A method and system for detecting forest vegetation diseases and pests based on unmanned aerial vehicles (UAVs)

Abstract

This application relates to the field of computer vision technology and provides a method for detecting forest vegetation diseases and pests based on unmanned aerial vehicles (UAVs). The method includes: acquiring forest disease and pest images using a UAV; preprocessing the images to generate an initial input feature map; performing feature enhancement on the initial input feature map to obtain a first feature map; performing saliency gate calibration and feature encoding on the first feature map to generate a multi-scale fusion feature map; performing multi-level fusion on the multi-scale fusion feature map to obtain first feature maps of different scales; and performing reparameterized large kernel convolution on each first feature map to output a target feature map, which can locate small disease and pest areas. This application introduces targeted improvements to the YOLO11 model at each processing stage to generate an MDFR-YOLO model, thereby enhancing the network's ability to detect forest vegetation diseases and pests in complex backgrounds.

Description

Technical Field

[0001] This application belongs to the field of computer vision technology, and in particular relates to a method and system for detecting forest vegetation diseases and pests based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Traditional forestry patrols rely on manual on-site surveys, which are inefficient, have limited coverage, and pose significant safety risks in complex terrain. With the continuous advancement of artificial intelligence and low-altitude remote sensing technology, a new solution has been developed for large-scale, high-frequency intelligent monitoring of forest pests and diseases, based on remote sensing images of forests taken by drones combined with deep learning algorithms.

[0003] Existing technologies have been extensively studied in the detection of forest vegetation and crop pests and diseases, but most results still rely on clear close-up images of the targets. When directly applied to large-scale aerial photography data from UAVs, existing solutions still have significant shortcomings: First, solutions to key challenges such as small target detection, multi-scale changes, and dense occlusion are often limited to improvements in individual technical aspects, lacking a systematic and comprehensive consideration of the entire processing flow; Second, when processing weak features from the UAV's perspective, the simple channel and spatial compression mechanisms of existing network structures easily destroy high-frequency local textures, leading to excessive smoothing and loss of minor pest and disease features, resulting in a persistently high false negative rate; In addition, existing attention mechanisms still have room for improvement in balancing the suppression of background noise in complex forest areas with the enhancement of target focus, and are insufficient in suppressing target confirmation bias caused by surrounding background interference.

[0004] Therefore, there is an urgent need for a method for detecting forest vegetation diseases and pests that can preserve the features of small targets without loss, effectively suppress interference from complex backgrounds, and have accurate global perception capabilities. Summary of the Invention

[0005] This application provides a method and system for detecting forest vegetation diseases and pests based on unmanned aerial vehicles (UAVs), which can solve one of the problems in the prior art mentioned above.

[0006] In a first aspect, embodiments of this application provide a method for detecting forest vegetation pests and diseases based on unmanned aerial vehicles (UAVs), including: Images of forest pests and diseases are collected by drones, and these images are preprocessed to generate an initial input feature map. The initial input feature map is enhanced to obtain a first feature map; The first feature map is subjected to saliency gate calibration and feature encoding to generate a multi-scale fused feature map; Multi-scale fusion feature maps are fused at multiple levels to obtain first feature maps at different scales; Each of the first feature maps is subjected to a reparameterized large kernel convolution to output a target feature map, which can locate areas of minor pests and diseases.

[0007] Furthermore, the step of performing feature enhancement on the initial input feature map to obtain the first feature map includes: The initial input feature map is divided into multiple local blocks, and the pixels of each local block are transferred to the channel dimension according to their spatial location. The channel dimensions are evaluated and calibrated using an attention mechanism to obtain high-quality features; The high-quality features are subjected to feature fusion and dimensionality reduction operations to output the first feature map.

[0008] Furthermore, the step of performing saliency gate calibration and feature encoding on the first feature map to generate a multi-scale fused feature map includes: The first feature map is reshaped along the channel dimension and divided into multiple channel groups; Channel weights are set for each of the channel groups, and based on the channel weights, saliency gate calibration is performed on the grouping features of each of the channel groups to obtain grouping calibration features; The grouped calibration features are subjected to bi-branch feature extraction to obtain the first branch features and the second branch features; Cross-dimensional feature aggregation is performed on the first branch features and the second branch features to obtain a fusion spatial attention map; The fused spatial attention map is non-linearly activated into an attention weight mask using the Sigmoid function; The attention weight mask is multiplied element-wise with the group calibration features to obtain a multi-scale fused feature map.

[0009] Furthermore, the step of setting channel weights for each channel group and performing salient gate calibration on the grouping features of each channel group based on the channel weights to obtain grouping calibration features includes: The grouped feature tensors are compressed in spatial dimension through a global adaptive average pooling layer, and information exchange between channels is performed through a single-kernel convolutional layer to obtain the output feature tensors of each channel group. The output feature tensor is normalized to a dynamic gating scaling factor using the Sigmoid activation function. The dynamic gating scaling factor is then multiplied element-wise with the grouping features of each channel group to obtain the grouping calibration features.

[0010] Furthermore, obtaining the first branch feature includes: One-dimensional adaptive average pooling is used to encode the grouped calibration features along the width and height directions of the space, respectively, to generate width-direction sensing features and height-direction sensing features. The width direction perception feature and the height direction perception feature are spliced ​​together in the spatial dimension, and information exchange between channels is performed to obtain directional attention features; The directional attention features and the group calibration features are weighted to obtain the first branch features.

[0011] Furthermore, for the second branch feature, the grouped calibration feature is further mined using 3×3 convolution to explore the spatial context information within the local receptive field.

[0012] Furthermore, the step of performing cross-dimensional feature aggregation on the first branch features and the second branch features to obtain a fused spatial attention map includes: Global average pooling and Softmax activation are performed on the first branch features and the second branch features respectively to obtain the first branch channel attention weights and the second branch channel attention weights; Based on the attention weights of the first branch channel and the second branch channel, matrix multiplication is used to cross-multiply the first branch features and the second branch features, and the results of the cross-multiplication are added together to obtain the fusion spatial attention map.

