A farmland pest and disease unmanned aerial vehicle inspection method based on remote sensing image recognition
Patent Information
- Application Number
- CN202611091430.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-22
- Publication Date
- 2026-10-30
AI Technical Summary
[0003]然而，现有基于遥感图像的病虫害识别方法仍存在以下主要问题：高分辨率遥感图像尺寸巨大（通常达到数亿像素），直接输入深度学习模型进行推理受限于GPU显存与计算资源，难以在实际巡检场景中高效部署；同时，现有方法在图像切分推理后，往往缺乏有效的坐标映射机制，无法准确地将局部检测结果还原至全局图像坐标系，更难以进一步转换为真实地理坐标以生成病虫害空间分布图
[0037]1. This invention introduces a multi-threaded asynchronous segmentation mechanism to efficiently divide remote sensing images into multiple segmented sub-maps, and combines them with a pre-trained lightweight pest and disease digital identification model for independent inference. This greatly reduces the peak usage of GPU memory and computing resources for a single forward inference, enabling complex remote sensing image pest and disease identification tasks to be efficiently deployed and run smoothly on resource-constrained actual inspection equipment.
Smart Images
Figure CN123016249A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and specifically to a method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition. Background Technology
[0002] In recent years, with the rapid development of drone technology and artificial intelligence, using agricultural drones equipped with multispectral or high-definition cameras for farmland pest and disease inspection has become an important development direction for smart agriculture. Drone inspection can quickly acquire high-resolution remote sensing images of large areas of farmland, and automatically identify the location of disease spots or pests in the images through deep learning models, thereby guiding subsequent precision pesticide application operations, which is of great significance for reducing pesticide use and increasing crop yields.
[0003] However, existing methods for identifying pests and diseases based on remote sensing images still have the following main problems: high-resolution remote sensing images are huge (usually reaching hundreds of millions of pixels), and directly inputting them into deep learning models for inference is limited by GPU memory and computing resources, making it difficult to deploy efficiently in actual inspection scenarios; at the same time, after image segmentation and inference, existing methods often lack an effective coordinate mapping mechanism, making it impossible to accurately restore local detection results to the global image coordinate system, and even more difficult to further convert them into real geographic coordinates to generate a spatial distribution map of pests and diseases. Summary of the Invention
[0004] To address the aforementioned shortcomings of existing technologies, this invention provides a method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition.
[0005] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for inspecting farmland pests and diseases using unmanned aerial vehicles (UAVs) based on remote sensing image recognition, including: Step 1: UAV flight path planning, multimodal high-resolution remote sensing image acquisition, and UAV flight attitude acquisition;
[0006] Step 2: Using a multi-threaded asynchronous segmentation mechanism, the acquired high-resolution remote sensing image is divided into multiple segmentation sub-maps, and the absolute pixel anchor point coordinates of each segmentation sub-map in the multimodal high-resolution remote sensing image are calculated.
[0007] Step 3: Input the segmented graph into the pre-trained lightweight pest and disease digital recognition model for inference to obtain a structured prediction result set. The structured prediction result set includes pest and disease categories, prediction confidence scores for each pest and disease category, and local bounding box coordinates in the segmented graph coordinate system.
[0008] Step 4: Using the global coordinate aggregation module, perform matrix aggregation operations on the structured prediction result set and the absolute pixel anchor coordinates of each segmented sub-map in the multimodal high-resolution remote sensing image to reconstruct and identify the global bounding box coordinates of pests and diseases in the multimodal high-resolution remote sensing image. Then, convert the global bounding box pixel coordinates into real geographic coordinates in the global coordinate system, and generate and output the spatial distribution data of pests and diseases in combination with the corresponding pest and disease categories.
[0009] Furthermore, step two includes:
[0010] The dynamic calculation of the segmentation step size of the sliding window in the horizontal and vertical directions is based on the preset sliding window overlap rate and sliding window size; where the sliding window size includes the window length and window height;
[0011] The acquired multimodal high-resolution remote sensing image is divided into multiple segmented sub-images according to the preset sliding window overlap rate and the calculated step size.
[0012] Furthermore, step three includes:
[0013] Normalization is performed on the segmented graph, and the normalized segmented graphs are stacked into batch data according to the cache size of the UAV edge computing platform to obtain a standardized floating-point three-dimensional data tensor.
