Intelligent Crop Disease Identification Method and System Based on UAV Imagery

Abstract

This invention discloses a method and system for intelligent crop disease identification based on UAV imagery, belonging to the field of smart agriculture technology. The method includes multi-dimensional data acquisition, early disease identification, spatiotemporal disease extrapolation, and crop management recommendations. This invention employs early disease identification based on a spectral-sensitive vegetation index and an attention-guided dual-branch network. It constructs a spectral-sensitive vegetation index that amplifies weak lesion features, highlighting weak disease characteristics through multispectral information enhancement and temporal feature suppression. Simultaneously, it utilizes an attention mechanism to focus on potential disease areas, thereby significantly improving the accuracy and stability of early disease identification. Furthermore, it employs spatiotemporal disease extrapolation combining propagation potential energy and a spatiotemporal graph convolutional network. Considering environmental conditions, disease temporal evolution, and spatial neighborhood influences, it dynamically simulates disease propagation paths, quantifies future risks, and generates intuitive risk prediction maps and propagation vector maps, thereby assisting in optimizing agricultural management.

Description

Technical Field

[0001] This invention belongs to the field of smart agriculture technology, specifically referring to a method and system for intelligent identification of crop diseases based on UAV imagery. Background Technology

[0002] Intelligent crop disease identification based on UAV imagery combines modern remote sensing and artificial intelligence technologies. By efficiently collecting large-scale, high-resolution farmland images using UAVs, it enables early and accurate detection and risk assessment of crop diseases. The aim is to improve the timeliness and accuracy of disease identification, achieve quantifiable and traceable farmland disease management, and provide technical support for efficient inspection of large-scale farmland and the digital, green, and sustainable development of agriculture.

[0003] However, existing intelligent crop disease identification processes suffer from several challenges. Early-stage disease lesion signals are weak and scattered, easily interfered with by background noise and changes in light, making timely detection difficult. Traditional methods rely heavily on single images or visible symptoms, failing to effectively capture potential signals related to spectral variations, moisture levels, and chlorophyll changes, leading to missed detections and misjudgments. Furthermore, while early disease identification can pinpoint lesion locations, it struggles to accurately predict future disease spread paths and intensity. It also fails to adequately consider the impact of environmental factors such as topography, weather, and crop growth status on disease transmission, resulting in inaccurate spread path predictions, hindering timely control measures, and potentially causing missed optimal prevention and control windows. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a method and system for intelligent crop disease identification based on UAV imagery. It creatively employs an early disease identification method based on a spectrally sensitive vegetation index and an attention-guided dual-branch network. This constructs a spectrally sensitive vegetation index capable of amplifying weak lesion features, enhancing multispectral information and suppressing temporal features to highlight subtle disease characteristics. Simultaneously, it utilizes an attention mechanism to focus on potential disease areas, thereby significantly improving the accuracy and stability of early disease identification. Furthermore, it creatively employs a disease spatiotemporal extrapolation method combining propagation potential energy and a spatiotemporal graph convolutional network. Considering environmental conditions, disease temporal evolution, and spatial neighborhood influences, it dynamically simulates disease propagation paths, quantifies future risks, and generates intuitive risk prediction maps and propagation vector maps. This provides a scientific and quantifiable decision-making basis for farmland disease control, assists in optimizing agricultural management, and significantly improves crop production safety and management efficiency.

[0005] The technical solution adopted by this invention is as follows: The intelligent crop disease identification method based on UAV imagery provided by this invention includes the following steps:

[0006] Step S1: Multidimensional data collection;

[0007] Step S2: Early disease identification;

[0008] Step S3: Spatiotemporal simulation of the disease;

[0009] Step S4: Crop Management Recommendations.

[0010] Furthermore, in step S1, the multi-dimensional data acquisition specifically involves collecting soil data and meteorological monitoring data, acquiring multi-source image data of farmland using drones, generating a digital elevation model by combining it with lidar, and then performing data preprocessing.

