Rice blast disease resistance screening method based on unmanned aerial vehicle and deep learning

Abstract

The present application relates to the technical field of agricultural disease detection and crop disease resistance screening, and in particular to a rice bacterial leaf blight disease resistance screening method based on a UAV and deep learning, which has the technical scheme of: an optimized YOLOv11-OBB model is composed of a C3k2FC module, an SPPF_LSKA module, a SlimNeck module and a LiteHead module, is responsible for extracting features in an image and locating bacterial leaf blight spots, and can efficiently and accurately detect bacterial leaf blight spots; a lightweight deep learning model is used, can efficiently run on a resource-limited UAV platform, and reduces device and computing costs; through cooperation of multiple optimization modules, the detection accuracy of bacterial leaf blight spots is significantly improved, and stable detection effects can be maintained; the operation process is simplified, and full automation of bacterial leaf blight disease resistance screening is realized; therefore, by using the lightweight and optimized YOLOv11-OBB model and combining the module design with high computing efficiency, low-cost, efficient, automated and high-precision bacterial leaf blight disease resistance screening is realized, and the method is suitable for large-scale field application.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural disease detection and crop disease resistance screening, and particularly relates to a rice bacterial leaf blight disease resistance screening method based on a UAV and deep learning. BACKGROUND

[0002] Rice is one of the most important food crops in the world, and bacterial leaf blight is one of the main diseases affecting rice yield and quality. Traditional bacterial leaf blight disease resistance screening methods rely on manual observation and recording, which have problems such as high labor intensity, strong subjectivity, and poor data consistency. In addition, existing bacterial leaf blight disease resistance screening methods, although they have improved screening efficiency, still face limitations such as high equipment costs, high computing resource requirements, and complex operations. Specifically, existing UAV platforms rely on high-precision sensors, resulting in high equipment costs; traditional deep learning models require strong computing resources and cannot efficiently run on low-cost, computing resource-limited UAV platforms; existing technologies usually rely on professional personnel for data analysis and operation, and lack efficient automated screening systems.

[0003] In view of this, we propose a rice bacterial leaf blight disease resistance screening method based on a UAV and deep learning to solve the existing problems. SUMMARY

[0004] The present application aims to provide a rice bacterial leaf blight disease resistance screening method based on a UAV and deep learning to solve the problems raised in the background.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solution: a rice bacterial leaf blight disease resistance screening method based on a UAV and deep learning, the operation steps of which include:

[0006] S1, UAV image acquisition: the UAV flies at a predetermined flight route and height, and takes pictures of bacterial leaf blight lesion images in the rice field;

[0007] S2, bacterial leaf blight lesion detection and positioning: after inputting the image data, the YOLOv11-OBB model is used to detect the bacterial leaf blight lesions; the improved YOLOv11-OBB model is composed of a C3k2FC module, a SPPF_LSKA module, a SlimNeck module, and a LiteHead module, which are responsible for extracting features in the image and positioning the bacterial leaf blight lesions;

[0008] S3, target tracking and multi-target identification: a Kalman filter is used to track the bacterial leaf blight lesions in each frame of image for multi-target tracking, avoiding the situation of image repetition of the same target during flight;

[0009] S4, Disease incidence calculation and disease resistance classification: the disease incidence of rice is calculated according to the number of detected disease spots, and the rice is classified into different disease resistance grades by using the threshold value of the incidence, so as to screen the rice varieties according to the disease resistance;

[0010] S5, screening result output and display: the final screening result is output and presented to the user in the form of a graphical interface or a report, helping the breeding researchers to make disease-resistant variety selection;

[0011] In S2, the overall structure of the improved YOLOv11-OBB model includes backbone, neck and detection head; the improvement contents include: in the backbone network, the C3k2FC module integrates PConv and CGLU into the C3k2 structure; in the backbone network, the SPPF_LSKA module includes the combination of LSKA pooling; in the neck, the SlimNeck module uses GSConv and VoVGSCSP to optimize feature fusion while minimizing FLOP; in the detection head, the LiteHead module adopts separate batch normalization and dynamic anchor frame; in addition, LAMP pruning implements adaptive global pruning to compress model parameters.