[0014] Geometric edges and semantic features are extracted from standardized floating-point 3D data tensors to obtain baseline feature maps at multiple different scales;
[0015] Adaptive reshaping and cross-scale fusion of recalibrated baseline feature maps at multiple different scales are performed to obtain a multi-level fused feature map pyramid.
[0016] The multi-level fused feature map pyramid is split by scale, and each split is independently decoded into specific disease categories and spatial coordinates to obtain accurate prediction tensors at different scales.
[0017] Spatial coordinate system mapping and non-maximum suppression processing are performed on the accurate prediction tensors at different scales to obtain a structured prediction result set. The structured prediction result set includes pest and disease categories, prediction confidence scores corresponding to each pest and disease category, and local bounding box coordinates in the tangent molecular graph coordinate system.
[0018] Furthermore, step four includes:
[0019] Extract the pest category, prediction confidence level, and local bounding box coordinates in the segmentation graph coordinate system for each identified pest from the structured prediction results. Then, perform matrix addition aggregation operation on the local bounding box coordinates in the segmentation graph coordinate system and the absolute pixel anchor coordinates of the corresponding segmentation graph in the multimodal high-resolution remote sensing image to obtain the global bounding box coordinates of the identified pest in the global pixel coordinate system of the multimodal high-resolution remote sensing image.
[0020] In the global pixel coordinate system, the global bounding boxes of the identified pests and diseases are sorted in descending order according to the prediction confidence of the pest and disease category corresponding to the pest and disease; and the cross-union ratio between overlapping global bounding boxes is calculated. Redundant global bounding boxes with cross-union ratio greater than a preset suppression threshold are removed to obtain a deduplicated set of global bounding boxes.
[0021] Based on the flight attitude of the UAV, a spatial affine transformation matrix is constructed. Using the spatial affine transformation matrix, the coordinate system projection transformation is performed on the center point pixel coordinates of each global bounding box in the deduplicated global bounding box set to obtain the real geographic coordinates. The real geographic coordinates are then associated and encapsulated with the pest category to generate and output structured pest spatial distribution data.
[0022] Furthermore, the process of dividing the acquired high-resolution remote sensing image into multiple segmented sub-images according to a preset sliding window overlap rate and a calculated step size includes:
[0023] Construct horizontal and vertical index variables within a sliding window to perform sliding traversal in the pixel coordinate system of the high-resolution remote sensing image;
[0024] The absolute pixel anchor point coordinates of each segmented sub-map in the multimodal high-resolution remote sensing image are calculated using a boundary adaptive truncation mechanism.
[0025] Using the calculated coordinates of each absolute pixel anchor point as the starting point for slicing, and with the window length and window height as the fixed slicing range, a local slicing operation is performed on the high-resolution remote sensing image to obtain a sliced sub-map.
[0026] Furthermore, the lightweight model based on multi-level self-distillation includes: a data processing module, a backbone network, a neck network, a decoupled detection head, and an output terminal.
[0027] Furthermore, the extraction of geometric edges and semantic features from standardized floating-point 3D data tensors through a backbone network to obtain multiple benchmark feature maps at different scales includes:
[0028] Channel upscaling, spatial sliding extraction, and linear dimensionality reduction compression are performed on standardized floating-point 3D data tensors to obtain geometric features with basic contours and textures of multimodal high-resolution remote sensing images. The geometric features are then abstracted step by step to obtain primary feature maps with multi-level information.
[0029] The bidirectional coordinate attention module performs one-dimensional global average pooling on the primary feature map along the vertical Y-axis and the horizontal X-axis respectively. The pooled directional features are then concatenated to reduce the dimensionality and generate a spatial attention weight matrix. The spatial attention weight matrix is then applied to the primary feature map for feature fusion by multiplying element by element to obtain a recalibrated feature map.
[0030] The recalibration feature maps of multiple pre-defined specific downsampling depth nodes are extracted and organized into independent and parallel multi-scale feature sets. These sets are then distributed to subsequent networks through multiple independent parallel data lines to serve as recalibration reference feature maps at multiple different scales.