[0013] Furthermore, the reparameterized large-kernel convolution on the first feature map includes: The first feature map is input into a large-scale convolutional kernel and a small-scale convolutional kernel; In the small-scale convolutional kernel, zero padding is performed in the spatial aspect to align its size with that of the large-scale convolutional kernel, and the padded small-scale weight matrix is ​​obtained. The small-scale weight matrix and the large-scale weight matrix are added element-wise to form an equivalent large convolution kernel.

[0014] Secondly, embodiments of this application provide a forest vegetation pest and disease detection system based on unmanned aerial vehicles (UAVs), including: First processing module: used to collect images of forest pests and diseases through drones, and to preprocess the images of forest pests and diseases to generate an initial input feature map; The second processing module is used to perform feature enhancement on the initial input feature map to obtain the first feature map. The third processing module is used to perform salient gate calibration and feature encoding on the first feature map to generate a multi-scale fused feature map. The fourth processing module is used to perform multi-level fusion of multi-scale fusion feature maps to obtain first feature maps at different scales. The fifth processing module is used to perform reparameterized large kernel convolution on each of the first feature maps to output a target feature map, which can locate the area of ​​small pests and diseases.

[0015] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for detecting forest vegetation diseases and pests based on unmanned aerial vehicles.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium, including a computer program stored in the computer-readable storage medium, which, when executed by a processor, implements the above-described method for detecting forest vegetation diseases and pests based on unmanned aerial vehicles.

[0017] The beneficial effects of the embodiments in this application compared with the prior art are: This application discloses a method for detecting forest vegetation diseases and pests based on unmanned aerial vehicles (UAVs). By introducing targeted improvements to the YOLO11 model at each processing stage, an MDFR-YOLO model is generated to enhance the network's ability to detect forest vegetation diseases and pests in complex backgrounds. Specifically, based on depth-to-space convolution technology, a SEFM module with spatial-to-depth convolution enhancement fusion is proposed. This module enhances forest disease and pest features while preserving subtle features without loss. Furthermore, a CWC-EMA module with a dynamic feature selection mechanism is proposed. This module deeply reshapes spatial and channel features, strengthening the network's focus on key disease and pest features while effectively suppressing background noise in forest areas. Addressing the issue of traditional detection heads having small receptive fields, which easily lead to missed or false detections of forest diseases and pests, a RepLK module with reparameterized large kernel convolution is introduced. By expanding the receptive field, the accuracy of identification is further improved in the target confirmation stage. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for detecting forest vegetation pests and diseases based on unmanned aerial vehicles (UAVs) according to an embodiment of the present invention. Figure 2 yes Figure 1 The overall network architecture diagram of the MDFR-YOLO model in the embodiment shown is as follows; Figure 3 yes Figure 1 A flowchart of the SEFM module in the illustrated embodiment; Figure 4 yes Figure 1 A flowchart illustrating the CWC-EMA module of the embodiment shown; Figure 5 yes Figure 1 A flowchart of the RepLK module in the illustrated embodiment; Figure 6 These are the feature map activation heatmaps output by the YOLO11 model and the MDFR-YOLO model in the test experiments of this invention; wherein, Figure (a) shows the original image; Figure (b) shows the feature map activation heatmap output by the baseline model YOLO11; Figure (c) shows the feature map activation heatmap output by the YOLO11+SEFM+CWC_EMA+RepLK model of this application; Figure 7 This is the convergence curve of the training loss and evaluation metric of the MDFR-YOLO model in the test experiment of this invention; wherein, Figure (a) shows the curve of training bounding box loss as a function of the number of iterations; Figure (b) shows the curve of training classification loss as a function of the number of iterations; Figure (c) shows the curve of the training distribution focus loss as a function of the number of iterations; Figure (d) shows the curve of the validation bounding box loss as a function of the number of iterations; Figure (e) shows the curve of the validation classification loss as a function of the number of iterations; Figure (f) shows the curve of the validation distribution focus loss as a function of the number of iterations; Figure (g) shows the curve of accuracy as a function of the number of iterations; Figure (h) shows the recall rate as a function of the number of iterations; Figure (i) shows the curve of the mean accuracy mAP@0.5 as a function of the number of iterations; Figure 8 This refers to the actual detection results of various models in the test experiments of this invention; among them, Figure (a) shows the original drone aerial imagery; Figure (b) shows the actual detection performance of the benchmark model YOLO11; Figure (c) shows the actual detection performance of the YOLO11 + SEFM model; Figure (d) shows the actual detection performance of the YOLO11 + CWC_EAM model; Figure (e) shows the actual detection performance of the YOLO11 + RepLK model; Figure (f) shows the actual detection performance of the YOLO11 + SEFM + CWC_EAM model; Figure (g) shows the actual detection performance of the YOLO11 + SEFM + RepLK model; Figure (h) shows the actual detection performance of the YOLO11 + CWC_EAM + RepLK model; Figure (i) shows the actual detection performance of the YOLO11+SEFM+CWC_EMA+RepLK model in this application; Figure 9 This is a schematic diagram of a forest vegetation pest and disease detection system based on an unmanned aerial vehicle (UAV) according to an embodiment of the present invention. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0021] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0022] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0023] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0024] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0025] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0026] like Figure 2 As shown, the UAV-based forest vegetation pest and disease detection method of this application is based on the YOLO11 model and introduces targeted improvement measures at different stages to generate a high-precision detection network for the entire process of UAV forest vegetation pest and disease detection. Specifically, by improving the YOLO11 model in multiple stages, a multi-dimensional feature reconstruction detection model is generated to detect forest vegetation pests and diseases. This effectively prevents the loss of weak features of pest and disease targets in large-scale real-world scenarios, achieving lossless preservation and feature enhancement of weak features. Furthermore, while strengthening the network's focus on the main features of pests and diseases, it effectively suppresses background noise in forest areas and reduces the interference of the actual surrounding background on target confirmation, significantly improving the accuracy of identification.