[0012] Further, for the C3k2FC module, two 1x1 point convolutions in the FasterNet block are replaced with the CGLU module to form an improved FC structure; at the same time, PConv is integrated into the deep convolution section of the C3k2 bottleneck part, and finally an improved C3k2FC module is formed.

[0013] Further, for the SPPF_LSKA module, the structure first extracts multi-resolution features through multiple maximum pooling branches, and then applies the LSKA module to attention-weighted spliced features; the horizontal and vertical branches of LSKA process spatial information, and the depth separable convolution further optimizes channel mixing; finally, the fused features are output through 1x1 convolution.

[0014] Further, for the SlimNeck module, the GSConv unit introduces a Shuffle mechanism to fuse the channel-related data obtained from SC with the spatial output from DSC.

[0015] Further, for the SlimNeck module, the VoVGSCSP unit divides the input matrix channel into two parts; one part is preprocessed through convolution and extracts features through a series of GS bottleneck units, while the other part is used as a residual path and only performs single convolution transformation.

[0016] Further, the core components of the LiteHead detection head include shared convolution layers, separate BN modules and dynamic anchor frame generators.

[0017] Further, LAMP determines the pruning threshold by calculating the adaptive importance score of each layer.

[0018] Further, the core components of GSConv units include Conv, DWConv, Concat, and Shuffle operations.

[0019] Further, VoVGSCSP units utilize the CSP framework to enhance feature reuse, minimize data loss through a one-time aggregation approach, and integrate batch normalization and SiLU activation functions to maintain feature consistency.

[0020] Further, shared convolutional layers are used for multi-task reuse; separate BN modules apply BN independently for each detection branch; and a dynamic anchor generator calculates anchor parameters in real-time based on feature map statistics.

[0021] Compared with the prior art, the present application has the following beneficial effects:

[0022] The optimized YOLOv11-OBB model provided by the present application can efficiently and accurately detect white leaf blight spots, is suitable for automatic screening in large-scale rice fields, uses a lightweight deep learning model that can efficiently run on a resource-limited unmanned aerial vehicle platform, reduces equipment and computing costs, significantly improves the detection accuracy of white leaf blight spots through the cooperation of multiple optimization modules, especially in complex environments, leaf occlusion, and light changes, etc., can maintain stable detection effect, simplifies the operation process, realizes the full automation of white leaf blight resistance screening, and provides a more efficient solution for rice breeding; therefore, by using the lightweight and optimized YOLOv11-OBB model, combined with the module design with high computing efficiency, low-cost, efficient, automatic, and high-precision white leaf blight resistance screening is realized, which is suitable for large-scale field application. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 Fig. 1 is a structural schematic diagram of the white leaf blight detection model of the present application;

[0024] Figure 2 Fig. 2 is a structural schematic diagram of the C3k2FC module of the present application;

[0025] Figure 3 Fig. 3 is a structural schematic diagram of the SPPF_LSKA module of the present application;

[0026] Figure 4 Fig. 4 is a structural schematic diagram of the SlimNeck module of the present application;

[0027] Figure 5 Fig. 5 is a schematic diagram of the LAMP process of the present application. DETAILED DESCRIPTION

[0028] The technical solutions of the present application are further described below in combination with the drawings and specific embodiments. Example one

[0029] The operation steps of the rice bacterial leaf blight resistance screening method based on the unmanned aerial vehicle and deep learning include:

[0030] S1, unmanned aerial vehicle image acquisition: the unmanned aerial vehicle flies at a predetermined flight route and height, and shoots the images of the bacterial leaf blight spots in the rice field; the flight height is set to 0.6 meters to 1.5 meters, and the image resolution is 1920x1080 pixels; the images captured by the unmanned aerial vehicle are transmitted to the ground end computing device through wireless or wired mode; a lightweight unmanned aerial vehicle such as DJI Mini 2 or DJI Mini 4 Pro is used for image acquisition of bacterial leaf blight infection in the rice field, to ensure full coverage, data consistency and accuracy;

