FHBDSR-Net: automated measurement of diseased spikelet rate of Fusarium Head Blight on wheat spikes

Ze Wu, Huijun Zhao, Zeyu Chen, Yang Suo, Stephen Joseph, Xiaohui Yuan, Caixia Lan, Weizhen LiuaBIOTECH 2025 · DOI

TL;DR

This paper introduces FHBDSR-Net, a novel lightweight deep learning framework for automatically measuring the diseased spikelet rate (DSR) of Fusarium Head Blight (FHB) in wheat. By developing a new spikelet-level annotated dataset and a specialized network architecture, FHBDSR-Net achieves high accuracy in detecting diseased spikelets, paving the way for efficient, on-site plant phenotyping and accelerated breeding programs.

What Problem Does This Paper Address?

Fusarium Head Blight (FHB) is a devastating fungal disease that significantly impacts global wheat production and food security. Accurately quantifying the severity of FHB is crucial for breeding resistant wheat varieties. The standard method for this, the Diseased Spikelet Rate (DSR), involves tedious, manual, and destructive counting of infected spikelets on wheat spikes. While deep learning has shown promise for automated disease detection, existing methods struggle with FHB due to a lack of detailed spikelet-level annotated data, difficulty in representing subtle diseased spikelet features, and challenges in encoding spatial relationships within densely packed spikelets. This paper aims to bridge this gap by developing an automated, accurate, and lightweight system for spikelet-level DSR measurement.

Methods

The core of this work is the FHBDSR-Net, a lightweight object detection network designed for precise spikelet detection. Its architecture incorporates three key innovations:

  • Multi-scale Feature Enhancement (MFE) module: This module dynamically combines lesion textures, morphological features, and lesion-awn contrast using adaptive multi-scale kernels. It aims to suppress background noise and improve the representation of diseased spikelets by focusing on relevant features at different scales.
  • Inner-EfficiCIoU loss function: This novel loss function is designed to reduce localization errors, particularly for small targets in dense environments. It combines aspects of Inner-IoU (for accelerated regression) and EfficiCIoU (for aspect ratio considerations) to improve bounding box accuracy.
  • Scale-Aware Attention (SAA) module: This module uses dilated convolutions and self-attention mechanisms to encode multi-scale pathological patterns and spatial distributions. It's crucial for enhancing the resolution and understanding of densely packed spikelets, effectively capturing their spatial relationships.

The overall pipeline involves training FHBDSR-Net on the new dataset, using its predictions to count diseased and healthy spikelets, and then calculating the DSR.

Datasets

The authors constructed a novel dataset specifically for this research. It comprises:

  • 620 high-resolution RGB images of wheat spikes.
  • 5,222 spikelet-level annotations, providing detailed bounding boxes for individual diseased and healthy spikelets.

This dataset is a significant contribution, addressing the critical lack of spikelet-level annotations that hampers existing FHB detection methods. The dataset is split into training, validation, and testing sets at an 8:1:1 ratio.

Key Insights and Ideas

  • Spikelet-level annotation is critical: The paper strongly emphasizes that existing datasets and models often focus on spike-level detection, which is insufficient for precise DSR measurement. The creation of a spikelet-level annotated dataset is a foundational step for advancing FHB research.
  • Dense spikelet detection is a unique challenge: The close proximity and overlapping nature of spikelets require specialized attention mechanisms and loss functions to accurately distinguish and localize individual units, especially when diseased.
  • Lightweight design is key for practical deployment: The authors prioritize a compact model (7.2M parameters) that can run on resource-constrained devices, enabling real-time, on-site phenotyping in agricultural settings.
  • Multi-scale feature fusion is essential for small objects: The MFE and SAA modules highlight the importance of effectively integrating information from different scales to accurately detect and characterize small, subtle features like diseased spikelets.

Key Results

  • High Detection Accuracy: FHBDSR-Net achieved an average precision of 93.8% for detecting diseased spikelets and 89.2% for healthy spikelets on their custom dataset.
  • Lightweight and Efficient: The model boasts a small parameter count of 7.2 million, making it significantly lighter than many state-of-the-art detectors.
  • Strong Correlation with Experts: The DSR predictions from FHBDSR-Net showed a strong correlation with expert evaluations, with a Pearson correlation coefficient of 0.901.
  • Competitive Performance: Compared to other lightweight detectors like YOLOv5-S and GELAN-S, FHBDSR-Net offers a better balance of accuracy and scale. It also achieved significant parameter reduction (e.g., 80% compared to YOLOv7) while improving mAP.
  • Robustness: Visual comparisons demonstrated FHBDSR-Net's ability to effectively mitigate overlapping and missed detections in complex backgrounds, outperforming several other detectors.

Limitations and Future Work

  • Dataset Bias: While the dataset includes complex backgrounds and lighting variations, it was primarily collected in greenhouse environments. The authors acknowledge that field samples present more diverse challenges (pathogen infection, disease symptoms) that could impact generalization.
  • Single-Image Inference: The current model relies on single-image inference, which has inherent perspective limitations.
  • Future Directions: The authors suggest expanding the dataset to include more diverse field conditions and uncontrolled factors like lighting and weather. They also propose developing multi-view, multi-dimensional FHB detection models for even higher precision and exploring more efficient learning paradigms to reduce dependence on data quality and costs.

Summary updated

Abstract

Fusarium Head Blight (FHB), a fungal wheat (<i>Triticum aestivum</i>) disease that threatens global food security, requires precise quantification of diseased spikelet rate (DSR) as a phenotypic indicator for resistance breeding. Most techniques for measuring DSR rely on manual spikelet-by-spikelet observation and counting, which is inefficient and destructive. Although deep learning offers great promise for automated DSR measurement, existing intelligent detection algorithms are hampered by the lack of spikelet-level annotated data, insufficient feature representation for diseased spikelets, and weak spatial encoding of densely arranged spikelets. To address these challenges, we constructed a dataset of 620 high-resolution RGB images of wheat spikes with 5,222 spikelet-level annotations to systematically analyze spikelet size distributions to fill small-object detection data gaps in this field. We designed FHBDSR-Net, a light framework for automated DSR measurement centered on diseased spikelet detection, which features (1) multi-scale feature enhancement architecture that dynamically combines lesion textures, morphological features, and lesion-awn contrast through adaptive multi-scale kernels to suppress background noise; (2) the Inner-EfficiCIoU loss function to reduce small-target localization errors in dense contexts; and (3) a scale-aware attention module using dilated convolutions and self-attention to encode multi-scale pathological patterns and spatial distributions to enhance dense spikelet resolution. FHBDSR-Net detected diseased spikelets with an average precision of 93.8% with a lightweight design of 7.2 M parameters. The results were strongly correlated with expert evaluations, with a Pearson correlation coefficient of 0.901. Our method is suitable for deployment on resource-constrained mobile devices, facilitating portable plant phenotyping and smart breeding.
BibTeX
@inproceedings{wu2025fhbdsr,
  author = {Ze Wu and Huijun Zhao and Zeyu Chen and Yang Suo and Stephen Joseph and Xiaohui Yuan and Caixia Lan and Weizhen Liu},
  title = {{FHBDSR-Net: automated measurement of diseased spikelet rate of Fusarium Head Blight on wheat spikes}},
  booktitle = {aBIOTECH},
  year = {2025},
  doi = {10.1007/s42994-025-00245-0},
}
Figure