[0027] It is worth noting that YOLO11 is a real-time object detection model, optimized based on lightweight CNN modules such as C3K2 and C2PSA. It boasts high computational density and maintains extremely low latency across various hardware platforms, remaining one of the fastest models in the YOLO series to date. Furthermore, YOLO11 is not demanding on dataset size, converging quickly and exhibiting low overfitting even on small-sample agricultural and forestry pest and disease datasets. Based on these reasons, this application selects the YOLO11 framework as its foundation and introduces targeted improvements at each processing stage to generate an improved multidimensional feature reconstruction detection model, namely the MDFR-YOLO model, further enhancing the network's ability to detect forest vegetation pests and diseases in complex contexts.

[0028] Specifically, the YOLO11 model architecture includes a backbone network, a neck network, and a head. The backbone network extracts image features step by step through convolution and downsampling operations and generates feature maps of different scales. For example, when the input image size is 640×640, the backbone network may generate feature maps such as P2 (160×160), P3 (80×80), P4 (40×40), and P5 (20×20). Meanwhile, the neck network further integrates feature maps of different scales, combining shallow and deep features through upsampling and lateral connections to enhance the model's ability to detect targets at different scales. Finally, the multi-scale feature maps from the neck network are sequentially passed through independent head detectors for target detection.

[0029] Please see Figure 1 As shown, this invention is a method for detecting forest vegetation diseases and pests based on unmanned aerial vehicles (UAVs), comprising the following steps: S100. Collect images of forest pests and diseases using a drone, and preprocess the images to generate an initial input feature map. In this embodiment, a high-precision drone aerial forest image dataset is constructed for training and testing the improved multi-dimensional feature reconstruction detection model. This high-precision drone aerial forest image dataset is collected by drones in real complex forestry scenarios and includes images of forest pests and diseases under various lighting conditions, different flight altitudes, and foliage cover conditions.

[0030] Specifically, before training and testing using the images in the aforementioned high-precision UAV aerial forest image dataset, the images need to be uniformly adjusted to a size of 640×640 pixels as the initial input feature map to conform to the standard input resolution of the YOLO11 model.

[0031] S200: Perform feature enhancement on the initial input feature map to obtain a first feature map; Based on depth-to-space convolution technology, this application proposes a spatial-to-depth convolution enhanced fusion SEFM module, which can enhance the features of forest pests and diseases while achieving lossless preservation of small features.

[0032] In some embodiments, step S200 above includes: The initial input feature map is divided into multiple local blocks, and the pixels of each local block are transferred to the channel dimension according to their spatial location. The channel dimensions are evaluated and calibrated using an attention mechanism to obtain high-quality features; The high-quality features are subjected to feature fusion and dimensionality reduction operations to output the first feature map.

[0033] In the existing YOLO11 network architecture, the backbone network typically uses ordinary convolution or pooling operations with a stride of 2 for image downsampling. This conventional strided convolution, while reducing the spatial resolution of the feature map, irreversibly discards some pixel information, causing fine-grained features such as edge textures and color spots to rapidly degrade or even disappear completely during network deepening, leading to severe missed detections. Therefore, to effectively address the problem of feature loss in small targets, this application introduces a spatial-to-depth convolutional enhancement fusion (SEFM) module into the backbone network to replace the original downsampling layer, performing feature enhancement processing on the initial input feature map and improving its expressive power.

[0034] Specifically, the aforementioned SEFM module couples spatial-to-depth pixel rearrangement technology with a dynamic feature selection mechanism, enabling immediate feature recovery and enhancement during the downsampling and recombination stage. Its core idea is to losslessly convert spatial dimension information to channel dimension through "spatial-to-depth" pixel rearrangement operation, and then use the internally integrated attention module for feature calibration, thereby achieving downsampling without loss of information.

[0035] Specifically, such as Figure 3 As shown, assuming the initial input feature map has dimensions H×W×C, where H and W represent the height and width in spatial resolution, respectively, and C represents the number of input channels, and the scale factor is set to scale=2, the SPD operation will systematically sample the initial input feature map spatially according to a predetermined stride, dividing the initial input feature map into multiple 2×2 local blocks. Pixels within each local block are then separated based on their spatial location and recombined into channels. Understandably, through the above operation, the spatial resolution of the initial input feature map is precisely halved, becoming... However, since no pixels were discarded, this spatial information was completely transferred to the channel dimension, causing the number of channels to multiply, becoming C. This process physically achieves zero information loss in the feature downsampling process. After lossless segmentation, the graph dimension of the initial input features becomes... .

[0036] To address the potential for redundant background noise in the numerous recombined channels, the SEFM module innovatively embeds an attention mechanism at this stage instead of using convolutional dimensionality reduction. This attention mechanism directly acts on the intermediate feature tensor output by the pixel rearrangement operation, adaptively evaluating and calibrating the channels containing rich spatial details through a gating mechanism. This allows for the precise extraction of texture and color representations highly correlated with pests and diseases before feature channels are compressed, while suppressing ineffective background channels and obtaining high-quality features. Subsequently, a standard convolutional layer with a stride of 1 and a kernel size of k is applied to the attention-enhanced expanded channels without changing the feature map space size. This performs cross-channel feature fusion and dimensionality reduction operations, ultimately outputting a downsampled feature map mapped to the target number of channels. Therefore, in the backbone network, this SEFM module can downsample the initial input feature map to obtain first feature maps of different sizes after feature enhancement.

[0037] S300. Perform saliency gate calibration and feature encoding on the first feature map to generate a multi-scale fused feature map; This application proposes a CWC-EMA module that integrates a dynamic feature selection mechanism for channel weight calibration and efficient multi-scale attention. This module replaces the C3k2 module in the existing YOLO11 model. It deeply reshapes spatial and channel features, strengthens the network's focus on the main features of pests and diseases, and effectively suppresses background noise in forest areas.