[0011] Further, in step S2, the early disease identification is used to achieve accurate location and diagnosis in the early stage of disease occurrence. Specifically, based on multi-source image data, early disease identification is performed using a dual-branch network based on spectral sensitive vegetation index and attention guidance to obtain a probability distribution map of early crop diseases, including the following steps:

[0012] Step S21: Multimodal feature enhancement, specifically, by introducing temporal suppression based on multi-source image data, calculating the spectral sensitive vegetation index, and performing multi-scale enhancement processing to obtain the spectral enhancement index, then by combining the spectral enhancement index with the RGB channels to construct a multi-channel input tensor, and generating an attention mask based on the normalization result of the spectral enhancement index;

[0013] Step S22: Construct a dual-branch feature extraction network to extract structural, spectral, and spatial information. Specifically, the network is encoded in parallel by a backbone feature extraction branch and an attention-guided branch. The backbone feature extraction branch uses EfficientNet-V2 to encode the multi-channel input tensor to obtain a structural semantic feature map. The attention-guided branch performs convolution and sigmoid activation on the attention mask to generate an early disease attention weight map. Then, the output size of the early disease attention weight map is aligned with the structural semantic feature map.

[0014] Step S23: Attention-guided feature fusion, which is used to fuse structural semantic information and attention information to improve the response sensitivity of early crop disease areas. Specifically, the structural semantic features are weighted and amplified element by element based on the early disease attention weight map to construct a spectral response attention map. Then, the spectral response attention map is fused with multiple layers of features through a joint optimization block of channel attention and spatial attention to obtain a spectral spatial fusion feature map.

[0015] Step S24: Disease occurrence probability assessment, specifically, extracting global features from the spectral spatial fusion feature map through global average pooling, then calculating the probability of early crop disease occurrence through a fully connected layer and Sigmoid activation, and generating a probability distribution map of early crop diseases.

[0016] Further, in step S3, the spatiotemporal disease simulation is used to dynamically simulate and predict the disease's transmission path and future risks. Specifically, based on the early crop disease probability distribution map, soil data, meteorological monitoring data, and digital elevation model, a spatiotemporal disease simulation combining propagation potential energy and a spatiotemporal graph convolutional network is used to obtain disease spatiotemporal simulation information, including the following steps:

[0017] Step S31: Disease and environment fusion modeling, used to construct a spatiotemporal input tensor reflecting the coupling effect between disease occurrence and environmental conditions. Specifically, by spatiotemporally registering the probability distribution map of early crop diseases at consecutive time points, the disease time-series vector of each pixel is obtained, and the terrain feature vector and meteorological comprehensive index are extracted. Then, the disease time-series vector, terrain feature vector and meteorological comprehensive index are concatenated to form a spatiotemporal input tensor.

[0018] Step S32: Construct a propagation graph to reflect the potential energy and directional characteristics of disease propagation. Specifically, each pixel is treated as a node. By constructing a propagation potential energy function, the propagation potential energy between any two nodes is calculated. For each node, the k nodes with the largest propagation potential energy are selected as neighboring nodes. Propagation edges are constructed to establish a disease propagation topology graph. The disease propagation adjacency matrix is ​​obtained through normalization.

[0019] The propagation potential energy function is specifically constructed by comprehensively considering vegetation activity, spatial distance, humidity difference, wind direction angle and early disease gradient;

[0020] Step S33: Predict future disease risk. Specifically, based on the spatiotemporal input tensor and the disease propagation adjacency matrix, a spatiotemporal attention graph convolutional network combining graph convolution, temporal convolution and an improved self-attention mechanism is constructed to perform spatial neighborhood aggregation and temporal evolution trend capture for each node, thereby obtaining the future disease risk of the node.

[0021] The improved self-attention mechanism specifically calculates the query, key, and value mapping only within the node's neighborhood and adjusts the attention coefficient in conjunction with the propagation weight to achieve adaptive attention to high-impact nodes.

[0022] Step S34: Risk visualization, specifically, mapping the future disease risk of each node to the farmland grid, generating a future disease risk prediction map, and calculating the disease transmission flux between nodes based on the disease transmission adjacency matrix, retaining only the top k edges with the largest disease transmission flux, generating a future disease transmission vector map for visualizing the main transmission direction and intensity.

[0023] Step S35: Generation of spatiotemporal projection information for diseases. Specifically, by executing steps S31 to S34, spatiotemporal projection information for diseases is obtained. The spatiotemporal projection information for diseases includes a future disease risk prediction map and a future disease propagation vector map.