[0031] Furthermore, the adaptive reshaping and cross-scale fusion of multiple recalibrated reference feature maps at different scales to obtain a multi-level fused feature map pyramid includes:
[0032] Parallel spatial two-dimensional offset calculation is performed on multiple recalibrated reference feature maps of different scales using a deformable convolution module. Based on the calculated spatial two-dimensional offset, the recalibrated reference feature maps of different scales are processed to obtain deformable feature maps of different scales.
[0033] Convolution is performed on deformable feature maps at different scales to generate intrinsic feature maps at different scales. Then, linear transformations are applied to intrinsic feature maps at different scales to generate "phantom" copies at different scales. Finally, intrinsic feature maps at different scales are concatenated with "phantom" copies at different scales to obtain lightweight feature maps at different scales with half the number of parameters and computational cost without losing feature richness.
[0034] A bidirectional feature transfer path is constructed, which is both top-down and bottom-up. Learnable weights are introduced to perform cross-scale weighted concatenation of lightweight feature maps at different scales, so as to obtain a multi-level fused feature map pyramid that integrates deep strong semantics and shallow high-resolution details.
[0035] Secondly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for inspecting farmland pests and diseases using unmanned aerial vehicles based on remote sensing image recognition.
[0036] Compared with the prior art, the beneficial effects of the present invention are:
[0037] 1. This invention introduces a multi-threaded asynchronous segmentation mechanism to efficiently divide remote sensing images into multiple segmented sub-maps, and combines them with a pre-trained lightweight pest and disease digital identification model for independent inference. This greatly reduces the peak usage of GPU memory and computing resources for a single forward inference, enabling complex remote sensing image pest and disease identification tasks to be efficiently deployed and run smoothly on resource-constrained actual inspection equipment.
[0038] 2. This invention directly performs matrix aggregation operations on the local bounding box coordinates of the structured prediction result set and the anchor point coordinates after inference, thereby quickly reconstructing the global bounding box coordinates of pests and diseases in the original high-resolution remote sensing image. This scheme utilizes pure numerical tensor matrix operations to completely replace the time-consuming and memory-intensive pixel-level image stitching operation in traditional schemes, achieving efficient and lossless restoration of local detection results to the global image coordinate system with zero image stitching burden.
[0039] 3. Based on the reconstructed global bounding box coordinates, this invention further transforms the physically meaningless two-dimensional pixel coordinates into real geographic coordinates within the global coordinate system, and structurally binds them with the corresponding pest and disease categories and prediction confidence levels. The resulting spatial distribution data of pests and diseases not only clarifies "what kind of disease it is," but also precisely pinpoints "which latitude and longitude on Earth," providing direct and reliable data support for subsequent precise targeted variable spraying operations by agricultural drones, truly realizing a closed loop from "visual perception" to "precision agricultural operations." Attached Figure Description
[0040] Figure 1 This is a flowchart of a method for inspecting farmland pests and diseases using unmanned aerial vehicles based on remote sensing image recognition, according to an embodiment of the present invention. Detailed Implementation
[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0042] Example 1:
[0043] like Figure 1 As shown, this embodiment provides a method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition, including:
[0044] S1: UAV flight path planning, multimodal high-resolution remote sensing image acquisition, and UAV flight attitude acquisition;
[0045] Specifically, the UAV is equipped with an edge computing platform and a multimodal camera. According to a pre-set flight path and aerial photography overlap rate, it acquires high-resolution remote sensing images of farmland areas, including visible light and near-infrared bands. At the same time, it acquires the UAV's flight attitude when acquiring multimodal high-resolution remote sensing images. The flight attitude includes at least the heading angle θ, ground sampling distance GSD, flight altitude h, and geographic coordinates of the acquisition center point.
[0046] Step S2: In the UAV edge computing platform, using a multi-threaded asynchronous segmentation mechanism, the acquired high-resolution remote sensing image is divided into multiple segmentation sub-maps, and the absolute pixel anchor point coordinates of each segmentation sub-map in the multimodal high-resolution remote sensing image are calculated.