[0038] In some embodiments, step S300 above includes: The first feature map is reshaped along the channel dimension and divided into multiple channel groups; Channel weights are set for each of the channel groups, and based on the channel weights, saliency gate calibration is performed on the grouping features of each of the channel groups to obtain grouping calibration features; The grouped calibration features are subjected to bi-branch feature extraction to obtain the first branch features and the second branch features; Cross-dimensional feature aggregation is performed on the first branch features and the second branch features to obtain a fusion spatial attention map; The fused spatial attention map is non-linearly activated into an attention weight mask using the Sigmoid function; The attention weight mask is multiplied element-wise with the group calibration features to obtain a multi-scale fused feature map.

[0039] Furthermore, after obtaining the first feature maps of different sizes, feature fusion is required to obtain multi-scale fused features. In the backbone network of existing technologies, the C3k2 module is used to achieve multi-scale feature fusion. Specifically, the C3k2 module adopts a dual-branch design. One branch uses 3×3 convolution to capture local features, and the other branch uses 1×1 convolution for channel interaction. Multi-scale feature representations are fused through residual connections. This C3k2 module enhances feature extraction capabilities while reducing the number of parameters, enabling the model to maintain high detection accuracy even in complex backgrounds. However, when processing feature maps, the C3k2 module treats both channel and spatial aspects equally, lacking the ability to selectively focus on key local information. From the perspective of aerial view from a drone, the forest background is very complex. The color and texture of healthy branches and leaves are often very similar to areas infected by pests and diseases. At the same time, due to factors such as uneven lighting and foliage occlusion, the effective features of pests and diseases are easily submerged in background noise. Therefore, the C3k2 module in existing technologies cannot strengthen the focus on the main features of pests and diseases. Therefore, this application improves the C3k2 module into the CWC-EMA module, which integrates a dynamic feature selection mechanism and channel weight calibration for efficient multi-scale attention. Specifically, the CWC-EMA module introduces saliency gating on the basis of the original multi-scale attention architecture, while the EMA module uses multiple parallel branches to capture spatial orientation information and local receptive field features respectively. Through cross-spatial feature aggregation and multi-scale feature extraction, the feature representation is enhanced, thereby achieving effective purification of subtle pest and disease features and accurate removal of background noise.

[0040] Furthermore, to reduce computational overhead and enable the network to learn diverse local semantic representations, such as... Figure 4 As shown, the CWC-EMA module first reshapes the feature tensor of the input first feature map along the channel dimension and divides it into G non-overlapping channel groups. This channel grouping mechanism not only eliminates the redundancy of full-channel convolution but also lays the foundation for subsequent cross-spatial feature aggregation. Then, to compensate for the lack of channel importance evaluation in the backbone network of existing technologies after feature grouping, the CWC-EMA module adds a saliency gate after the grouping operation to perform saliency gate calibration.

[0041] In some embodiments, the step of setting channel weights for each of the channel groups and performing significant gate calibration on the grouping features of each of the channel groups based on the channel weights to obtain grouping calibration features includes: The grouped feature tensors are compressed in spatial dimension through a global adaptive average pooling layer, and information exchange between channels is performed through a single-kernel convolutional layer to obtain the output feature tensors of each channel group. A dynamic gating scaling factor is generated by using the Sigmoid activation function. The dynamic gating scaling factor is then multiplied element-wise with the grouping features of each channel group to obtain the grouping calibration features.

[0042] In this embodiment, the grouped feature tensors are first compressed in space dimension by a global adaptive average pooling layer, and then a single-kernel convolutional layer is used to perform information interaction between channels without changing the number of channels to generate an output feature tensor. The output feature tensor is then generated by a Sigmoid activation function, which generates a dynamic gating scaling factor with a value between 0 and 1.

[0043] Specifically, the shape of the grouped feature tensor is [B,C,H,W], where B is the batch size, C is the number of channels, and H / W represents the spatial dimension. If a first feature map is divided into G groups, then each group contains C / G channels, thereby reducing the amount of computation and allowing different groups to learn independent feature patterns.

[0044] Furthermore, global average pooling is performed on the spatial dimension of the grouped features for each channel group to output the spatial mean of each channel, thereby compressing spatial information and obtaining the global statistics for each channel. The pooling window size is automatically determined by the spatial size of the input features. Then, a 1×1 convolution is applied to the pooled features. Through the weight sharing of the convolution kernel, the features of different channels can influence each other without changing the number of channels, keeping the feature dimensions consistent, and finally obtaining the output feature tensor.

[0045] Furthermore, the output feature tensor of the convolution is applied to the Sigmoid function, mapping the values ​​to a dynamic gating scaling factor in the [0,1] interval. This dynamic gating scaling factor is then multiplied element-wise with the grouping features of each channel group to achieve adaptive calibration of channel importance, obtaining the grouped calibration features for the corresponding channel group. Thus, by introducing a dynamic gating scaling factor to dynamically assign corresponding weights to different channels, redundant background channels similar to health features are effectively suppressed, highlighting the key features of pests and diseases, thereby achieving a precise dynamic feature selection function.

[0046] Furthermore, after obtaining the group calibration features of each channel group, the feature encoding stage is entered, and the features are decoupled and input to two parallel extraction branches, namely the direction sensing branch and the local feature branch, so as to obtain the first branch features and the second branch features respectively.

[0047] In some embodiments, obtaining the first branch feature includes: One-dimensional adaptive average pooling is used to encode the grouped calibration features along the width and height directions of the space, respectively, to generate width-direction sensing features and height-direction sensing features. The width direction perception feature and the height direction perception feature are spliced ​​together in the spatial dimension, and information exchange between channels is performed to obtain directional attention features; The directional attention features and the group calibration features are weighted to obtain the first branch features.

[0048] Specifically, the orientation-aware branch uses one-dimensional adaptive average pooling to encode the group calibration features along the height and width directions in space, generating two sets of orientation-aware features: width orientation features and height orientation features. Subsequently, the width and height orientation features are concatenated in the spatial dimension and used for inter-channel information exchange through 1×1 convolution. Then, the output tensor after the 1×1 convolution is divided into two parts along the channel dimension: a width tensor and a height tensor. After processing through the Sigmoid activation function, width and height attention features are obtained. The orientation attention features are multiplied by the group calibration features to complete the initial feature weighting, obtaining the first group features.

