Hyperspectral Analysis for Discriminating Herbicide Site of Action: A Novel Approach for Accelerating Herbicide Research

Zhongzhong Niu, Tanzeel U. Rehman, Julie M. Young, William G. Johnson, Takayuki Yokoo, Bryan G. Young, Jian JinSensors 2023 · DOI

TL;DR

This paper introduces a novel approach using hyperspectral imaging and machine learning to rapidly and accurately classify herbicide "sites of action" (SOAs) in plants. By analyzing spectral signatures, the method can identify herbicide effects much earlier than traditional methods, potentially accelerating the discovery of new herbicides and combating weed resistance.

What Problem Does This Paper Address?

The agricultural industry faces a critical challenge: the development of herbicide-resistant weeds. For decades, innovation in herbicide modes of action (MOAs) has stagnated, largely due to the high cost and time involved in discovering and registering new active ingredients. While MOA describes the general mechanism of a herbicide, the site of action (SOA) is a more specific target within a plant's biochemical pathways. Identifying unique SOAs is crucial for developing new herbicides that can overcome existing resistance. Traditional methods for this are slow and expensive, hindering progress. This paper aims to bridge this gap by leveraging hyperspectral imaging and machine learning to provide a faster, more efficient way to characterize herbicide effects and identify novel SOAs.

Methods

The core of this research lies in combining hyperspectral imaging with machine learning for herbicide SOA classification.

  • Hyperspectral Imaging System: The researchers developed an automated, high-throughput imaging system. This system uses a line-scanning hyperspectral camera (MSV-500) capturing data from 380-1030 nm, mounted on a conveyor-based plant transportation system within a controlled greenhouse environment. This setup aims to minimize microclimate variability and ensure high-quality data.
  • Image Processing: Raw hyperspectral images were calibrated to correct for illumination non-uniformity. A red-edge segmentation algorithm was then used to isolate plant pixels from the background, followed by the extraction of mean spectral reflectance for each plant. Data was further pre-processed using a Savitzky-Golay smoothing filter.
  • Feature Selection: A Random Forest algorithm was employed to identify the most informative spectral bands for differentiating between herbicide treatments. This was done using a "one-vs-all" approach, where each treatment was considered a separate class.
  • Machine Learning Classification: Two algorithms were compared: Partial Least Squares Discriminant Analysis (PLS-DA) and Support Vector Machines (SVM). Both models underwent rigorous pre-processing and were trained using leave-one-out cross-validation. The study explored both linear and radial basis function (RBF) kernels for SVM. For classifying multiple SOAs, a one-vs-one (OVO) paradigm with soft voting was used.
  • Day-to-Day Validation: To assess model robustness, a day-to-day validation strategy was implemented, training models on data from one day and testing them on data from subsequent days.

Datasets

The study utilized a dataset comprising hyperspectral images of barnyard grass (Echinochloa crus-galli) treated with eight different herbicides, each representing distinct SOAs, plus a non-treated control group.

  • Herbicide Treatments: Eight herbicides were selected to represent four MOA groups, with at least two unique SOAs per group. These included Atrazine, Dinoseb, Flumioxazin, Paraquat, Chlorimuron, Glufosinate, Glyphosate, and Indaziflam.
  • Experimental Design: Plants were grown in pots and treated 10 days after sowing. Hyperspectral imaging was conducted daily from the second hour after treatment (1 DAT) up to 7 DAT. The experiment was conducted in two rounds, resulting in a cumulative dataset of 320 samples across nine treatments (eight herbicides + control).

Key Insights and Ideas

  • Early Detection is Key: Hyperspectral imaging can detect herbicide stress signatures as early as 1 day after treatment (DAT), significantly accelerating the screening process compared to traditional methods that might take weeks.
  • Spectral Signatures are Unique: Different herbicide SOAs leave distinct spectral fingerprints on plants, allowing for their differentiation. This is a fundamental principle that the study exploits.
  • Machine Learning is Crucial: Advanced machine learning models, particularly SVM, are essential for interpreting the complex spectral data and achieving high classification accuracy.
  • Temporal Dynamics Matter: The effectiveness of classification varies with time after treatment. Some herbicides show rapid effects (e.g., Paraquat), while others have slower, cumulative impacts (e.g., Glufosinate), requiring careful consideration of data collection timing.
  • SVM Outperforms PLS-DA: The study found SVM models generally achieved higher accuracy than PLS-DA, likely due to SVM's ability to handle the non-linear relationships between spectral responses and herbicide SOAs.