[0031] S2, bacterial leaf blight spot detection and positioning: after inputting the image data, the YOLOv11-OBB model is used for bacterial leaf blight spot detection; the improved YOLOv11-OBB model is composed of C3k2FC module, SPPF_LSKA module, SlimNeck module and LiteHead module, which is responsible for extracting features in the image and positioning the bacterial leaf blight spots; based on the YOLOv11-OBB model, the C3k2FC, SPPF_LSKA, SlimNeck and LiteHead modules are optimized, which significantly improves the detection accuracy and calculation efficiency of the bacterial leaf blight spots;

[0032] S3, target tracking and multi-target identification: the Kalman filter is used for multi-target tracking of the bacterial leaf blight spots in each frame of image, to avoid the situation that the same target appears repeatedly in the image during flight due to occlusion; the position of the disease spot is predicted by the Kalman filter, to reduce the error and repetition in detection;

[0033] S4, disease incidence calculation and disease resistance classification: the disease incidence of the rice is calculated according to the number of detected disease spots; the rice is divided into different disease resistance grades (L1, L2, L3, L4) by using the threshold value of the incidence, so as to screen out rice varieties with strong disease resistance; the disease resistance screening is automatically completed by the algorithm, to avoid human error;

[0034] S5, screening result output and display: the final screening result is output and presented to the user in the form of graphical interface or report, to help the breeding researchers to make disease resistance variety selection; the rice varieties with strong bacterial leaf blight resistance are screened according to the disease resistance grades, for use in breeding research.

[0035] YOLOv11 is an advanced object detection model that comes in multiple versions, including the standard YOLOv11, YOLOv11-OBB, YOLOv11-Seg for segmentation tasks, and YOLOv11-Pose for pose estimation. Additionally, YOLOv11 has multiple size variants: n, s, l, and x, each optimized for different computational efficiency and performance.

[0036] Since the drone captures images from above, the target is usually presented at different angles. In addition, white leaf blight often appears as elongated lesions. Considering these factors, the YOLOv11-OBB model was chosen. To optimize model efficiency and reduce computational complexity, the smallest n model was chosen as the basis for improvement. As shown in Figure 1 The overall structure of the improved model includes three main parts: Backbone, Neck, and Head.

[0037] The improvements include: In the backbone network, the C3k2FC module integrates PConv and CGLU into the C3k2 structure to reduce redundancy and improve efficiency; In the backbone network, the SPPF_LSKA module combines LSKA pooling to enhance multi-scale feature aggregation and reduce noise; In the neck, the SlimNeck module uses GSConv and VoVGSCSP to optimize feature fusion while minimizing FLOP; In the detection head, the LiteHead module uses separate batch normalization (BN) and dynamic anchor boxes to better handle rotation; In addition, LAMP pruning implements adaptive global pruning to compress model parameters while maintaining accuracy.

[0038] The C3k2 module in YOLOv11-OBB is based on the Cross Stage Partial (CSP) structure, which divides the input feature map into two parallel branches. This design enhances feature reuse and gradient flow by repeating bottleneck units, supports flexible layer and channel configurations, and balances model depth and computational efficiency. However, this module has redundant convolution operations, increasing computational cost and making deployment on resource-constrained drone embedded systems complex.

[0039] To address these issues, this application integrates the partial convolution (PConv) of FasterNet block and the TransNext CGLU module into the C3k2 module to enhance the backbone and neck network. This improvement aims to achieve lightweight design, reduce computational redundancy, and minimize parameter size, making it more suitable for deployment on resource-constrained UAV embedded systems. The CGLU module is an evolved version of the GLU Gated Linear Unit, which expands local spatial awareness by inserting a 3x3 depth convolution before the activation function. The module consists of two linear transformation paths: one path applies a 3x3 depth convolution followed by a gated activation function to enhance local feature capture, and the other path generates a baseline output. These two paths are merged by element-wise multiplication, adjusting feature weights, and finally generating the final output through a linear projection and residual connection with the input.

