Wild rice germplasm resources from the Institute of Crop Science, Chinese Academy of Agricultural Sciences, were planted at the Potianyang Base, Yazhou District, Sanya City, Hainan Province, China (N: 18°39′84.13″, E: 109°17′51.68″). On 1 March 2023, a total of 120 wild rice samples were artificially inoculated with the pathogen of rice bacterial blight. Following the “The technique rules for identification of rice variety resistance against bacterial blight (Xanthomonas oryzae pv. oryzae),” the inoculation was performed using the pathogenic strain PXO99A of Xanthomonas oryzae pv. oryzae, which is known to cause rice bacterial blight. The inoculation method involved manually cutting leaves during the tillering stage (Administration, A.P.M.S, 2017; Tang et al., 2017).

The camera-equipped UAV was utilized to acquire RGB images of wild rice at different times after infection with bacterial blight. The image acquisition method of wild rice bacterial blight is shown in Figure 1 . The type of UAV equipment is Dji Mini2 (DJI Inc, Shenzhen, China), the flying altitude is 0.6–1.5 m above the surface of the wild rice field, the camera head pitch angle is −90° to −60°, and the output image resolution is 1,920 × 1,080, which is saved in JPG format. Image acquisition was conducted at different time intervals, specifically on the fifth, seventh, ninth, 12th, 14th, 16th, and 18th days after infection. The acquisition periods were from 8:00 AM to 11:00 AM and from 4:30 PM to 6:30 PM. The weather conditions during image acquisition included sunny, cloudy, and overcast days, with environmental temperatures ranging from 22°C to 30°C. All images were captured in the natural field environment, utilizing natural lighting without the use of flash. The images contained various types of interference, such as different levels of occlusion, water surface reflections, and overexposure, as well as weeds, withered leaves, bird droppings, and field debris. Each image contained one-fourth to two wild rice plants. In total, 750 images of infected wild rice with bacterial blight were collected under the UAV viewpoint. The days since inoculation, figures, and image samples in the wild rice bacterial blight dataset are shown in Figure 2 . These images were further identified and confirmed by two experts specializing in wild rice germplasm identification. The open-source software roLabelImg (cgvict, 2017) was used to manually annotate wild rice leaf blight spot areas, and the annotation information was saved as a file in.xml format using this software. The dataset was divided and split in the ratio of 6:2:2, with 450 images randomly selected as the training set, 150 as the validation set, and 150 as the test set. Furthermore, the inclusion of mosaic augmentation is notable during training. Mosaic augmentation involves the random extraction of four images, where each image contributes only a portion of its content and corresponding detection box information. These fragments are then combined into a single image, serving as input for the network. This technique significantly diversifies the training data, effectively guarding against overfitting by introducing greater variability into the learning process.

Given the intricate backgrounds inherent in wild rice paddies—encompassing water, shadows, and reflections—and the dynamic alterations in the manifestation of visual disease attributes due to the interplay of UAV propeller airflow and variable weather conditions, the imperative for real-time detection underscores our decision to employ YOLOv8 as the baseline model for this research. In this context, Xoo-YOLO emerges as an enhanced iteration built upon YOLOv8, specifically tailored for the task of detecting wild rice bacterial blight from the UAV perspective in the field. The architecture of Xoo-YOLO is visualized in Figure 3 .

The model comprises four principal components: input, backbone, neck, and head. The enhancements incorporated are outlined as follows:

By amalgamating these advancements, Xoo-YOLO is equipped to address the intricacies associated with wild rice bacterial blight detection from the UAV viewpoint, offering improved accuracy and robustness. The whole algorithm is summarized in the pseudo-lcode in Algorithm 1 .

YOLOv8, a prominent object detection algorithm, was introduced by Ultralytics in January 2023. This algorithm has exhibited commendable outcomes in both speed and accuracy (Terven and Cordova-Esparza, 2023). The backbone of YOLOv8 centers around the C2f module, drawing inspiration from the ELAN module. The C2f module is constructed from two Conv modules and multiple Darknet BottleNeck modules, interconnected by Split and Concat modules. Additionally, YOLOv8 incorporates the Spatial Pyramid Pool Fusion (SPPF) module. This innovation transforms feature maps of varying sizes into fixed-size feature vectors, effectively preserving the original image’s features and positional information to the maximum extent. The YOLOv8’s neck network incorporates the Path Aggregation Network (PANet), similar to YOLOv7 (Wang et al., 2023), which fuses three effective feature layers obtained from the backbone network across layers of features, with the three effective feature layers located in the middle, lower middle, and bottom layers of the backbone network.

