The wild rice germplasm was provided by the Yazhou Bay National Laboratory and the Institute of Crop Science, Chinese Academy of Agricultural Sciences, and is stored at the National Wild Rice Germplasm Repository in Sanya, Hainan Province. Samples were collected from regions including Guangxi, Hainan, and Yunnan, encompassing both ordinary and medicinal wild rice varieties. Disease plots were established on February 18, 2023, and March 3, 2024, at the Potianyang base in Yazhou District, Sanya City (Latitude 18°23′36″, Longitude 109°9′52″). Based on their growth status, tillering, and population distribution, 120 and 50 representative and diverse wild rice samples were transplanted. After transplanting, compound fertilizer was applied one week later, with regular water changes, field cleaning, and pest control carried out.
After tillering reached the reproductive stage, wild rice was inoculated with the bacterial blight pathogen Xanthomonas oryzae pv. oryzae (strain PXO99, a highly virulent, widely pathogenic Philippine race 6) following the Rice Bacterial Blight Resistance Identification Technical Standards. The bacteria, strain PXO99, known for its high virulence and broad pathogenicity, were cultured on PSA medium (10 g/L tryptone, 10 g/L sucrose, 1 g/L monosodium glutamate, 15 g/L agar, pH 7.0) for 2 days, then washed with sterile water and diluted to an OD600 of 0.6. Using scissors dipped in the bacterial solution, leaves of individual wild rice plants were cut 3–5 cm from the leaf tip, with at least 7–10 leaves inoculated per sample. Jingang 30, a susceptible variety, was used as a positive control, with inoculation considered valid only if the control exhibited symptoms.
RGB images of rice infected with BB were captured at different time points using DJI Mini 2 (DJI Innovations, Shenzhen, China) and DJI Mini 4 Pro (DJI Innovations, Shenzhen, China) UAVs. The flight altitude was set between 0.6 m and 1.5 m above the wild rice field, with the camera gimbal pitch adjusted to -90° to -60°. The images had a resolution of 1920×1080 pixels and were saved in JPG format. Image acquisition occurred between 5 and 21 days post-infection, during the morning (8:00–11:00) and afternoon (16:30–18:30), under weather conditions such as sunny, cloudy, and overcast, with ambient temperatures ranging from 22 °C to 30 °C. All images were taken under natural field conditions with ambient lighting, without the use of a flash. The images may contain varying degrees of obstruction, water surface reflection, overexposure, as well as noise from weeds, dead leaves, bird droppings, and other field debris. Each image contained between 1/4 and 2 wild rice plants. In total, 2,035 images were captured. Their quality and accuracy for representing BB symptoms were verified by two wild rice germplasm identification experts, confirming that the depicted lesions were definitively BB.
Data processing of the UAV images was carried out to ensure high-quality input for the model. To mitigate noise, overexposure, and occlusion, specific preprocessing techniques were applied. A Gaussian blur with a kernel size of 5×5 and a standard deviation (σ) of 1.5 was applied to reduce high-frequency noise. Morphological opening operations using a 3×3 rectangular structuring element were employed to clean small debris and refine the boundaries of detected lesions. Additionally, Z-score normalization was performed to standardize lighting conditions across the images by subtracting the mean and dividing by the standard deviation of the dataset, thereby reducing the effect of varying weather and light conditions during UAV capture.
The process of material preparation and image acquisition.
To meet the input and training requirements of deep learning models, this study employed an image cropping strategy, dividing the raw UAV-captured images into multiple 640×640 pixel blocks. A sliding window method with a fixed step size was used to ensure that each cropped region covered different parts of the original image, creating independent training samples. Additionally, to enhance dataset diversity and model robustness, data augmentation techniques were applied, including random rotations, horizontal and vertical flips, scaling, and color adjustments. This resulted in a total of 12,210 images. The dataset, comprising images collected from both the 2023 and 2024 growing seasons, was pooled together and randomly shuffled. It was then split into training (9,768 images), validation (1,221 images), and testing (1,221 images) sets in an 8:1:1 ratio. This random splitting strategy ensures that the model learns from a diverse range of environmental conditions present in both years. The BB lesion areas on wild rice were manually annotated using the open-source software roLabelImg, with the annotation information saved in.xml format.
YOLOv11, introduced in 2024, is an advanced object detection model that offers various versions, including the standard YOLOv11, YOLOv11-OBB, YOLOv11-Seg for segmentation tasks, and YOLOv11-Pose for pose estimation (
The UAV captures images from above, often presenting targets at various angles. Additionally, bacterial blight (BB) typically manifests as elongated lesions. Considering these factors, the YOLOv11-OBB model was selected. To optimize model efficiency and reduce computational complexity, the smallest variant, the n model, was chosen as the foundation for improvements. As shown in
Structure of the BB detection model.