[0049] Furthermore, the aforementioned first group features need to undergo group normalization to accelerate network convergence and alleviate internal covariate bias.

[0050] In some embodiments, for the second branch feature, the grouped calibration feature is further mined using 3×3 convolution to explore the spatial context information within the local receptive field.

[0051] In this embodiment, the second branch feature is the acquisition of local features. Corresponding to the local feature branch, 3×3 convolution is used in this branch to deeply mine the spatial context information in the local receptive field of the grouped calibration features, so as to preserve the high-frequency texture and color distribution details of the pest and disease features to the greatest extent.

[0052] In some embodiments, the step of performing cross-dimensional feature aggregation on the first branch features and the second branch features to obtain a fused spatial attention map includes: Global average pooling and Softmax activation are performed on the first branch features and the second branch features respectively to obtain the first branch channel attention weights and the second branch channel attention weights; Based on the attention weights of the first branch channel and the second branch channel, matrix multiplication is used to cross-multiply the first branch features and the second branch features, and the results of the cross-multiplication are added together to obtain the fusion spatial attention map.

[0053] In this embodiment, bi-branch feature extraction is performed on group calibration scales of different scales to obtain first-branch features and second-branch features of different sizes. Then, a fused feature attention map is obtained based on a cross-dimensional feature aggregation mechanism using matrix multiplication. Specifically, global average pooling is performed on the two branch features to compress the spatial dimension, while Softmax activation is performed to extract the channel attention weights of the corresponding branches, namely the first-branch channel attention weights and the second-branch channel attention weights.

[0054] Furthermore, through matrix multiplication, the channel attention weights of one branch are cross-multiplied with the flattened two-dimensional spatial feature matrix of the other branch. This cross-multiplication operation not only forces the probability distribution normalization within a feature group, enabling the network to focus its attention on the high-frequency disease and pest areas with the highest scores when facing complex forest backgrounds, but also achieves the cross-aggregation of the two feature paths.

[0055] Furthermore, the results of the cross-products of the two branches are summed to generate a multi-scale fused spatial attention map. Finally, this fused spatial attention map is non-linearly activated by the Sigmoid function to become an attention weight mask, and then multiplied element-wise with the grouped calibration features after saliency gate calibration. At this time, the features of high-weight regions are preserved or enhanced, while the features of low-weight regions are suppressed, generating a multi-scale fused feature map. This completes the adaptive recalibration of multi-scale features and reshapes them back to the original input tensor dimension for subsequent processing.

[0056] Specifically, such as Figure 2 The improved framework shown includes improvements to the backbone network of the YOLO11 model. Based on the SEFM module for spatial-to-depth convolutional enhancement fusion and the CWC-EMA module for efficient multi-scale attention with channel weight calibration, multi-stage downsampling and multi-scale feature fusion are performed on the initial input feature map. Specifically, in stage P1, the 640×640 initial input feature map is subjected to a convolution operation with a stride of 2. Then, in stages P2-P5, based on the SEFM module and the CWC-EMA module, progressive downsampling and multi-scale feature fusion are performed to obtain a multi-scale fused feature map.

[0057] S400. Perform multi-level fusion on the multi-scale fusion feature maps to obtain first feature maps at different scales; like Figure 2 As shown, for the shallow features of stage P3, the mid-level features of stage P4, and the deep features of stage P5, it is also necessary to combine the multi-scale fusion feature maps generated by the backbone network and input them into the neck network to perform fusion at different levels in order to improve detection performance.

[0058] Specifically, the neck network employs a PAN-FPN architecture. FPN fuses high-level semantic information with low-level detail information through upsampling and lateral connections, generating a multi-scale feature pyramid. Specifically, upsampling enlarges the size of high-level feature maps for fusion with low-level feature maps, and concatenates feature maps from different levels along the channel dimension to integrate multi-scale information. PAN enhances and adjusts attention through SEFM and CWC-EMA modules, while concatenation fuses features from different paths, ensuring sufficient information flow between different levels. Finally, the neck network generates first feature maps corresponding to different scales.

[0059] S500. Perform reparameterized large kernel convolution on the first feature map to output a target feature map, which can locate areas of small pests and diseases.

[0060] This application addresses the issue that the receptive field of the existing YOLO11 model is small, which easily leads to missed detections and false detections in the detection of forest pests and diseases. It introduces the RepLK module with reparameterized large kernel convolution to further improve the accuracy of identification in the target confirmation stage by expanding the receptive field.

[0061] In some embodiments, performing reparameterized large-kernel convolution on the first feature map includes: The first feature map is input into a large-scale convolutional kernel and a small-scale convolutional kernel; In the small-scale convolutional kernel, zero padding is performed in the spatial aspect to align its size with that of the large-scale convolutional kernel, and the padded small-scale weight matrix is ​​obtained. The small-scale weight matrix and the large-scale weight matrix are added element-wise to form an equivalent large convolution kernel.

[0062] In this embodiment, the first feature maps of different scales obtained in step S500 are respectively input into independent detection heads for target detection to obtain target feature maps containing target bounding boxes, thereby locating the small pest and disease areas.

[0063] Specifically, in complex forest scenes captured by drones, the boundaries between pest-infested areas and healthy vegetation often appear gradual and blurred. To accurately pinpoint pest-infested areas and reduce false detection rates, detection networks typically need sufficient global spatial awareness to understand the relative relationship between pests and their surrounding branches and leaves. However, the existing YOLO11 network primarily uses small 3×3 convolutional kernels for local feature extraction in its detection head, resulting in a very limited receptive field.

[0064] Therefore, in order to improve the receptive field, the multi-dimensional feature reconstruction detection model of this application introduces a reparameterized large kernel convolution module, namely the RepLK module, into the deep network. During the training phase, this module adopts a parallel dual-branch topology of 9×9 large kernel and 1×1 small kernel to capture rich feature gradient flow; while during the inference deployment phase, the linear homogeneity theorem of the convolution operator is used to fuse the multi-branch into a single cylindrical convolutional layer.