[0024] Further, in step S4, the crop management suggestion is used to generate an executable agricultural decision-making scheme. Specifically, based on the early crop disease probability distribution map and the spatiotemporal inference information of diseases, a comprehensive risk score matrix is ​​obtained by integrating the early crop disease probability distribution map with the future disease risk prediction map. Based on the preset risk score function, the farmland grid is divided into risk levels to generate high-risk, medium-risk and low-risk areas. Then, according to the risk level, the corresponding crop management measures are automatically matched.

[0025] The intelligent crop disease identification system based on UAV imagery provided by this invention includes: a multi-dimensional data acquisition module, an early disease identification module, a disease spatiotemporal extrapolation module, and a crop management suggestion module;

[0026] The data multidimensional acquisition module is used for data multidimensional acquisition. Through data multidimensional acquisition, soil data, meteorological monitoring data, digital elevation model and multi-source image data are obtained, and the multi-source image data is sent to the early disease identification module. The soil data, meteorological monitoring data and digital elevation model are sent to the disease spatiotemporal inference module.

[0027] The early disease identification module is used for early disease identification. Through early disease identification, it obtains a probability distribution map of early crop diseases and sends the probability distribution map of early crop diseases to the disease spatiotemporal inference module and the crop management suggestion module.

[0028] The disease spatiotemporal projection module is used for disease spatiotemporal projection. Through disease spatiotemporal projection, disease spatiotemporal projection information is obtained, and the disease spatiotemporal projection information is sent to the crop management suggestion module.

[0029] The crop management suggestion module is used to provide crop management suggestions and generate crop management measures.

[0030] The beneficial effects achieved by the present invention using the above solution are as follows:

[0031] (1) In the existing intelligent identification process of crop diseases, there are problems such as weak and scattered early disease lesion signals, which are easily interfered with by background noise and light changes, making it difficult to detect diseases in time. Traditional methods rely on single images or visible symptoms, which cannot effectively capture the potential signals of diseases in terms of spectrum, water and chlorophyll changes, and the identification results are prone to missed detection and misjudgment. This solution creatively adopts early disease identification based on spectral sensitive vegetation index and attention-guided dual-branch network. It constructs a spectral sensitive vegetation index that can amplify the characteristics of weak lesions, highlights the weak features of diseases through multispectral information enhancement and temporal feature suppression, and uses the attention mechanism to focus on potential disease areas, thereby significantly improving the accuracy and stability of early disease identification.

[0032] (2) In the existing intelligent identification process for crop diseases, although early identification of diseases can determine the location of lesions, it is difficult to accurately predict the future spread path and intensity of the disease. Furthermore, it does not fully consider the influence of environmental factors such as topography, weather and crop growth status on disease spread, resulting in inaccurate prediction of disease spread path, difficulty in timely formulation of control measures, and easy to miss the best prevention and control window. This solution creatively adopts disease spatiotemporal extrapolation that combines propagation potential energy and spatiotemporal graph convolutional network. Based on considering factors such as environmental conditions, disease temporal evolution and spatial neighborhood influence, it dynamically simulates disease spread path, quantifies future risks, and generates intuitive risk prediction maps and propagation vector maps. This provides scientific and quantifiable decision-making basis for farmland disease prevention and control, assists in optimizing agricultural management, and significantly improves crop production safety and management efficiency. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the intelligent crop disease identification method based on UAV imagery provided by the present invention.

[0034] Figure 2 A schematic diagram of the modules of the intelligent crop disease identification system based on UAV imagery provided by the present invention;

[0035] Figure 3 This is a flowchart illustrating step S2;

[0036] Figure 4 This is a flowchart illustrating step S3.

[0037] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0038] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0040] Example 1, see Figure 1 The present invention provides a method for intelligent identification of crop diseases based on UAV imagery, which includes the following steps:

[0041] Step S1: Multidimensional data collection;

[0042] Step S2: Early disease identification;

[0043] Step S3: Spatiotemporal simulation of the disease;

[0044] Step S4: Crop Management Recommendations.