[0040] The FasterNet block uses a reverse residual structure to simplify the network, first using a 1x1 point convolution for channel expansion and feature compression. Then the PConv module is inserted, which applies a 3x3 standard convolution to a quarter of the input channels for spatial feature mixing, while the remaining channels are passed through an identity mapping to reduce full-channel computation. Next, the feature distribution and nonlinearity are maintained through a normalization layer and ReLU activation function, and finally the channels are restored and integrated through another 1x1 point convolution.

[0041] As shown in Figure 2 , this application replaces the two 1x1 point convolutions in the FasterNet block with the CGLU module, forming an improved FC structure. At the same time, the PConv is integrated into the depth convolution section of the C3k2 bottleneck part, finally forming an improved C3k2 FC module.

[0042] This structure optimizes feature extraction through partial channel convolution and gating mechanisms, reducing floating-point operations and memory access. The FC block with the PConv module extracts multi-scale features from the input image while maintaining low computational overhead. Combined with the CGLU module, this design supports dynamic feature weight allocation. The improved C3k2 FC module integrates these components to capture detailed and global information, supporting rice white leaf blight detection and positioning under the UAV perspective, suitable for deployment on edge devices.

[0043] The SPPF module in YOLOv11 is based on the Spatial Pyramid Pooling (SPPF) structure, which 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. This module aims to reduce information loss and enhance adaptability to different scales of targets, supporting efficient feature representation. However, the pooling operations of this module have limitations, leading to low efficiency in feature aggregation in complex scenes and potentially introducing background noise. This is particularly challenging when processing rice field images captured by drones, as the images contain different lighting, leaf occlusions, and resolution fluctuations, which can cause fine-grained details to be overlooked or irrelevant features to be amplified, leading to false detections and performance degradation.

[0044] To address these issues, this application integrates a large separable convolution attention mechanism (LSKA) into the SPPF module, enhancing multi-scale feature extraction and local attention capabilities. The LSKA module is based on separable convolution, which decomposes large kernel convolution into 1D convolution in horizontal and vertical directions to simulate a large receptive field attention mechanism. This approach combines depthwise separable convolution, reducing the number of parameters.

[0045] As shown in Figure 3 The SPPF_LSKA module integrates the LSKA mechanism after the pooling layer of the traditional SPPF. This structure first extracts multi-resolution features through multiple maximum pooling branches, then applies the LSKA module to attention-weighted concatenated features. The horizontal and vertical branches of LSKA handle spatial information, and depthwise separable convolution further optimizes channel mixing. Finally, the fused features are output through a 1x1 convolution to enhance global and local semantic representation.

[0046] This application replaces the original SPPF module in YOLOv11-OBB with the improved SPPF_LSKA module, optimizing the integration of multi-scale features through attention mechanisms to support rice white leaf blight detection and positioning from the perspective of drones.

[0047] The standard convolution (SC) component in YOLOv11-OBB handles multi-channel data through parallel multi-core operations, which, although promotes inter-channel deep interaction, also leads to parameter bloating 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 greatly reduces feature integration and extraction. This limitation is particularly difficult for detecting rice bacterial leaf blight spots from a UAV perspective, as there may be variations in lighting, leaf occlusion, and resolution in the image, which can cause small spots to be missed or inaccurately located.

[0048] To accelerate computation without compromising detection accuracy, the present application integrates the GSConv composite convolution unit into the neck structure of YOLOv11-OBB. The GSConv unit introduces a Shuffle mechanism to fuse channel-related data obtained from SC with spatial outputs from DSC. Compared to using only DSC, GSConv alleviates the computational burden while preserving potential channel correlations, optimizing accuracy and response time, making it very suitable for crop disease detection on edge hardware.

[0049] The core components of the GSConv unit include Conv, DWConv, Concat, and Shuffle operations. Its structure is as follows: the input feature matrix contains C1 channels; apply depthwise separable convolution (DSC) to half of the channels to capture spatial details, while apply standard convolution (SC) to the other half of the channels for channel fusion; the outputs on both sides are concatenated along the channel axis; apply Shuffle operation to the concatenated result to promote random reorganization of channels to enhance data interaction; the final output matrix contains C2 channels.