[0065] Specifically, such as Figure 5 As shown, during the model training phase, the reparameterized large kernel convolution module is a parallel dual-branch structure. For the input feature map X, this reparameterized large kernel convolution module simultaneously runs a 9×9 large convolution kernel and a 1×1 small convolution kernel. The large convolution kernel captures long-range dependencies and global macro-features, while the small convolution kernel preserves local fine-grained features and promotes cross-channel information interaction. Its forward propagation process can be represented as: Where * represents the convolution operation, and These are the weight matrices for the large and small convolutional kernels, respectively. BN represents batch normalization. Through the design of multiple branches in parallel, it not only provides the network with diverse multi-scale feature gradient flows, but also effectively reduces the gradient vanishing that is very easy to occur in the initial stage of training with large kernel convolution.

[0066] More specifically, when the model training converges, the reparameterized large kernel convolution module performs a structural reparameterization transformation based on the linear homogeneity of the convolution operation. Specifically, it first absorbs the scaling and offset parameters of the batch normalization (BN) layer into the corresponding convolutional layer weights and biases. Then, it performs zero-padding on the 1×1 convolutional kernel to align its size to a 9×9 kernel. Finally, it adds the padded 1×1 weight matrix to the original 9×9 weight matrix element-wise, merging them into a single equivalent large convolutional kernel. Specifically, the transformed forward propagation formula is greatly simplified to: Where Y represents the output feature map, X represents the input feature map, and W... fused B represents the fusion weight. fused This indicates the fusion bias.

[0067] Therefore, the aforementioned multi-dimensional feature reconstruction detection model replaces all the grouped convolutions in the deep network of the detection head in the YOLO11 standard model with heavily parameterized large kernel convolutional modules, allowing the network to fully enjoy the high-precision spatial localization capabilities brought by the large receptive field of multiple branches during training; during the inference process, the network scheme is losslessly compressed into a very concise single-layer convolution. This improvement measure achieves receptive field enhancement with zero additional inference time in terms of physics, accurately meeting the dual strict requirements of high precision and high frame rate for drone aerial pest detection.

[0068] In one embodiment of this application, a test experiment is provided for the above-mentioned method for detecting forest vegetation pests and diseases based on unmanned aerial vehicles (UAVs) to test the performance of the multi-dimensional feature reconstruction detection model for UAVs based on YOLO11 improved in this application.

[0069] The experimental dataset used in this test experiment is the PDT dataset (Pests and Diseases TreeDataset), the first high-precision tree pest and disease target detection dataset based on drones. It addresses the performance limitations of AI pest and disease detection models in agriculture due to a lack of professionally labeled data. Specifically, the PDT dataset is a high-precision drone aerial forest image dataset, collected by drones in real, complex forestry scenarios, including images of forest pests and diseases under various lighting conditions, flight altitudes, and foliage shading. Furthermore, before the test experiment, all images were uniformly resized to 640×640 pixels to meet the standard input resolution of the YOLO11 model.

[0070] In this test experiment, in order to verify the application effect of the above multi-dimensional feature reconstruction detection model in the actual detection of pests and diseases in forest areas, a system comparison experiment was conducted under the same training parameter settings. The basic version of the YOLO11 model was used as the benchmark, and the detection results of the PDT dataset were used as the reference. The detailed experimental results are shown in Table 1 below. Table 1. Comparison of pest and disease detection performance of different models:

[0071] As shown in Table 1 above, compared with the benchmark model YOLO11, the multi-dimensional feature reconstruction detection model proposed in this application is superior in all major indicators of detection capability. The precision is improved by approximately 3.4% (89.81%-86.87%) / 86.87%, and similarly, the recall is improved by 3.1%, the mean precision is improved by 2.4%, and the overall F1 score is improved by 3.2%. This is mainly due to the improvements made in each processing stage.

[0072] Meanwhile, compared with the YOLO-DP model used when the PDT dataset was proposed, the recall rate is improved by (90.14%-88.0%) / 88.0%≈2.4%. This is mainly because the multi-dimensional feature reconstruction detection model of this application introduces spatial-to-depth convolution technology, which can better achieve lossless retention of weak target features, and improves the synchronization to achieve feature enhancement, reducing the false negatives of targets, thus significantly improving the recall rate. In terms of precision, the YOLO-DP model is basically equivalent to the model of this application, with a slight decrease of (90.2%-89.81%) / 89.81≈0.4%. This is due to the slight difference brought about by the attention mechanism chosen by each. However, in terms of overall detection ability, the average precision of this application is improved by 1.2%, and the overall F1 score is improved by 1.1%.

[0073] Therefore, this application improves the detection model based on the YOLO model from multiple dimensions such as feature information transmission, extraction and confirmation. Thus, the multi-dimensional feature reconstruction detection model proposed in this application is still superior to the YOLO-DP model. The multi-dimensional feature reconstruction detection model based on multi-dimensional feature reconstruction is more practical and robust in the detection of pests and diseases in the face of complex natural forest scenarios.

[0074] Furthermore, to more intuitively demonstrate the changes in the network's global perception and focus capabilities brought about by the proposed multi-dimensional feature reconstruction detection model, Figure 6 The feature map activation heatmaps output by the baseline model YOLO11 and the multidimensional feature reconstruction detection model MDFR-YOLO of this application are shown. It can be found that the multidimensional feature reconstruction detection model can more accurately locate small pest and disease areas and effectively reduce the activation response in complex forest backgrounds.