[0047] Specifically, the following steps are included:
[0048] S21: Calculate the dynamic segmentation step size of the sliding window in the horizontal and vertical directions based on the preset sliding window overlap rate and sliding window size; wherein, the sliding window size includes the window length and the window height;
[0049] ;
[0050] ;
[0051] in, The horizontal step size; The length of the window; The overlap ratio of the sliding window and ; The vertical step size; The height of the window;
[0052] S22: Divide the acquired high-resolution remote sensing image into multiple segmented sub-images according to the preset sliding window overlap rate and the calculated step size;
[0053] Specifically, the following steps are included:
[0054] S22-1: Construct horizontal and vertical index variables within the sliding window to perform sliding traversal in the pixel coordinate system of the high-resolution remote sensing image;
[0055] The initial values of the horizontal and vertical index variables are 0, and the increment is 1.
[0056] S22-2: Calculate the absolute pixel anchor point coordinates of each segmented sub-map in the multimodal high-resolution remote sensing image through a boundary adaptive truncation mechanism;
[0057] ;
[0058] ;
[0059] in, Used as a horizontal index variable; This is a vertical index variable;
[0060] Solving for pixel anchor coordinates using a minimum value function helps ensure that the size of the segmented subplots located at the far right and bottom of the image remains strictly equal to the preset sliding window size.
[0061] S22-3: Calculated coordinates of each absolute pixel anchor point To extract the starting point, based on the window length and window height To fix the cropping range, a local tiling operation is performed on the high-resolution remote sensing image, and the cropping size is [size missing]. Independent image data blocks are used as segmented sub-maps.
[0062] S22-4: Match each segmented sub-image with its corresponding absolute pixel anchor coordinates. The data is bound together, encapsulated into a segmented data structure, and pushed into a shared queue using a multi-threaded asynchronous mechanism to complete the global overlap segmentation of the high-resolution remote sensing image.
[0063] Step S3: Input the segmented graph into the pre-trained lightweight pest and disease digital recognition model for inference, and output the pest and disease category and the local bounding box coordinates in the segmented graph coordinate system;
[0064] The lightweight model based on multi-level self-distillation includes: a data processing module, a backbone network, a neck network, a decoupled detection head, and an output terminal.
[0065] Specifically, the following steps are included:
[0066] S3-1: The data processing module performs a normalization operation on the segmented sub-maps and stacks the normalized segmented sub-maps into batch data according to the cache size of the UAV edge computing platform. This results in a model-readable normalized floating-point 3D data tensor that eliminates exposure extreme value interference and has a consistent underlying data distribution. ;
[0067] S3-2: From standardized floating-point 3D data tensors via the backbone network Geometric edges and semantic features are extracted to obtain baseline feature maps at multiple different scales.
[0068] The backbone network includes an inverted residual module and a bidirectional coordinate attention module;
[0069] Specifically, the following steps are included:
[0070] S3-2-1: Processing normalized floating-point 3D data tensors using the Inverted Residual Block Channel upscaling, spatial sliding extraction, and linear dimensionality reduction compression are performed to obtain the geometric features with basic contours and textures of multimodal high-resolution remote sensing images. The geometric features are then abstracted step by step to obtain primary feature maps with multi-level information.
[0071] For example, the residual module handles normalized floating-point 3D data tensors. Channel upscaling is performed using 1×1 convolution, spatial sliding extraction is performed using 3×3 depthwise separable convolution, and linear dimensionality reduction and compression are performed using 1×1 convolution.
[0072] S3-2-2: The bidirectional coordinate attention module performs one-dimensional global average pooling on the primary feature map along the vertical Y-axis and the horizontal X-axis respectively, and concatenates the pooled directional features to reduce the dimensionality to generate a spatial attention weight matrix. Then, the spatial attention weight matrix is applied to the primary feature map by element-wise multiplication to perform feature fusion and obtain a recalibrated feature map.
[0073] S3-2-3: Extract recalibration feature maps of multiple pre-defined specific downsampling depth nodes, organize the extracted recalibration feature maps into independent and parallel multi-scale feature parallel sets, and then distribute them to the subsequent network through multiple independent parallel data lines to serve as recalibration reference feature maps of multiple different scales.
[0074] S3-3: Adaptive reshaping and cross-scale fusion of recalibrated baseline feature maps at multiple different scales are performed through the neck network to obtain a multi-level fused feature map pyramid.
[0075] This multi-level fusion feature map pyramid includes sufficient semantics and texture to obtain from both extremely small insect infestations and extremely large lesions.