Owing to the intricate setting of wild rice paddies, encompassing water, mud, algae, weeds, bird droppings, shadows, and other intricate backgrounds, coupled with the relatively diminutive size of wild rice bacterial blight as viewed from UAV, relying on limited contextual information often triggers incorrect detection. For instance, it may lead to the misinterpretation of white streaks of debris within the wild rice paddies as instances of wild rice bacterial blight. Simultaneously, the varying viewpoints and distances of the UAV introduce distinct contextual information requirements for accurate detection. However, introducing an excess of contextual information can inadvertently obscure target features’ specifics while exacerbating the model’s complexity.

To address the aforementioned challenges, the presented Xoo-YOLO model incorporates the LSKNet (Li et al., 2023). LSKNet dynamically adjusts its large spatial receptive field, enabling the model to flexibly employ varying large kernels and modify the receptive field according to the specific spatial requirements of each target. This adaptability is especially vital for detecting wild rice bacterial blight amid the UAV viewpoint and intricate conditions of the field.

LSKNet is implemented through a spatial selection mechanism, which is implemented by effectively weighting the features processed by a sequence of large depth-wise convolutional kernels and spatially merging them. The weights of this kernel are dynamically determined based on the input, allowing the model to adaptively use different large kernels and adjust the receptive field for each target in space as needed.

LSKNet mainly consists of two sub-blocks, Large Kernel Selection (LK Selection) and Feed-Forward Network (FFN). FFN is used for channel mixing and feature refinement and consists of a sequence of a fully connected layer, a depth-wise convolution, a GELU activation, and another fully connected layer; LK Selection consists of a fully connected layer, LSK sub-block, a GELU activation, and another fully connected layer.

The central component of LSKNet is the LSK sub-block, which comprises a series of large kernel convolutions alongside a spatial kernel selection mechanism. Larger kernel convolution constructs by explicitly decomposing it into a sequence of depthwise convolutions with a large growing kernel and increasing dilation. To elaborate, the expansion of various parameters, such as kernel size “k,” dilation rate “d,” and the receptive field “RF,” within the ith depthwise convolution within the sequence, is characterized by the following definitions:

The increasing kernel size and dilation rate ensure that the receptive field expands quickly enough, and at the same time, an upper bound on the dilation rate is set to avoid the dilation convolution introducing gaps between feature maps. This approach makes the later kernel selection easier and also reduces the number of parameters significantly. Simultaneously, a series of decomposed depthwise convolutions with different receptive fields are used to obtain features with contextual information at different ranges, allowing channel blending for each spatial feature vector. The calculation is shown in Eqs. (2) and (3).

Based on the above, LSKNet employs a spatial kernel selection mechanism to enhance the network’s ability to focus on the most relevant spatial context regions, spatially selecting feature mappings from large convolutional kernels of different scales.

The outlined implementation steps are as follows:

(1) Concatenate features obtained from different kernels with different ranges of receptive fields:

(2) Efficient spatial relation extraction using channel-based average pooling and maximum pooling:

ally pooled features and use a convolutional layer ℱ 2 → N (·) to transform the pooled features (with two channels) into N spatial attention maps:

(4) A sigmoid activation function is used to obtain a separate spatial selection mask for each decomposed kernel:

(5) The features in the decomposed macronucleus sequence are weighted with the corresponding spatially selective masks and fused by a convolutional layer ℱ (·) to obtain the attention feature S :

(6) The final output of the LSK module is an element-by-element product between input features X and S:

LSKNet fulfills the requirement for detecting bacterial blight in wild rice from the perspective of UAV in the field while addressing the necessity for a wider and adaptable contextual understanding without bells and whistles. Therefore, we add LSKNet to the backbone network to enhance the feature extraction capability of this model. The structural diagram of LSKNet is illustrated in Figure 4 .

The standard convolution (SC) module used in YOLOv8 applies different convolutional kernels to multiple channels simultaneously, which leads to an increase in the number of parameters required and high FLOP. On the other hand, although lightweight networks using depthwise separable convolutions (DSC) such as MobileNet (Howard et al., 2017) and ShuffleNet (Zhang et al., 2018) can effectively solve this problem and greatly improve the detection performance, DSC separates channel information from the input image during computation, leading to a significant reduction in the feature extraction and fusion capabilities and resulting in a decrease in the detection performance of the model. It cannot meet the real-time requirements of disease resistance identification for the detection of wild rice bacterial blight.

To enhance computation speed without compromising detection accuracy, the Xoo-YOLO model integrates the GSConv hybrid convolution module (Li et al., 2022). This module incorporates Shuffle to infuse information produced by SC into the information generated by DSC. In contrast to DSC, the strength of GSConv lies in its ability to maintain hidden connections while operating with reduced complexity. This approach adeptly strikes a model equilibrium between accuracy and speed, ensuring a judicious trade-off between the two.