[0050] VoVGSCSP is an iterative bottleneck combination of GSConv, which divides the input matrix channels into two parts. One part is pre-processed through convolution and extracts features through a series of GS bottleneck units, while the other part serves as a residual path and only undergoes a single convolution transformation. This unit uses the CSP framework to enhance feature reuse, minimizes data loss through a one-time aggregation method, and integrates batch normalization and SiLU activation functions to maintain feature consistency.

[0051] As Figure 4As shown, the present application uses GSConv and VoVGSCSP to reconstruct the neck framework of YOLOv11-OBB, forming a SlimNeck structure. These modifications reduce the computational burden of the model, improving inference speed and data processing capacity, while enhancing the integration of multi-resolution features. This optimization scheme achieves an effective balance between speed and accuracy.

[0052] The detection head of YOLOv11-OBB uses a shared convolution mechanism to process multi-scale feature outputs for bounding box prediction. Although this approach promotes parameter reuse, using uniform batch normalization (BN) can lead to inaccurate moving averages when there are significant statistical differences between feature layers, affecting training stability and generalization performance. In addition, the static anchor box design is difficult to adapt to the challenges of leaf occlusion, resolution fluctuations, and changes in lighting conditions when dealing with the rotation and scale changes of white leaf blight spots, especially when captured by drones, often leading to positioning errors or missing small targets, thus limiting overall detection performance.

[0053] To address these issues, the present application proposes a LiteHead detection head. This optimization design aims to improve computational efficiency, reduce parameter size and operational burden, making the system more suitable for resource-constrained drone embedded systems. LiteHead avoids BN-related issues by separately handling statistical differences between different levels of features, ensuring independent sliding averages, thereby preventing the accumulation of bias in shared parameters. In addition, a dynamic anchor box generator dynamically adjusts the scale, angle, and aspect ratio of anchor boxes based on input features, enhancing the alignment capability for rotated targets.

[0054] The core components of the LiteHead detection head include: shared convolution layers for multi-task reuse, reducing redundant parameters; separate BN modules, applying BN independently to each detection branch to optimize feature distribution and accelerate convergence; and a dynamic anchor box generator that calculates anchor box parameters in real time based on feature map statistics, supporting the prediction of rotated bounding boxes (OBB). The workflow is as follows: multi-scale feature maps are processed through shared convolution to extract bounding box regression and classification scores; separate BN standardizes the output of each layer; the dynamic anchor box module generates adaptive anchor boxes in real time based on feature clustering or gradient guidance, and finally outputs OBB predictions.

[0055] In the present application, LiteHead is integrated into the detection head of YOLOv11-OBB, replacing standard shared convolution and static anchor boxes, optimizing the feature processing pipeline, reducing floating point operations and memory access, and enhancing the ability to capture rotated details and global information. The LiteHead architecture is very suitable for white leaf blight spot detection and positioning in rice field images captured by drones.

[0056] Despite the introduction of lightweight modules, the YOLOv11-OBB model still faces problems of parameter redundancy and computational intensity, which leads to an increase in floating-point operations (GFLOPs) and memory requirements. This problem is particularly serious for resource-constrained UAV embedded systems, which can cause inference delays and deployment bottlenecks.

[0057] To address these challenges, the present application introduces a layer-adaptive magnitude pruning (LAMP) strategy to further optimize the YOLOv11-OBB model. LAMP reduces the parameter size and computational overhead of the model through global pruning, making it more suitable for resource-constrained UAV platforms. LAMP is a magnitude-based pruning strategy that determines the pruning threshold by calculating an adaptive importance score (LAMP score) for each layer. It avoids performance degradation due to uniform sparsity and dynamically adjusts the sparsity of each layer based on the statistical characteristics of its weight distribution. Earlier layers retain more connections to capture fine-grained features, while later layers undergo more aggressive pruning to reduce computational complexity.