[0075] Furthermore, in order to fully examine the contribution of each improvement to enhancing the detection capability of pests and diseases, the YOLO11 model was used as the benchmark, and a progressive ablation experiment was carried out under the same training parameter settings. The detailed experimental results are shown in Table 2 below. Table 2 Ablation test results:

[0076] First, let's examine the impact of improvements in individual steps. As shown in Table 2 above, the results of Model (B) indicate that by introducing the SEFM module to replace traditional convolution, feature loss caused by conventional downsampling and pooling is avoided, resulting in a slight improvement in precision while recall remains approximately unchanged. This is mainly because information from weak features is preserved. However, due to the continued difficulty in extracting weak features, the overall performance is not significantly improved. The results of Model (C) show a slight increase in precision, reflecting the effect of introducing the CWC_EMA module into the backbone network. However, the recall rate decreases to some extent. This is because in standard CNNs, downsampling can cause small targets to remain with only a few pixels or even disappear in deep feature maps. The CWC_EMA module cannot apply attention to "non-existent" features, leading to an increase in missed detections. Therefore, the mean precision is not significantly improved. The results of Model (D) show that the overall metrics have increased slightly. Specifically, among the four core metrics mentioned above, precision, recall, and F1 all increased, and the overall performance score F1 also increased. This is mainly because by introducing RepLK, the target can be judged more accurately in a larger receptive field, further improving the false negative detection and reducing the false positive detection. However, since this improvement is in the final detection stage, and the previous input information has not been significantly improved, the overall improvement is not significant.

[0077] Secondly, the effects of combination improvement were examined through pairwise combinations. Table 2 shows that all combined models showed some improvement in overall performance compared to the baseline models. Specifically, comparing model (E) with models (B) and (C), it can be seen that the recall rate of the combined model was further improved, demonstrating the advantages of the combined model in feature information transmission and feature extraction, and the false negative rate was further improved. However, the precision decreased slightly, possibly due to insufficient bounding box regression accuracy, which is also supported by the increase in the overall score F1. Comparing model (F) with models (B) and (D), it can be seen that although SEFM and RepLK have their own advantages, the lack of an attention mechanism reduces feature extraction capability, leading to a slight decrease in the overall performance of the model. Comparing model (G) with models (C) and (D), we can see that in terms of recall, model (G) has a slightly higher recall than model (C). This is mainly because RepLK performs target identification in a larger receptive field, improving false negatives. However, compared to model (D), model (G) has a slightly lower recall, primarily due to the potential loss of subtle information after introducing CWC_EMA. In terms of precision, model (G) has a slightly lower precision than model (C), possibly because the secondary filtering of CWC_EMA disrupts the stable feature distribution learned by RepLK, introducing unnecessary feature fluctuations. Compared to model (D), model (G) has a slightly higher precision, mainly due to the introduction of the improved and efficient multi-scale attention mechanism CWC_EMA, which enhances the target feature extraction capability.

[0078] Therefore, as can be seen from the above ablation experiment process, this application starts from the entire processing flow and introduces improvement measures at each stage, and finally the multi-dimensional feature reconstruction detection model constructed achieves comprehensive superior performance.

[0079] Furthermore, in order to analyze the stability of the multi-dimensional feature reconstruction detection model during the training process, Figure 7 The figures show the decreasing trend of the loss function and the changing trend of the core evaluation metrics during the training period. Figures (a)-(c) show the curves of the training bounding box loss, classification loss, and distribution focus loss as a function of the number of iterations, respectively. Figures (d)-(f) show the curves of the validation bounding box loss, classification loss, and distribution focus loss as a function of the number of iterations, respectively. Figures (g)-(i) show the curves of the change of precision, recall, and mean precision mAP@0.5 as a function of the number of iterations, respectively.

[0080] like Figure 7As shown, in the initial stage of network training, both the bounding box loss and classification loss exhibit a rapid and steep decreasing trend, indicating that the model is efficiently learning the macroscopic characteristics of forest pests and diseases. With increasing iterations, the loss curves of the training and validation sets gradually flatten and closely match, without any abnormal rebound in the validation set loss. This demonstrates that the model has good generalization ability, and that the SEFM module, while preserving small target features, does not cause the model to overfit due to background noise.

[0081] Meanwhile, the accuracy evaluation curves show that the model's precision, recall, and mean precision rapidly increased in the early stages of iteration, and then entered a stable oscillating convergence period around 100 epochs. Specifically, thanks to the efficient multi-scale attention mechanism of the CWC-EMA module, which powerfully purifies features, the model's precision quickly remained high, with mAP@0.5 eventually converging stably at around 94.27%. Dynamic data throughout the training process demonstrates that the multi-dimensional feature reconstruction detection model constructed in this application exhibits fast convergence speed and strong robustness.

[0082] Furthermore, to more intuitively demonstrate the performance of the multi-dimensional feature reconstruction detection model in real-world application scenarios, some representative drone aerial images of complex forest areas were randomly selected from the test dataset for a comparative visualization study. The selected scenes included representative drone inspection scenarios such as densely packed small targets and severe foliage obstruction. Figure 8 The figure shows the actual detection performance of various models.

[0083] Depend on Figure 8 As shown, in areas where pests and diseases are small and densely distributed, other models exhibit varying degrees of false detection when faced with complex background interference and similar color occlusion. The model used in this application benefits from the unique lossless downsampling mechanism of the SEFM module, which perfectly preserves the fine-grained high-frequency characteristics of small objects. Simultaneously, it introduces the efficient multi-scale attention mechanism of the CWC-EMA module and uses the Softmax activation function for grouped feature purification, demonstrating strong background suppression capabilities and superior visual performance. The boundary alignment between the predicted bounding boxes and the actual pest and disease areas is significantly improved. With the help of the wide receptive field provided by the reparameterized large kernel convolution, the multi-dimensional feature reconstruction detection model can better understand the contextual topological relationships of the pest and disease area and its surrounding environment. When dealing with irregular, contiguous pest areas, the generated bounding boxes are more complete and accurate, without the situation where a single pest and disease area is segmented into multiple small boxes.

[0084] Overall, the visualized qualitative results are highly consistent with the quantitative indicators presented earlier. Faced with the complex environmental interference caused by drone aerial photography, the multi-dimensional feature reconstruction detection model constructed in this application demonstrates superior noise resistance stability and small target acquisition capability.

