Sensors},
  year = {2023},
  url = {https://openreview.net/forum?id=Q9tCNQEtal6},
}
Figure