[0045] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the multi-dimensional data acquisition is used to obtain multi-source data covering the entire farmland area. Specifically, it involves collecting soil data and meteorological monitoring data, collecting multi-source image data of farmland through drones, generating a digital elevation model by combining it with lidar, and then performing data preprocessing.

[0046] The multi-source image data includes RGB images, multispectral images, and thermal infrared images;

[0047] The soil data includes soil moisture, soil type, and nutrient distribution;

[0048] The meteorological monitoring data includes temperature, humidity, rainfall, and wind speed;

[0049] The data preprocessing includes image registration, illumination correction, interpolation normalization, and spatiotemporal alignment.

[0050] Example 3, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S2, the early disease identification is used to achieve accurate location and diagnosis in the early stage of disease occurrence. Specifically, it uses multi-source image data and adopts early disease identification based on spectral sensitive vegetation index and attention-guided dual-branch network to obtain a probability distribution map of early crop diseases, including the following steps:

[0051] Step S21: Multimodal feature enhancement, used to fuse and enhance multi-source image data acquired by UAV, thereby highlighting the weak features of early lesions. Specifically, based on multi-source image data, temporal suppression is introduced, a spectral sensitive vegetation index is calculated, and multi-scale enhancement processing is performed to obtain a spectral enhancement index. Then, by combining the spectral enhancement index with the RGB channels, a multi-channel input tensor is constructed, and an attention mask is generated based on the normalization result of the spectral enhancement index.

[0052] The formula for calculating the spectrally sensitive vegetation index is as follows:

[0053] ;

[0054] In the formula, SVI t It is a spectrally sensitive vegetation index at time step t, MCARI t NDWI is the chlorophyll absorbance reflectance index at time step t. t It is the normalized moisture index at time step t. It is a time-series suppression weight. It is the difference in spectral-sensitive vegetation index between adjacent time points;

[0055] The spectral enhancement index, used to characterize faint lesion texture, is calculated using the following formula:

[0056] ;

[0057] In the formula, It is the spectral enhancement index at time step t. This is the enhancement factor, with a value range of [0.5, 1.2], used to control the amplification level of high-frequency components. It is a Gaussian blur operator used to preserve low-frequency information;

[0058] The generation of the attention mask is specifically achieved by normalizing the spectral enhancement index and taking the lowest 10% of the pixel regions in the normalized values ​​as candidate lesion regions.

[0059] Step S22: Construct a dual-branch feature extraction network to extract structural, spectral, and spatial information. Specifically, the network is encoded in parallel by a backbone feature extraction branch and an attention-guided branch. The backbone feature extraction branch uses EfficientNet-V2 to encode the multi-channel input tensor to obtain a structural semantic feature map. The attention-guided branch performs convolution and sigmoid activation on the attention mask to generate an early disease attention weight map. Then, the output size of the early disease attention weight map is aligned with the structural semantic feature map.

[0060] Step S23: Attention-guided feature fusion, which is used to fuse structural semantic information and attention information to improve the response sensitivity of early crop disease areas. Specifically, the structural semantic features are weighted and amplified element by element based on the early disease attention weight map to construct a spectral response attention map. Then, the spectral response attention map is fused with multiple layers of features through a joint optimization block of channel attention and spatial attention to obtain a spectral spatial fusion feature map.

[0061] The formula for constructing the spectral response attention map is as follows:

[0062] ;

[0063] In the formula, It is a spectral response attention map. M is the spectral response amplification factor, with a value range of [0.5, 1], used to control the amplification amplitude. att This is a weighted map of early disease concerns. It is the element-wise multiplication symbol, F rgb-svi It is a structural semantic feature map;

[0064] The formula for calculating the spectral spatial fusion feature map is as follows:

[0065] ;

[0066] In the formula, F final It is a spectral spatial fusion feature map, and CBAM(·) is a joint optimization block of channel attention and spatial attention;

[0067] The joint optimization block of channel attention and spatial attention includes a channel attention block and a spatial attention block. Specifically, the channel attention block extracts channel statistics through parallel global average pooling and global max pooling, generates a channel attention weight map through a shared multilayer perceptron and sigmoid activation, and then performs channel weighting to obtain a channel weighted feature map. Specifically, the spatial attention block performs global average pooling and global max pooling on the channel weighted feature map, extracts spatial context relationships through 7×7 convolution operation, generates a spatial attention weight map through sigmoid activation, and then multiplies the spatial attention weight map and the channel weighted feature map element-wise to obtain a spectral spatial fusion feature map.