[0085] Please see Figure 9 As shown, the present invention also provides a forest vegetation pest and disease detection system based on unmanned aerial vehicles (UAVs), the system comprising: First processing module 901: used to collect images of forest pests and diseases through drones, and to preprocess the images of forest pests and diseases to generate an initial input feature map; The second processing module 902 is used to perform feature enhancement on the initial input feature map to obtain a first feature map; The third processing module 903 is used to perform saliency gate calibration and feature encoding on the first feature map to generate a multi-scale fused feature map; The fourth processing module 904 is used to perform multi-level fusion of multi-scale fusion feature maps to obtain first feature maps at different scales; The fifth processing module 905 is used to perform reparameterized large kernel convolution on each of the first feature maps and output a target feature map, which can locate the area of ​​small pests and diseases.

[0086] It is understandable that, such as Figure 1 The content of the UAV-based forest vegetation pest and disease detection method embodiments shown is applicable to the UAV-based forest vegetation pest and disease detection system embodiments. The specific functions implemented by the UAV-based forest vegetation pest and disease detection system embodiments are the same as those shown in the examples. Figure 1 The embodiment of the UAV-based forest vegetation pest and disease detection method shown is the same, and the beneficial effects achieved are the same as those described above. Figure 1 The beneficial effects achieved by the illustrated embodiment of the UAV-based forest vegetation pest and disease detection method are the same.

[0087] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0088] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0089] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for detecting forest vegetation diseases and pests based on unmanned aerial vehicles (UAVs), characterized in that, include: Images of forest pests and diseases are collected by drones, and these images are preprocessed to generate an initial input feature map. The initial input feature map is enhanced to obtain a first feature map; The first feature map is subjected to saliency gate calibration and feature encoding to generate a multi-scale fused feature map; Multi-scale fusion feature maps are fused at multiple levels to obtain first feature maps at different scales; Each of the first feature maps is subjected to a reparameterized large kernel convolution to output a target feature map, which can locate areas of minor pests and diseases.

2. The method as described in claim 1, characterized in that, The step of performing feature enhancement on the initial input feature map to obtain a first feature map includes: The initial input feature map is divided into multiple local blocks, and the pixels of each local block are transferred to the channel dimension according to their spatial location. The channel dimensions are evaluated and calibrated using an attention mechanism to obtain high-quality features; The high-quality features are subjected to feature fusion and dimensionality reduction operations to output the first feature map.

3. The method as described in claim 1, characterized in that, The step of performing salient gate calibration and feature encoding on the first feature map to generate a multi-scale fused feature map includes: The first feature map is reshaped along the channel dimension and divided into multiple channel groups; Channel weights are set for each of the channel groups, and based on the channel weights, saliency gate calibration is performed on the grouping features of each of the channel groups to obtain grouping calibration features; The grouped calibration features are subjected to bi-branch feature extraction to obtain the first branch features and the second branch features; Cross-dimensional feature aggregation is performed on the first branch features and the second branch features to obtain a fusion spatial attention map; The fused spatial attention map is non-linearly activated into an attention weight mask using the Sigmoid function; The attention weight mask is multiplied element-wise with the group calibration features to obtain a multi-scale fused feature map.

4. The method as described in claim 3, characterized in that, The step of setting channel weights for each channel group and performing salient gate calibration on the grouping features of each channel group based on the channel weights to obtain grouping calibration features includes: The grouped feature tensors are compressed in spatial dimension through a global adaptive average pooling layer, and information exchange between channels is performed through a single-kernel convolutional layer to obtain the output feature tensors of each channel group. The output feature tensor is normalized to a dynamic gating scaling factor using the Sigmoid activation function. The dynamic gating scaling factor is then multiplied element-wise with the grouping features of each channel group to obtain the grouping calibration features.

5. The method as described in claim 3, characterized in that, The process of obtaining the first branch feature includes: One-dimensional adaptive average pooling is used to encode the grouped calibration features along the width and height directions of the space, respectively, to generate width-direction sensing features and height-direction sensing features. The width direction perception feature and the height direction perception feature are spliced ​​together in the spatial dimension, and information exchange between channels is performed to obtain directional attention features; The directional attention features and the group calibration features are weighted to obtain the first branch features.

6. The method as described in claim 3, characterized in that, For the second branch feature, the grouped calibration feature is further mined using 3×3 convolution to explore the spatial context information within the local receptive field.

7. The method as described in claim 3, characterized in that, The step of performing cross-dimensional feature aggregation on the first branch features and the second branch features to obtain a fused spatial attention map includes: Global average pooling and Softmax activation are performed on the first branch features and the second branch features respectively to obtain the first branch channel attention weights and the second branch channel attention weights; Based on the attention weights of the first branch channel and the second branch channel, matrix multiplication is used to cross-multiply the first branch features and the second branch features, and the results of the cross-multiplication are added together to obtain the fusion spatial attention map.

8. The method as described in claim 1, characterized in that, The step of performing reparameterized large kernel convolution on the first feature map includes: The first feature map is input into a large-scale convolutional kernel and a small-scale convolutional kernel; In the small-scale convolutional kernel, zero padding is performed in the spatial aspect to align its size with that of the large-scale convolutional kernel, and the padded small-scale weight matrix is ​​obtained. The small-scale weight matrix and the large-scale weight matrix are added element-wise to form an equivalent large convolution kernel.

9. A forest vegetation pest and disease detection system based on unmanned aerial vehicles (UAVs), characterized in that, include: First processing module: used to collect images of forest pests and diseases through drones, and to preprocess the images of forest pests and diseases to generate an initial input feature map; The second processing module is used to perform feature enhancement on the initial input feature map to obtain the first feature map. The third processing module is used to perform salient gate calibration and feature encoding on the first feature map to generate a multi-scale fused feature map. The fourth processing module is used to perform multi-level fusion of multi-scale fusion feature maps to obtain first feature maps at different scales. The fifth processing module is used to perform reparameterized large kernel convolution on each of the first feature maps to output a target feature map, which can locate the area of ​​small pests and diseases.

Images