The key improvements are outlined below:
C3k2FC Module in Backbone: Integrates PConv and CGLU into the C3k2 structure to reduce redundancy and enhance efficiency.
SPPF_LSKA Module in Backbone: Incorporates LSKA post-pooling to improve multi-scale feature aggregation and reduce noise.
SlimNeck in Neck: Utilizes GSConv and VoVGSCSP to optimize feature fusion while minimizing FLOPs.
LiteHead in Detection Head: Employs separated Batch Normalization (BN) and dynamic anchors for improved handling of rotations.
LAMP Pruning: Implements adaptive global pruning to compress model parameters while maintaining accuracy.
The C3k2 module in YOLOv11-OBB, based on the Cross-Stage Partial (CSP) structure, splits the input feature map into two parallel branches. This design enhances feature reuse and gradient flow by repeating bottleneck units, supporting flexible configurations of layers and channels to balance model depth and computational efficiency. However, this module suffers from redundant convolution operations, increasing computational cost and complicating deployment on resource-constrained UAV embedded systems.
To address these issues, this paper integrates the FasterNet Block’s Partial Convolution (PConv) and the TransNext CGLU module to enhance the C3k2 module in both the Backbone and Neck networks (
In this study, we replace the two 1×1 pointwise convolutions in the FasterNet Block with the CGLU module, forming an improved FC structure. We also integrate PConv into the depthwise convolution section of the C3k2 bottleneck, resulting in the improved C3k2FC module, as shown in
Structure of C3k2FC.
The SPPF module in YOLOv11, based on the spatial pyramid pooling (SPPF) structure, aggregates multi-scale features by applying maximum pooling operations of different sizes in parallel. These features are then fused through channel concatenation and convolution to extract global context. The module aims to reduce information loss and enhance adaptability to targets of varying scales, supporting efficient feature representation. However, the pooling operations in this module have limitations, leading to inefficient feature aggregation in complex scenes and the potential introduction of background noise. This is especially problematic when dealing with UAV-captured rice field images that involve varying lighting, leaf occlusion, and resolution fluctuations, which may cause fine-grained details to be overlooked or irrelevant features to be amplified, resulting in false detections and degraded performance.
To address these challenges, we propose the integration of the Large Separable Kernel Attention (LSKA) mechanism to improve the SPPF module in the Backbone network, achieving a lightweight design while enhancing multi-scale feature extraction and local attention capabilities. The LSKA module, based on separable convolutions, decomposes large kernel convolutions into 1D separable operations in the horizontal and vertical directions to simulate large receptive field attention. This approach, combined with depth-wise separable convolutions, reduces parameter count.
The SPPF_LSKA module integrates the LSKA mechanism (
Structure of SPPF_LSKA.
This study replaces the original SPPF in YOLOv11-OBB with the improved SPPF_LSKA module, optimizing multi-scale feature integration through attention mechanisms to support wild rice BB detection and localization from UAV perspectives.
The standard convolution (SC) component in YOLOv11-OBB captures features by processing multi-channel data through parallel multi-kernel operations. While this promotes deep interactions between channels, it also leads to parameter inflation and a significant increase in floating-point operations (FLOPs), which limits real-time performance, especially on resource-constrained UAV embedded systems. On the other hand, efficient architectures like MobileNet and ShuffleNet alleviate the computational burden by using depthwise separable convolution (DSC). However, DSC isolates channel data during computation, which significantly reduces feature integration and extraction effectiveness. This limitation is problematic for wild rice BB detection from UAV perspectives, where noise from lighting changes, leaf occlusion, and resolution fluctuations can lead to missed or mislocalized small lesions.
To accelerate computation without compromising detection accuracy, this study integrates the GSConv composite convolution unit into the Neck structure of YOLOv11-OBB. The GSConv unit incorporates a Shuffle mechanism to merge the channel association data derived from SC with the spatial output from DSC (
The core components of the GSConv unit include Conv, DWConv, Concat, and Shuffle operations. The architecture is constructed through the following process: The input feature matrix with C1 channels is first processed by applying DSC to half of the channels to capture spatial details, while SC is applied to the other half for channel fusion. The outputs from both sides are then concatenated along the channel axis. Subsequently, a Shuffle operation is applied to the concatenated result to promote random channel reorganization and enhance data interaction. The final output matrix contains C2 channels. The VoVGSCSP, an iterative bottleneck combination for GSConv, splits the input matrix channels into two segments. One segment undergoes Conv preprocessing and is passed through a series of GS bottleneck units to extract features, while the other segment serves as a residual path with only a single Conv transformation. This unit, leveraging the CSP framework, enhances feature reuse, employs a one-shot aggregation method to minimize data loss, and integrates batch normalization and SiLU activation to maintain feature consistency.