[0058] The LAMP pruning process includes the following steps: pre-training the full model to obtain initial weights; calculating global LAMP scores, sorting all weights, and setting thresholds based on the target sparsity; applying magnitude-based pruning at the layer level to remove connections below the threshold; fine-tuning the pruned model to restore accuracy, often using techniques such as knowledge distillation or learning rate scheduling.

[0059] Compared to traditional magnitude pruning methods, LAMP maintains better accuracy after pruning by using a layer-adaptive mechanism. This allows for a significant reduction in parameters and GFLOPs without sacrificing detection performance. Figure 5 The process of LAMP is demonstrated.

[0060] In this application, LAMP is applied to the backbone, neck, and detection head of YOLOv11-OBB. After iterative pruning and fine-tuning, a lightweight variant is obtained. This optimized structure enhances resource utilization, supports capturing both detailed and global information from the UAV's perspective, and is suitable for rice white leaf blight detection and positioning. Embodiment Two

[0061] Since the UAV will continue to move while shooting video, some frames may not fully capture the rice plants, and the same white leaf blight target may appear in multiple frames. Relying solely on single-image disease detection can lead to missed detection, false detection, or repeated detection, affecting the accuracy and stability of the screening process. To address this problem, the present application proposes a screening method. This method includes image acquisition and transmission, disease target detection, Kalman filter prediction, detection and prediction matching, disease counting, disease grade classification, and germplasm screening.

[0062] In image acquisition and transmission, the user controls the flight of the UAV, adjusts the focal length of the camera and the angle of the gimbal during flight, and then activates the video recording function. The captured video data is transmitted to the edge computing device through wireless or wired connection.

[0063] In disease target detection, the video recorded by the UAV is processed by the white leaf blight detection model proposed in this application, and the corresponding bounding box is detected and generated.

[0064] In Kalman filter prediction, a multi-target tracking model is introduced to predict the position of the target using a Kalman filter. The Kalman filter is a recursive estimation algorithm that uses statistical inference of historical data to estimate the current position of the target. This prediction provides prior information for disease detection in subsequent frames, thereby reducing the bias of occlusion or detection errors.

[0065] The state update equation of the Kalman filter is where, represents the predicted state at time t, F is the state transition matrix, B is the control input matrix, and u t is the control vector. This prediction helps improve the accuracy of subsequent frame detection.

[0066] In detection and prediction matching, the Hungarian algorithm is used to match the detection results of the current frame with the predicted position of the Kalman filter. The calculation formula of the matching score is where, represents the intersection over union of the bounding box, represents the similarity of the re-identified features, and λ is the weight parameter that balances spatial overlap and appearance features.

[0067] In disease counting, the white leaf blight targets in the tracking list are tracked, and each target has a unique identifier for counting the number of white leaf blight occurrences.

[0068] In disease grade classification, according to the disease count and the number of inoculations of rice, the disease incidence is calculated using the formula, and then the rice is classified into white leaf blight resistance grades according to the incidence, as shown in Table 1.

[0069] Table 1 Classification criteria for white leaf blight-resistant rice germplasm

[0070]

[0071] In germplasm screening, rice with L1 or L2 grade is selected as the preferred candidate material for white leaf blight resistance.

[0072] The above specific embodiments are only several preferred embodiments of the present application, and based on the technical solutions of the present application and the related inspirations of the above embodiments, the person skilled in the art can make various alternative improvements and combinations on the above specific embodiments.