[0076] The neck network includes: a deformable convolutional module, a Ghost convolutional module, and a BiFPN bidirectional fusion module;
[0077] Specifically, the following steps are included:
[0078] S3-3-1: Parallel spatial two-dimensional offset calculation is performed on multiple recalibrated reference feature maps of different scales through deformable convolution modules. Based on the calculated spatial two-dimensional offset, multiple recalibrated reference feature maps of different scales are processed to obtain deformable feature maps of different scales.
[0079] The deformation feature map perfectly matches the overlapping and irregular edges of crops.
[0080] S3-3-2: Convolve the deformable feature maps at different scales using the Ghost convolution module to generate intrinsic feature maps at different scales. Then, apply linear transformations to the intrinsic feature maps at different scales to generate "phantom" copies at different scales. Finally, concatenate the intrinsic feature maps at different scales with the "phantom" copies at different scales to obtain lightweight feature maps at different scales with half the number of parameters and computational cost without losing feature richness.
[0081] S3-3-3: A bidirectional feature transfer path is constructed from top to bottom and bottom to top through the BiFPN bidirectional fusion module, and learnable weights are introduced to perform cross-scale weighted concatenation of lightweight feature maps of different scales to obtain a multi-level fused feature map pyramid that integrates deep strong semantics and shallow high-resolution details.
[0082] It should be noted that the multi-level fusion feature map pyramid described in this invention... In the image, the subscript number represents the downsampling index of the feature map relative to the original input image. Among these, correspond Deep low-resolution feature maps with a downsampling rate of 100 times; and correspond Shallow high-resolution feature maps with a downsampling rate of 10 times can preserve the original pixel spatial details of tiny targets to the greatest extent due to their extremely low downsampling rate, thereby providing physical-level data support for downstream detection branches to prevent missed detections.
[0083] S3-4: By decoupling the detection head, the multi-level fused feature map pyramid is split by scale, and each split is independently decoded into specific disease categories and spatial coordinates to obtain accurate prediction tensors at different scales.
[0084] The decoupled detection head includes a small target-specific detection branch, a regular-scale target detection branch, and a classification and regression sub-branch.
[0085] Specifically, the following steps are included:
[0086] S3-4-1: Independent reasoning is performed on feature maps with extremely low sampling rates through a dedicated detection branch for tiny targets, so as to obtain dedicated branch features that can identify tiny pests and diseases and eliminate blind spots in the detection of extremely small targets.
[0087] S3-4-2: The feature maps except for the feature map with the lowest sampling rate are fused by the conventional scale target detection branch to obtain conventional branch features for identifying medium and large scale diseases.
[0088] S3-4-3: Calculate the class probability and coordinate offset of the exclusive branch features and the regular branch features through the classification and regression sub-branches respectively to obtain the classification probability matrix and the regression offset tensor. Then, spatially align and stitch the classification probability matrix and the regression offset tensor in the channel dimension to merge them into accurate prediction tensors at different scales.
[0089] S3-5: Perform spatial coordinate system mapping and non-maximum suppression (NMS) processing on the accurate prediction tensors at different scales to obtain a structured prediction result set. The structured prediction result set includes the pest and disease categories, the prediction confidence corresponding to each pest and disease category, and the local bounding box coordinates in the tangent graph coordinate system.
[0090] Step S4: The structured prediction result set and the absolute pixel anchor point coordinates of each segmented sub-map in the multimodal high-resolution remote sensing image are matrix aggregated by the global coordinate aggregation module to reconstruct and identify the global bounding box coordinates of the pests in the original high-resolution remote sensing image. The global bounding box pixel coordinates are converted into real geographic coordinates in the global coordinate system, and the spatial distribution data of pests are generated and output in combination with the corresponding pest categories.
[0091] Specifically, the following steps are included:
[0092] S4-1: Local to global pixel coordinate aggregation (coordinate system translation):
[0093] Extract the pest category, prediction confidence level, and local bounding box coordinates in the segmentation graph coordinate system for each identified pest from the structured prediction results. Then, perform matrix addition aggregation operation on the local bounding box coordinates in the segmentation graph coordinate system and the absolute pixel anchor coordinates of the corresponding segmentation graph in the multimodal high-resolution remote sensing image to obtain the global bounding box coordinates of the identified pest in the global pixel coordinate system of the multimodal high-resolution remote sensing image.