Finally, this study reconstructs the Neck framework of YOLOv11-OBB using GSConv and VoVGSCSP to form the SlimNeck architecture. These modifications reduce the model’s computational burden, enabling faster inference and data processing while enhancing multi-resolution feature integration. This optimization strikes an effective balance between speed and accuracy.
The detection head of YOLOv11-OBB uses a shared convolution mechanism to process multi-scale feature outputs for boundary box predictions. While this promotes parameter reuse, the use of unified Batch Normalization (BN) can lead to inaccurate moving averages when there are significant statistical differences between features at different levels. This can negatively affect training stability and generalization performance. Additionally, the static anchor box design struggles to adapt to the rotation and scale variations of BB lesions, under the challenges posed by UAVs, such as leaf occlusion, resolution fluctuations, and changing lighting conditions. This often results in localization errors or missed small targets, limiting overall detection performance.
To address these issues, this paper introduces the LiteHead detection head. This optimization is designed to improve computational efficiency, reduce parameter size and operational load, and make the system more suitable for resource-constrained UAV embedded systems. The LiteHead detection head mitigates BN-related issues by separately processing statistical differences between features at different levels, ensuring independent sliding averages and preventing bias accumulation in shared parameters. Additionally, the dynamic anchor box generator adapts the scale, angle, and aspect ratio of anchors based on input features, enhancing the alignment with rotating targets.
The core components of the LiteHead detection head include: a shared convolution layer for multi-task reuse to reduce redundant parameters; a separated BN module that applies BN independently to each detection branch, optimizing feature distribution and speeding up convergence; and a dynamic anchor box generator that computes anchor parameters in real-time using feature map statistics, supporting the prediction of rotated bounding boxes (OBBs). The workflow is as follows: multi-scale feature maps are processed by the shared convolution to extract boundary box regression and classification scores; separated BN standardizes each layer’s output; and the dynamic anchor box module generates adaptive anchors based on feature clustering or gradient guidance, ultimately outputting the OBB predictions.
In this study, LiteHead is integrated into the detection head of YOLOv11-OBB by replacing the standard shared convolution and static anchor boxes. This design optimizes the feature processing pipeline, reduces floating-point operations and memory access, and enhances the ability to capture rotational details and global information. The LiteHead architecture is well-suited for BB detection and localization in UAV-captured rice field images.
Despite the introduction of lightweight modules, the YOLOv11-OBB model still faces parameter redundancy and computational intensity, resulting in increased floating-point operations (GFLOPs) and memory requirements. This issue is particularly problematic for resource-constrained UAV embedded systems, where it can lead to inference delays and deployment bottlenecks.
To address these challenges, this study introduces Layer-adaptive Magnitude-based Pruning (LAMP) to further optimize the YOLOv11-OBB model (
Process of LAMP.
The LAMP pruning process proceeds in four stages: pre-training the complete model to obtain initial weights; calculating the global LAMP scores to sort all weights and set thresholds according to the target sparsity; applying magnitude-based pruning layer-by-layer to remove connections below the threshold; and finally fine-tuning the pruned model to restore accuracy, typically using techniques like knowledge distillation or learning rate scheduling.
Compared to traditional magnitude pruning, LAMP improves accuracy retention after pruning by using the layer-adaptive mechanism. This allows for significant reductions in parameters and GFLOPs without sacrificing detection performance.
In this study, LAMP is applied to the Backbone, Neck, and Detection Head of YOLOv11-OBB, resulting in a lightweight variant through iterative pruning and fine-tuning. This optimized structure enhances resource utilization and supports the capture of both fine details and global information, making it well-suited for wild rice BB detection and localization from UAV perspectives.
The continuous movement of UAVs during video capture often results in partial occlusion of wild rice plants or the redundancy of the same BB target appearing across multiple frames. Consequently, relying solely on static detection from individual images can lead to missed, incorrect, or duplicate counts, thereby compromising the accuracy and stability of the screening process. To address these challenges, we propose an integrated screening workflow. As illustrated in
The overall framework of the proposed XooNet method.
The user controls the UAV to fly over the wild rice disease field, adjusting the camera’s focus and gimbal angle, and then activates the recording function. The captured video data is transmitted to the edge computing device via either wireless or wired connection.