Key Results

  • High Accuracy Achieved: The final models, trained on a combined dataset from two rounds, achieved an overall accuracy of 81.5% for classification 1 DAT.
  • Early Classification Success: The study demonstrated the potential for high-precision models to screen numerous compounds in as little as 7 days.
  • SVM Performance: SVM models consistently showed higher accuracy, particularly for the control group (exceeding 90%), with most misclassifications involving confusing healthy plants with herbicide-treated ones.
  • Specific Herbicide Performance: Indaziflam (Cellulose synthesis) and Atrazine (PS II inhibition) showed particularly strong classification accuracy (up to 96.9% on 1 DAT).
  • Day-to-Day Validation Challenges: While overall accuracy varied, the study highlighted that some herbicides (e.g., Paraquat, Dinoseb) show rapid effects, making early detection crucial, while others (e.g., Chlorimuron) have slower effects, requiring later data points for optimal classification.

Limitations and Future Work

  • Limited Herbicide Range: While eight herbicides were tested, a broader spectrum of herbicides and SOAs would further validate the approach.
  • Spatial Information: The current study primarily relies on average spectral data. The authors suggest that incorporating spatial information (e.g., using tools like LeafSpec) could capture localized stress signals and improve accuracy, especially for herbicides that cause non-uniform damage.
  • Herbicide Dosage: The study used specific dosages. Future work could explore the impact of varying herbicide concentrations on spectral signatures and classification accuracy.
  • Data Collection Timing: The optimal time for classification varies by herbicide SOA. Future research could focus on developing models that integrate temporal data more effectively or identifying specific time windows for different herbicide classes.
  • Database Development: The authors emphasize the need for a well-curated hyperspectral database for herbicide SOAs to significantly accelerate herbicide development.
  • Focus on Early Detection: The paper highlights the potential of early detection, suggesting that future research should focus on data collection within the 1-2 week post-treatment window for many herbicides.

Summary updated

Abstract

In agricultural weed management, herbicides are indispensable, yet innovation in their modes of action (MOA)-the general mechanisms affecting plant processes-has slowed. A finer classification within MOA is the site of action (SOA), the specific biochemical pathway in plants targeted by herbicides. The primary objectives of this study were to evaluate the efficacy of hyperspectral imaging in the early detection of herbicide stress and to assess its potential in accelerating the herbicide development process by identifying unique herbicide sites of action (SOA). Employing a novel SOA classification method, eight herbicides with unique SOAs were examined via an automated, high-throughput imaging system equipped with a conveyor-based plant transportation at Purdue University. This is one of the earliest trials to test hyperspectral imaging on a large number of herbicides, and the study aimed to explore the earliest herbicide stress detection/classification date and accelerate the speed of herbicide development. The final models, trained on a dataset with nine treatments with 320 samples in two rounds, achieved an overall accuracy of 81.5% 1 day after treatment. With the high-precision models and rapid screening of numerous compounds in only 7 days, the study results suggest that hyperspectral technology combined with machine learning can contribute to the discovery of new herbicide MOA and help address the challenges associated with herbicide resistance. Although no public research to date has used hyperspectral technology to classify herbicide SOA, the successful evaluation of herbicide damage to crops provides hope to accelerate the progress of herbicide development.
BibTeX
@inproceedings{niu2023hyperspectral,
  author = {Zhongzhong Niu and Tanzeel U. Rehman and Julie M. Young and William G. Johnson and Takayuki Yokoo and Bryan G. Young and Jian Jin},
  title = {{Hyperspectral Analysis for Discriminating Herbicide Site of Action: A Novel Approach for Accelerating Herbicide Research}},
  booktitle = {Sensors},
  year = {2023},
  doi = {10.3390/s23239300},
}
Figure