Abstract
Wheat stripe rust is a major threat to global wheat production, causing significant yield reductions and economic losses. Traditional convolutional neural network (CNN) detection methods often struggle with the disease's characteristic long, thin, and arbitrarily oriented pustules. This paper presents WSRD-Net, a refined single-stage rotation detector designed to address these challenges. By incorporating a Feature Refinement Module (FRM) and a Skew IoU (S-IoU) loss function, WSRD-Net achieves superior localization and detection accuracy. Experimental results on the WSRD2021 dataset demonstrate that WSRD-Net achieves 60.8% Average Precision (AP), outperforming existing state-of-the-art oriented and horizontal object detection models.
Introduction
Wheat stripe rust (Puccinia striiformis) manifests as bright yellow, linear pustules on wheat leaves. Early and accurate detection is vital for precision agriculture and yield protection. However, standard object detection frameworks typically utilize horizontal bounding boxes (HBB), which are ill-suited for the linear morphology of stripe rust. HBBs often include excessive background noise (healthy leaf or soil) and fail to distinguish overlapping diseased areas in dense field environments. WSRD-Net addresses these limitations by employing oriented bounding boxes (OBB) to tightly encapsulate diseased regions regardless of their angle.
Methodology
WSRD-Net is built upon the RetinaNet architecture, enhanced with several novel components to handle the unique geometry of stripe rust:
Horizontal-to-Oriented Anchor Mechanism
To balance computational efficiency and precision, the model utilizes a two-stage regression process. It starts with standard horizontal anchors in a "coarse stage" and subsequently regresses them into oriented boxes. This approach maintains high speed while achieving the precision required for narrow objects.
Feature Refinement Module (FRM)
A significant challenge in rotation-based detection is feature misalignment. As a bounding box rotates or changes shape during regression, the underlying feature map grid may no longer align with the object's center. The FRM employs bilinear interpolation to re-sample and reconstruct the feature map, ensuring that the features used for classification are perfectly aligned with the predicted rotated box.
Skew IoU (S-IoU) Loss
Standard IoU metrics are insufficient for rotated rectangles. WSRD-Net implements Skew IoU loss, which uses triangulation to accurately calculate the overlap between oriented polygons. This is particularly critical for objects with high aspect ratios, where minor angular errors (e.g., 5 degrees) can lead to a massive drop in overlap and detection failure.
Dataset: WSRD2021
The study introduces the Wheat Stripe Rust Dataset 2021 (WSRD2021), comprising 979 high-resolution RGB images captured via mobile phones in real-world field conditions. The dataset is uniquely annotated with oriented bounding boxes, capturing the center, dimensions, and angle of each disease instance. Over 75% of the instances exhibit an aspect ratio greater than 3, reflecting the challenging nature of stripe rust detection.
Key Results
WSRD-Net was evaluated against several state-of-the-art oriented detectors, including Faster R-CNN OBB and RoI Transformer. WSRD-Net achieved the highest performance with 60.8% AP and 73.8% Recall.
As shown in the figure above, the oriented bounding boxes (B) provide a much tighter fit to the diseased areas compared to horizontal boxes (A). This improved localization is essential for quantifying disease severity and enabling automated precision spraying. Furthermore, WSRD-Net maintains a competitive inference speed of 6.8 FPS on an NVIDIA TITAN RTX, making it faster than complex two-stage oriented detectors while providing superior localization.
Conclusion
WSRD-Net represents a significant advancement in the automated detection of wheat stripe rust. By leveraging oriented bounding boxes and specialized refinement modules, the model overcomes the limitations of traditional horizontal detectors in complex field environments. Future work will focus on expanding dataset diversity and integrating automated severity quantification metrics to further assist agricultural scientists and farmers in disease management.
Summary updated