The GSconv module is primarily composed of Conv, DWConv, Concat, and Shuffle operations. This configuration is illustrated in Figure 5A and is constructed as follows:

VoVGSCSP is an iterative integration of the GS bottleneck utilizing the foundation of GSConv (Xu et al., 2023). The process involves segmenting the input feature map’s channel count into two segments. The initial portion traverses through convolution (Conv) for processing, following which the features undergo extraction through consecutively stacked GS bottleneck modules. On the other hand, the remaining segment is utilized as residuals, engaging in a single convolution operation. This module aptly harmonizes the model’s accuracy and speed, resulting in a reduction of both computational and complexity. Simultaneously, it sustains a commendable degree of accuracy and significantly augments the reutilization rate of extracted features. The structural representation of VoVGSCSP can be observed in the diagram depicted in Figure 5B .

Ultimately, the neck network underwent refinement through the amalgamation of GSConv and VoVGSCSP. These enhancements served to diminish the model’s overall computational overhead, resulting in swifter network operation and reduced information processing time. This harmonious adjustment better balances the trade-off between detection speed and accuracy. The augmented configuration of the neck network is visually presented in the provided Figure 6 .

The UAV was taken from above, and the target in its images usually appears at any angle. The wild rice leaves are striped and lanceolate in shape, and the bacterial blight is also striped on the leaves. This makes the horizontal bounding box of the original YOLOv8 model overlap too much and inevitably introduces too much background information into the horizontal bounding box of adjacent targets, which not only leads to the phenomenon of missed detection and wrong detection at the same time but also increases the difficulty of extracting features from the network (Zhang et al., 2021). Recently, oriented bounding boxes, which include an angular dimension, have been utilized to represent objects with different orientations (Liao et al., 2018). For instance, Zhang et al. (2023) utilized oriented bounding boxes to detect Fusarium head blight (FHB) in wheat and achieved notably high accuracy and robustness in predicting FHB levels. The use of oriented bounding boxes for detecting crop diseases from a UAV perspective has proven to be effective. In light of this, the model proposed in this study adopts oriented bounding box detection. This approach involves angular regression (Ren et al., 2015) and employs the long-edge definition method (Ma et al., 2018) to regress the minimum bounding rectangle of the target. This enhancement is crucial for ensuring reliable and accurate wild rice bacterial blight detection in field conditions.

The horizontal bounding box parameters are expressed as (x, y, w, and h). The four parameters indicate the horizontal and vertical coordinates of the center of the horizontal bounding box, width, and height, respectively. The oriented bounding box contains five parameters (x, y, w, h, and θ), with θ indicating the angle of rotation (Li et al., 2020). The structural diagram of the head is shown in Figure 7 . To prevent duality, the long-edge definition method ( Figure 8 ) is used, i.e., θ is between −90° and 90°, w is the longest edge, its neighbor is h, and θ represents the range of angles through which the x-axis is rotated to w. The model proposed in this paper adds a new rotation angle prediction channel to the head structure to implement the detection of the oriented bounding box, with a dimension of 3 × (5 + 1 + C), where 3 represents that each grid will be predefined with three predicted boxes of various aspect ratios, 5 represents that each predicted box will predict the parameter (x, y, w, h, and θ) of the border, 1 is used to determine whether each grid contains the object, and the final C parameters is used to determine the type of object each grid contains.

The loss function of the model proposed in this paper consists of three parts: Reg part, Obj part, and Cls part. The Reg part is the regression parameter judgment of the feature points, the Obj part is the judgment of whether the feature points contain objects, and the Cls part is the kind of objects contained in the feature points. In order to avoid the differences between the angular parameter θ and other parameters to bring difficulties to the training of the model, the Reg part uses KLD as the loss function (Yang et al., 2021), whose core idea is to convert the oriented bounding box into a two-dimensional Gaussian distribution, i.e. (x, y, w, h, and θ) into a two-dimensional Gaussian distribution of N ( μ , Σ ) , and the conversion is shown in Eq. (10). The KLD between the Gaussian distribution is calculated as the regression, the loss is calculated as shown in Eq. (11), where the subscripts p and t denote the predicted distribution results and the actual results respectively. The KLD loss function is scale-invariant and can dynamically adjust the gradient weights of the angle parameters according to the aspect ratio of the object, and this self-modulation optimization mechanism effectively promotes the accuracy of oriented bounding box detection. The Obj and Cls parts, on the other hand, adopt a binary cross-entropy loss function to reduce the computational complexity of training.