Evaluation of rice bacterial blight severity from lab to field with hyperspectral imaging technique

Xiulin Bai, Yujie Zhou, Xuping Feng, Mingzhu Tao, Jinnuo Zhang, Shuiguang Deng, Binggan Lou, Guofeng Yang, Qingguan Wu, Yu Li, Yang Liao, Yong HeFrontiers in Plant Science 2022 · DOI

TL;DR

This paper introduces a novel hyperspectral imaging approach to accurately assess rice bacterial blight (BB) severity, bridging the gap between controlled lab environments and complex field conditions. By combining a deep learning model with an attention mechanism and a new spectral index (NDSCI), the researchers achieved high accuracy in disease evaluation and demonstrated its scalability to real-world field data.

What Problem Does This Paper Address?

Rice bacterial blight (BB) is a significant threat to global rice production, causing substantial yield losses. While disease-resistant breeding is a crucial strategy for control, accurately and efficiently measuring disease severity is a bottleneck. Traditional methods are time-consuming and labor-intensive, and existing hyperspectral imaging studies are often confined to lab settings, limiting their practical application in dynamic field environments. This research aims to develop a robust and scalable method for evaluating BB severity that can transition from the lab to the field, facilitating faster and more accurate disease assessment for breeding programs.

Methods

The study employed a multi-faceted approach:

  • Hyperspectral Imaging (HSI): Visible/near-infrared (VIS-NIR) hyperspectral images were captured to obtain detailed spectral information from rice leaves.
  • Deep Learning Model (ATT-LSTM): A custom Long Short-Term Memory (LSTM) network integrated with an attention mechanism was developed. The attention block was crucial for identifying and weighting the most informative spectral bands, effectively filtering out redundant information.
  • Feature Extraction: Both univariate feature selection (UFS) and recursive feature elimination (RFE) were explored alongside the attention mechanism for identifying important wavelengths. Color features (first, second, and third-order moments of RGB channels) were also extracted.
  • Data Fusion: The study investigated the benefits of combining spectral data with color features to improve disease severity evaluation.
  • Novel Disease Index (NDSCI): A new index, the Normalized Difference Spectral Color Index (NDSCI), was constructed by fusing the most important wavelength identified by the ATT-LSTM model and the color feature most correlated with disease severity.

Datasets

The study utilized hyperspectral images of rice leaves from three cultivars (IR24, 3A26, and 4A37) infected with bacterial blight. These images were collected at various time points after inoculation in both a controlled experimental field and a real-world field environment using an unmanned aerial vehicle (UAV) platform. A total of 450 samples were collected from the experimental field.

Key Insights and Ideas

  • Attention Mechanism for Feature Selection: The attention mechanism in the ATT-LSTM model proved highly effective in identifying critical spectral bands for disease severity assessment, outperforming traditional feature selection methods in some cases. This highlights the power of deep learning in uncovering subtle spectral signatures.
  • Data Fusion Enhances Accuracy: Combining spectral data with color features significantly improved the accuracy of disease severity evaluation compared to using either data source alone. This suggests that a multi-modal approach captures a more comprehensive picture of the disease's impact.
  • NDSCI's Scalability: The newly developed NDSCI index demonstrated strong correlation with disease severity in both lab-derived and field-acquired data, showcasing its potential for bridging the gap between controlled experiments and real-world applications. This is a critical step towards practical field deployment.
  • Wavelength 682nm is Key: The study identified a specific wavelength around 682nm as particularly important for detecting BB severity, a finding supported by both ANOVA and the attention mechanism's weight distribution.

Key Results

  • The ATT-LSTM model achieved a maximum accuracy of 0.94 when fusing important wavelengths and color features.
  • When using only spectral data, the ATT-LSTM model achieved an accuracy of 0.89, outperforming the SVM model.
  • Data fusion improved the accuracy of the ATT-LSTM model by 4.49% and the SVM model by 1.16% compared to using spectral data alone.
  • The newly proposed NDSCI index showed a high correlation coefficient of -0.93 with disease severity in lab settings.
  • Crucially, the NDSCI demonstrated good scalability to field data, with a correlation coefficient of -0.84. This is significantly better than other common spectral indices tested for field application.

Limitations and Future Work

  • Environmental Variability: While the study demonstrated scalability, the authors acknowledge that field environments are complex and affected by factors like sunlight intensity and wind. The important features identified in the lab may differ from those in the field, though NDSCI showed robustness.
  • Data Fusion Exploration: The current data fusion was described as "primary," suggesting that more advanced feature-level and decision-level fusion strategies could be explored in future work to further enhance performance.
  • Further Field Validation: While the field validation was promising, continued research and validation in diverse field conditions are necessary to fully establish the robustness of the NDSCI and the overall approach.
  • Broader Application: The authors suggest that this work provides a foundation for future research into the rapid measurement of plant phenotypes in disease-resistant breeding, implying potential for application to other diseases and crops.

Summary updated

Abstract

Hyperspectral imaging technique combined with machine learning is a powerful tool for the evaluation of disease phenotype in rice disease-resistant breeding. However, the current studies are almost carried out in the lab environment, which is difficult to apply to the field environment. In this paper, we used visible/near-infrared hyperspectral images to analysis the severity of rice bacterial blight (BB) and proposed a novel disease index construction strategy (NDSCI) for field application. A designed long short-term memory network with attention mechanism could evaluate the BB severity robustly, and the attention block could filter important wavelengths. Best results were obtained based on the fusion of important wavelengths and color features with an accuracy of 0.94. Then, NSDCI was constructed based on the important wavelength and color feature related to BB severity. The correlation coefficient of NDSCI extended to the field data reached -0.84, showing good scalability. This work overcomes the limitations of environmental conditions and sheds new light on the rapid measurement of phenotype in disease-resistant breeding.
BibTeX
@inproceedings{bai2022evaluation,
  author = {Xiulin Bai and Yujie Zhou and Xuping Feng and Mingzhu Tao and Jinnuo Zhang and Shuiguang Deng and Binggan Lou and Guofeng Yang and Qingguan Wu and Yu Li and Yang Liao and Yong He},
  title = {{Evaluation of rice bacterial blight severity from lab to field with hyperspectral imaging technique}},
  booktitle = {Frontiers in Plant Science},
  year = {2022},
  doi = {10.3389/fpls.2022.1037774},
}
Figure