The video recorded by the UAV is processed through the BB detection model proposed in this paper, which detects and generates corresponding Oriented Bounding Boxes.
A multi-object tracking model is introduced, utilizing the Kalman Filter to predict the location of targets. The Kalman Filter is a recursive estimation algorithm that uses statistical inference based on historical data to estimate the current position of a target. This prediction provides prior information for disease detection in subsequent frames, reducing bias due to occlusion or detection errors.
The Kalman filter state update equation is expressed in
where
The Hungarian Algorithm is used to match the current frame’s detection results with the Kalman Filter’s predicted positions. The matching score, representing the similarity between targets, is calculated as shown in
where IoU_OBB represents the Intersection over Union calculated based on the Oriented Bounding Boxes (OBB) to accurately account for the rotation of the long, narrow lesions.
The BB targets in the tracking list, each with a unique identifier, are counted for BB occurrences.
Based on the disease count and the number of disease inoculations in wild rice, the disease incidence rate is calculated using
The incidence rate is then used to classify wild rice germplasm into BB-resistant categories, as shown in
Classification criteria for BB-resistant wild rice germplasm.
| Disease grade | Incidence rate (%) |
|---|---|
| L1 | 0≤ p ≤ 20 |
| L2 | 20< p ≤ 50 |
| L3 | 50< p ≤ 80 |
| L4 | 80< p ≤ 100 |
Wild rice classified as L1 or L2 is selected as prime candidates for resistance to BB.
The experiments were conducted on a Dell tower workstation (Dell, Inc., Round Rock, Texas, USA) running Windows 11. The system was powered by a 12th Gen Intel(R) Core(TM) i5–12500 processor with a clock speed of 3.00 GHz, supported by 64GB of RAM and a 1TB solid-state drive. For GPU-accelerated computations, an NVIDIA GeForce RTX 3080 graphics card (NVIDIA Corporation, Santa Clara, California, USA) with 10GB of video memory was used. The software environment included Python version 3.8.17, along with PyTorch 1.13.0, Torchvision 0.14.0, and CUDA 11.7.
During the model calibration process, hyperparameter optimization was conducted on the validation set. The hyperparameters tested included the learning rate, batch size, and optimizer settings. The Adam optimizer was used with an initial learning rate of 1e-3, a maximum learning rate of 1e-5, a momentum coefficient of 0.937, and a weight decay parameter of 5e-4. The batch size was set to 8, and the input images were resized to a resolution of 640×640 pixels.
For model validation, 5-fold cross-validation was performed to assess the model’s robustness. The dataset was divided into 5 subsets, and the model was trained and evaluated 5 times, each time using a different subset as the validation set and the remaining subsets for training. This process provided a reliable estimation of the model’s performance and ensured that the model generalizes well to unseen data.
The experimental procedure involved 300 epochs for each training run. The training parameters and dataset were kept consistent across all models throughout the training phase to ensure a fair comparison.
In the experiment on performance in edge computing devices, we first converted the model’s weight file to ONNX format, then optimized and added model nodes. Next, we used TensorRT to accelerate the model and deployed it on the Nvidia Jetson Nano. This compact, high-performance AI embedded development board, developed by Nvidia, features a quad-core ARM A57 CPU (1.43 GHz), a 128-core Maxwell GPU, and 4 GB of memory. The operating system used is Ubuntu 18.04, with the environment configured to JetPack 4.6, CUDA 10.2, cuDNN 8.2, and TensorRT 8.0.
To evaluate the performance of the proposed method, video data of 120 wild rice plants inoculated with BB, captured by UAVs in March 2023, were selected. The data were collected at 8, 10, 13, and 15 days post-inoculation. The results from the proposed method were compared with the manual counts performed by breeding experts on-site, providing a systematic assessment of the method’s accuracy.
To validate the effectiveness and stability of the screening method, a field application test was conducted on July 1, 2024, at the Potianyang Experimental Base in Yazhou District, Sanya City, Hainan Province. This test assessed the disease condition of BB in 50 different wild rice samples. The participants included eight individuals: rice disease resistance breeding researchers, experts, local farmers responsible for managing the wild rice base, and graduate students in related fields. During the test, participants were divided into two groups: the validation group, composed of four rice disease resistance breeding researchers and local farmers, performed manual counting to assess the disease condition of the wild rice; the test group, consisting of two graduate students, operated the UAV to collect wild rice images from a height of 0.6 to 1.5 meters and utilized the proposed method for disease condition evaluation. The tests were conducted in the field, with no more than a 30-minute interval between the counting processes of both groups. Two rice disease resistance breeding experts supervised the testing procedures and results.