[0068] The formula for calculating the channel attention weight map is as follows:

[0069] ;

[0070] In the formula, M c It is the channel attention weight map, Sig(·) is the Sigmoid activation function, MLP(·) is the multilayer perceptron, AvgPool(·) is the global average pooling, and MaxPool(·) is the global max pooling.

[0071] The formula for calculating the channel-weighted feature map is:

[0072] ;

[0073] In the formula, It is a channel-weighted feature map;

[0074] The formula for calculating the spatial attention weight map is as follows:

[0075] ;

[0076] In the formula, M s It is a spatial attention weight map, Conv 7×7 (·) is a 7×7 convolution operation, and [·,·] is a concatenation operation;

[0077] Step S24: Disease occurrence probability assessment, specifically, global features of the spectral spatial fusion feature map are extracted by global average pooling, and then the probability of early crop disease occurrence is calculated by using a fully connected layer and Sigmoid activation, and a probability distribution map of early crop disease is generated.

[0078] The formula for calculating the probability of early disease occurrence is as follows:

[0079] ;

[0080] In the formula, P disease is the probability of early disease occurrence, w is the weight of the fully connected layer, and b is the bias term of the fully connected layer.

[0081] By performing the above operations, this solution addresses the technical problems in existing intelligent crop disease identification processes, such as weak and scattered early-stage disease lesion signals, which are easily interfered with by background noise and changes in light, making it difficult to detect diseases in a timely manner. Traditional methods often rely on single images or visible symptoms, failing to effectively capture potential signals of disease in terms of spectral, moisture, and chlorophyll changes, leading to missed detections and misjudgments. This solution creatively adopts an early disease identification method based on a spectrally sensitive vegetation index and an attention-guided dual-branch network. It constructs a spectrally sensitive vegetation index that amplifies weak lesion features, highlights weak disease features through multispectral information enhancement and temporal feature suppression, and focuses on potential disease areas using an attention mechanism, thereby significantly improving the accuracy and stability of early disease identification.

[0082] Example 4, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S3, the spatiotemporal extrapolation of the disease is used to dynamically simulate and predict the spread path and future risks of the disease. Specifically, based on the early crop disease probability distribution map, soil data, meteorological monitoring data, and digital elevation model, spatiotemporal extrapolation of the disease combining propagation potential energy and spatiotemporal graph convolutional network is used to obtain spatiotemporal extrapolation information of the disease, including the following steps:

[0083] Step S31: Disease and environment fusion modeling, used to construct a spatiotemporal input tensor reflecting the coupling effect between disease occurrence and environmental conditions. Specifically, by spatiotemporally registering the probability distribution map of early crop diseases at consecutive time points, the disease time-series vector of each pixel is obtained, and the terrain feature vector and meteorological comprehensive index are extracted. Then, the disease time-series vector, terrain feature vector and meteorological comprehensive index are concatenated to form a spatiotemporal input tensor.

[0084] The terrain feature vector is specifically based on the digital elevation model, which calculates the slope, aspect and terrain humidity index of each pixel and then splices them together to form the terrain feature vector.

[0085] The meteorological comprehensive index is specifically based on meteorological monitoring data, which includes the temperature, humidity, rainfall, and wind speed of each pixel. These data are then weighted and fused to obtain the meteorological comprehensive index. The calculation formula is as follows:

[0086] ;

[0087] In the formula, Env is the meteorological comprehensive index, w1, w2, w3 and w4 are environmental sensitivity weights, Temp is temperature, Humid is humidity, Rain is rainfall, and Wind is wind speed.

[0088] Step S32: Construct a propagation graph to reflect the potential energy and directional characteristics of disease propagation. Specifically, each pixel is treated as a node. By constructing a propagation potential energy function, the propagation potential energy between any two nodes is calculated. For each node, the k nodes with the largest propagation potential energy are selected as neighboring nodes. Propagation edges are constructed to establish a disease propagation topology graph. The disease propagation adjacency matrix is ​​obtained through normalization.