[0094] The matrix addition aggregation operation is shown below:
[0095] ;
[0096] ;
[0097] in, and These are the coordinates of the global bounding box. and The coordinates of the center point of the local bounding box of the kth pest are given.
[0098] S4-2: In the global pixel coordinate system, the global bounding boxes of the identified pests and diseases are sorted in descending order according to the prediction confidence of the pest and disease category corresponding to the pest and disease; and the intersection-union ratio (IoU) between overlapping global bounding boxes is calculated. Redundant global bounding boxes with an IoU greater than a preset suppression threshold are removed to obtain a deduplicated set of global bounding boxes.
[0099] S4-3: Construct a spatial affine transformation matrix based on the UAV's flight attitude; use the spatial affine transformation matrix to determine the pixel coordinates of the center point of each global bounding box in the deduplicated global bounding box set. Perform coordinate system projection transformation to obtain the true geographic coordinates. The system then associates and encapsulates the real geographic coordinates with the pest category to generate and output structured spatial distribution data of pests and diseases.
[0100] Example 2:
[0101] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the farmland pest and disease drone inspection method based on remote sensing image recognition described in Embodiment 2.
[0102] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition, characterized in that, include: Step 1: Drone flight path planning, multimodal high-resolution remote sensing image acquisition, and drone flight attitude acquisition; Step 2: Using a multi-threaded asynchronous segmentation mechanism, the acquired high-resolution remote sensing image is divided into multiple segmentation sub-maps, and the absolute pixel anchor point coordinates of each segmentation sub-map in the multimodal high-resolution remote sensing image are calculated. Step 3: Input the segmented graph into the pre-trained lightweight pest and disease digital recognition model for inference to obtain a structured prediction result set. The structured prediction result set includes pest and disease categories, prediction confidence scores for each pest and disease category, and local bounding box coordinates in the segmented graph coordinate system. Step 4: Using the global coordinate aggregation module, perform matrix aggregation operations on the structured prediction result set and the absolute pixel anchor coordinates of each segmented sub-map in the multimodal high-resolution remote sensing image to reconstruct and identify the global bounding box coordinates of pests and diseases in the multimodal high-resolution remote sensing image. Then, convert the global bounding box pixel coordinates into real geographic coordinates in the global coordinate system, and generate and output the spatial distribution data of pests and diseases in combination with the corresponding pest and disease categories.
2. The method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition according to claim 1, characterized in that, Step two includes: The dynamic calculation of the segmentation step size of the sliding window in the horizontal and vertical directions is based on the preset sliding window overlap rate and sliding window size; where the sliding window size includes the window length and window height; The acquired multimodal high-resolution remote sensing image is divided into multiple segmented sub-images according to the preset sliding window overlap rate and the calculated step size.
3. The method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition according to claim 1, characterized in that, Step three includes: Normalization is performed on the segmented graph, and the normalized segmented graphs are stacked into batch data according to the cache size of the UAV edge computing platform to obtain a standardized floating-point three-dimensional data tensor. Geometric edges and semantic features are extracted from standardized floating-point 3D data tensors to obtain baseline feature maps at multiple different scales; Adaptive reshaping and cross-scale fusion of recalibrated baseline feature maps at multiple different scales are performed to obtain a multi-level fused feature map pyramid. The multi-level fused feature map pyramid is split by scale, and each split is independently decoded into specific disease categories and spatial coordinates to obtain accurate prediction tensors at different scales. Spatial coordinate system mapping and non-maximum suppression processing are performed on the accurate prediction tensors at different scales to obtain a structured prediction result set. The structured prediction result set includes pest and disease categories, prediction confidence scores corresponding to each pest and disease category, and local bounding box coordinates in the tangent molecular graph coordinate system.