Claims

1.A method for screening rice bacterial blight resistance based on unmanned aerial vehicle and deep learning, characterized in that, The operation steps comprise: S1, unmanned aerial vehicle image acquisition: the unmanned aerial vehicle flies at a predetermined flight route and height, and shoots a white leaf blast spot image in a rice field; S2, white leaf blast spot detection and positioning: after inputting image data, a YOLOv11-OBB model is used for white leaf blast spot detection; the improved YOLOv11-OBB model is composed of a C3k2FC module, an SPPF_LSKA module, a SlimNeck module and a LiteHead module, and is responsible for extracting features in the image and positioning the white leaf blast spot; S3, target tracking and multi-target identification: a Kalman filter is used for multi-target tracking of the white leaf blast spot in each frame of image, so as to avoid the situation that the same target repeatedly appears in the image during flight due to occlusion; S4, disease incidence calculation and disease resistance classification: the disease incidence of rice is calculated according to the number of detected disease spots, and the rice is classified into different disease resistance grades by using the threshold of the incidence, so as to select the rice varieties according to the disease resistance; S5, screening result output and display: the final screening result is output and presented to the user in the form of a graphical interface or a report, helping the breeding researchers to make disease-resistant variety selection; In S2, the overall structure of the improved YOLOv11-OBB model includes a backbone, a neck and a detection head; the improvement contents include: in the backbone network, the C3k2FC module integrates PConv and CGLU into the C3k2 structure; in the backbone network, the SPPF_LSKA module includes LSKA pooling; in the neck, the SlimNeck module uses GSConv and VoVGSCSP to optimize feature fusion while minimizing FLOP; in the detection head, the LiteHead module uses separate batch normalization and dynamic anchor boxes; in addition, LAMP pruning is implemented to adaptively prune the model parameters. 2.The unmanned aerial vehicle and deep learning-based screening method for rice bacterial leaf blight resistance according to claim 1, wherein: For the C3k2FC module, two 1x1 point convolutions in the FasterNet block are replaced with a CGLU module to form an improved FC structure; at the same time, PConv is integrated into the deep convolution section of the C3k2 bottleneck part, finally forming an improved C3k2FC module. 3.The unmanned aerial vehicle and deep learning-based screening method for rice bacterial leaf blight resistance according to claim 1, characterized in that: For the SPPF_LSKA module, the structure first extracts multi-resolution features through multiple maximum pooling branches, and then applies the LSKA module to attention-weighted spliced features; the horizontal and vertical branches of LSKA process spatial information, and the depth separable convolution further optimizes channel mixing; finally, the fused features are output through a 1x1 convolution. 4.The unmanned aerial vehicle and deep learning-based screening method for rice bacterial leaf blight resistance according to claim 1, wherein: For the SlimNeck module, the GSConv unit introduces a Shuffle mechanism to fuse the channel-related data obtained from SC with the spatial output from DSC. 5.The unmanned aerial vehicle and deep learning-based screening method for rice bacterial leaf blight resistance according to claim 1, characterized in that: For the SlimNeck module, the VoVGSCSP unit divides the input matrix channel into two parts; one part is preprocessed through convolution and extracts features through a series of GS bottleneck units, while the other part is a residual path and only performs a single convolution transformation. 6.The unmanned aerial vehicle and deep learning-based screening method for rice bacterial leaf blight resistance according to claim 1, wherein: The core components of LiteHead detection head include shared convolutional layers, separated BN modules, and dynamic anchor generator. 7.The unmanned aerial vehicle and deep learning-based screening method for rice bacterial leaf blight resistance according to claim 1, characterized in that: LAMP determines the pruning threshold by calculating the adaptive importance score of each layer. 8.The unmanned aerial vehicle and deep learning-based screening method for rice bacterial leaf blight resistance according to claim 4, characterized in that: The core components of GSConv unit include Conv, DWConv, Concat, and Shuffle operations. 9.The unmanned aerial vehicle and deep learning-based screening method for rice bacterial leaf blight resistance according to claim 5, characterized in that: VoVGSCSP unit utilizes the CSP framework to enhance feature reuse, minimizes data loss through a one-time aggregation method, and integrates batch normalization and SiLU activation functions to maintain feature consistency. 10.The unmanned aerial vehicle and deep learning-based screening method for rice bacterial leaf blight resistance according to claim 6, characterized in that: Shared convolutional layers are used for multi-task reuse; separated BN modules apply BN independently for each detection branch; and the dynamic anchor generator calculates anchor parameters in real time based on feature map statistics.

Images