[0089] The propagation potential energy function is specifically constructed by comprehensively considering vegetation activity, spatial distance, humidity difference, wind direction angle, and early disease gradient. The calculation formula is as follows:

[0090] ;

[0091] In the formula, A is the propagation potential energy between node i and node j, where i is the first index of node i and j is the second index of node j, i ≠ j. i A represents the vegetation activity at node i. j Let be the vegetation activity at node j, and e be the base of the natural logarithm. It is the humidity attenuation coefficient, TWI i It is the terrain humidity index (TWI) of node i. j It is the terrain humidity index of node j. It is the wind propagation enhancement coefficient. It is the angle between the direction of the line connecting node i and node j and the prevailing wind direction. It is the early disease gradient;

[0092] Preferably, the formula for calculating the vegetation activity is:

[0093] ;

[0094] In the formula, It is a weighting factor, NDVI i It is the normalized vegetation index (SVI) of node i. i It is the soil vegetation index of node i;

[0095] The formula for calculating the early disease gradient is as follows:

[0096] ;

[0097] In the formula, It is the gradient sensitivity coefficient. It is the difference in the probability of early disease occurrence between node i and node j;

[0098] The formula for calculating the disease propagation adjacency matrix is ​​as follows:

[0099] ;

[0100] In the formula, W ij It is the propagation weight between node i and node j, specifically the elements of the disease propagation adjacency matrix, where N(i) is the set of neighboring nodes of node i.

[0101] Step S33: Predict future disease risk. Specifically, based on the spatiotemporal input tensor and the disease propagation adjacency matrix, a spatiotemporal attention graph convolutional network combining graph convolution, temporal convolution and an improved self-attention mechanism is constructed to perform spatial neighborhood aggregation and temporal evolution trend capture for each node, thereby obtaining the future disease risk of the node.

[0102] The improved self-attention mechanism specifically calculates the query, key, and value mapping only within the node's neighborhood and adjusts the attention coefficient based on the propagation weight to achieve adaptive attention to high-influence nodes. The calculation formula is as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] In the formula, a ij It represents the attention weight of node i to node j, exp(·) is the natural exponential function, T is the transpose operation, and Q is the expression for the attention weight of node i to node j. i K is the query vector corresponding to node i. jIt is the key vector corresponding to node j, z is the third index of node, and K z W is the key vector corresponding to node z. iz It is the propagation weight between node i and node z. W is the feature vector of node i in the 1st layer (lay), where lay is the network layer index. Q It is the query mapping weight, W K It is the key-map weight. W is the feature vector of node j in the 1st layer. V Value mapping weights, Att(·) is a self-attention mapping;

[0107] Step S34: Risk visualization, specifically, mapping the future disease risk of each node to the farmland grid, generating a future disease risk prediction map, and calculating the disease transmission flux between nodes based on the disease transmission adjacency matrix, retaining only the top k edges with the largest disease transmission flux, generating a future disease transmission vector map for visualizing the main transmission direction and intensity.

[0108] The formula for calculating the disease transmission flux is:

[0109] ;

[0110] In the formula, F ij Let $\frac{i}{j}$ be the disease propagation flux from node $i$ to node $j$, and $max(·)$ be the maximum value function used to preserve the positive gradient. It is the difference in future disease risk between node i and node j;

[0111] The generation of the future disease transmission vector map specifically involves generating the disease transmission flux F. ij The top k largest edges are drawn as arrows, with the starting point of the arrow being node i and the ending point being node j. The length of the arrow is drawn according to the proportion of the disease transmission flux. The arrows are superimposed on the geographic information system platform to visualize the main transmission paths of the disease.

[0112] Step S35: Generation of spatiotemporal projection information for diseases. Specifically, by executing steps S31 to S34, spatiotemporal projection information for diseases is obtained. The spatiotemporal projection information for diseases includes a future disease risk prediction map and a future disease propagation vector map.