4. The method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition according to claim 1, characterized in that, Step four includes: Extract the pest category, prediction confidence level, and local bounding box coordinates in the segmentation graph coordinate system for each identified pest from the structured prediction results. Then, perform matrix addition aggregation operation on the local bounding box coordinates in the segmentation graph coordinate system and the absolute pixel anchor coordinates of the corresponding segmentation graph in the multimodal high-resolution remote sensing image to obtain the global bounding box coordinates of the identified pest in the global pixel coordinate system of the multimodal high-resolution remote sensing image. In the global pixel coordinate system, the global bounding boxes of the identified pests and diseases are sorted in descending order according to the prediction confidence of the pest and disease category corresponding to the pest and disease; and the cross-union ratio between overlapping global bounding boxes is calculated. Redundant global bounding boxes with cross-union ratio greater than a preset suppression threshold are removed to obtain a deduplicated set of global bounding boxes. Based on the flight attitude of the UAV, a spatial affine transformation matrix is constructed. Using the spatial affine transformation matrix, the coordinate system projection transformation is performed on the center point pixel coordinates of each global bounding box in the deduplicated global bounding box set to obtain the real geographic coordinates. The real geographic coordinates are then associated and encapsulated with the pest category to generate and output structured pest spatial distribution data.
5. The method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition according to claim 2, characterized in that, The process of dividing the acquired high-resolution remote sensing image into multiple segmented sub-images according to a preset sliding window overlap rate and a calculated step size includes: Construct horizontal and vertical index variables within a sliding window to perform sliding traversal in the pixel coordinate system of the high-resolution remote sensing image; The absolute pixel anchor point coordinates of each segmented sub-map in the multimodal high-resolution remote sensing image are calculated using a boundary adaptive truncation mechanism. Using the calculated coordinates of each absolute pixel anchor point as the starting point for slicing, and with the window length and window height as the fixed slicing range, a local slicing operation is performed on the high-resolution remote sensing image to obtain a sliced sub-map.
6. The method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition according to claim 1, characterized in that, The lightweight model based on multi-level self-distillation includes: a data processing module, a backbone network, a neck network, a decoupled detection head, and an output terminal.
7. The method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition according to claim 3, characterized in that, The process of extracting geometric edges and semantic features from standardized floating-point 3D data tensors through a backbone network to obtain multiple benchmark feature maps at different scales includes: Channel upscaling, spatial sliding extraction, and linear dimensionality reduction compression are performed on standardized floating-point 3D data tensors to obtain geometric features with basic contours and textures of multimodal high-resolution remote sensing images. The geometric features are then abstracted step by step to obtain primary feature maps with multi-level information. The bidirectional coordinate attention module performs one-dimensional global average pooling on the primary feature map along the vertical Y-axis and the horizontal X-axis respectively. The pooled directional features are then concatenated to reduce the dimensionality and generate a spatial attention weight matrix. The spatial attention weight matrix is then applied to the primary feature map for feature fusion by multiplying element by element to obtain a recalibrated feature map. The recalibration feature maps of multiple pre-defined specific downsampling depth nodes are extracted and organized into independent and parallel multi-scale feature sets. These sets are then distributed to subsequent networks through multiple independent parallel data lines to serve as recalibration reference feature maps at multiple different scales.
8. The method for unmanned aerial vehicle (UAV) inspection of farmland pests and diseases based on remote sensing image recognition according to claim 3, characterized in that, The adaptive reshaping and cross-scale fusion of recalibrated baseline feature maps at multiple different scales to obtain a multi-level fused feature map pyramid includes: Parallel spatial two-dimensional offset calculation is performed on multiple recalibrated reference feature maps of different scales using a deformable convolution module. Based on the calculated spatial two-dimensional offset, the recalibrated reference feature maps of different scales are processed to obtain deformable feature maps of different scales. Convolution is performed on deformable feature maps at different scales to generate intrinsic feature maps at different scales. Then, linear transformations are applied to intrinsic feature maps at different scales to generate "phantom" copies at different scales. Finally, intrinsic feature maps at different scales are concatenated with "phantom" copies at different scales to obtain lightweight feature maps at different scales with half the number of parameters and computational cost without losing feature richness. A bidirectional feature transfer path is constructed, which is both top-down and bottom-up. Learnable weights are introduced to perform cross-scale weighted concatenation of lightweight feature maps at different scales, so as to obtain a multi-level fused feature map pyramid that integrates deep strong semantics and shallow high-resolution details.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method for inspecting farmland pests and diseases by drone based on remote sensing image recognition as described in any one of claims 1 to 8.