[0113] By performing the above operations, this solution addresses the technical problems in existing intelligent crop disease identification processes. While early disease identification can pinpoint lesion locations, it struggles to accurately predict future disease spread paths and intensity. Furthermore, it fails to adequately consider the impact of environmental factors such as topography, weather, and crop growth status on disease transmission, leading to inaccurate disease transmission path predictions, difficulty in timely implementation of control measures, and a risk of missing the optimal prevention and control window. This solution creatively employs a disease spatiotemporal extrapolation method combining propagation potential energy and spatiotemporal graph convolutional networks. By considering environmental conditions, disease temporal evolution, and spatial neighborhood influences, it dynamically simulates disease transmission paths, quantifies future risks, and generates intuitive risk prediction maps and propagation vector maps. This provides a scientific and quantifiable basis for farmland disease control, assists in optimizing agricultural management, and significantly improves crop production safety and management efficiency.

[0114] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the crop management suggestion is used to generate an executable agricultural decision-making plan. Specifically, based on the early crop disease probability distribution map and disease spatiotemporal projection information, the early crop disease probability distribution map and the future disease risk prediction map are fused and calculated to obtain a comprehensive risk score matrix. Based on the preset risk score function, the farmland grid is divided into risk levels to generate high-risk, medium-risk and low-risk areas. Then, according to the risk level, the corresponding crop management measures are automatically matched.

[0115] The fusion calculation can be implemented using methods including but not limited to linear weighting, nonlinear functions, maximum value selection, and minimum value selection, and the comprehensive risk score matrix can be adjusted by referring to the main propagation paths recorded in the future disease propagation vector map.

[0116] The preset risk scoring function includes, but is not limited to, threshold segmentation, quantile division, and nonlinear mapping function;

[0117] The automatically matched crop management measures specifically include providing emergency response suggestions for high-risk areas, preventive management suggestions for medium-risk areas, and growth maintenance suggestions for low-risk areas.

[0118] The emergency response recommendations include, but are not limited to, precise spraying of designated types of pesticides, enhanced local ventilation, and immediate drainage; the preventative management recommendations include, but are not limited to, low-dose pesticide spraying, localized mulching, improved irrigation uniformity, and application of biological control agents; the growth maintenance recommendations include, but are not limited to, maintaining routine water and fertilizer management, regular inspections, and enhanced light control.

[0119] The crop management measures can be expanded according to crop type, growth period and meteorological conditions.

[0120] Example 6, see Figure 2 Based on the above embodiments, the intelligent crop disease identification system based on UAV imagery provided by the present invention includes: a multi-dimensional data acquisition module, an early disease identification module, a disease spatiotemporal extrapolation module, and a crop management suggestion module;

[0121] The data multidimensional acquisition module is used for data multidimensional acquisition. Through data multidimensional acquisition, soil data, meteorological monitoring data, digital elevation model and multi-source image data are obtained, and the multi-source image data is sent to the early disease identification module. The soil data, meteorological monitoring data and digital elevation model are sent to the disease spatiotemporal inference module.

[0122] The early disease identification module is used for early disease identification. Through early disease identification, it obtains a probability distribution map of early crop diseases and sends the probability distribution map of early crop diseases to the disease spatiotemporal inference module and the crop management suggestion module.

[0123] The disease spatiotemporal projection module is used for disease spatiotemporal projection. Through disease spatiotemporal projection, disease spatiotemporal projection information is obtained, and the disease spatiotemporal projection information is sent to the crop management suggestion module.

[0124] The crop management suggestion module is used to provide crop management suggestions and generate crop management measures.

[0125] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0126] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0127] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for intelligent identification of crop diseases based on UAV imagery, characterized in that: The method includes the following steps: Step S1: Multidimensional data acquisition to obtain soil data, meteorological monitoring data, digital elevation model and multi-source image data; Step S2: Early disease identification, used to achieve accurate location and diagnosis in the early stage of disease occurrence. Specifically, based on multi-source image data, early disease identification is carried out using a dual-branch network based on spectral sensitive vegetation index and attention-guided to obtain a probability distribution map of early crop diseases. This includes the following steps: Step S21 Multimodal feature enhancement, Step S22 Dual-branch feature extraction network construction, Step S23 Attention-guided feature fusion, and Step S24 Disease occurrence probability assessment. In step S21, the multimodal feature enhancement specifically involves introducing temporal suppression based on multi-source image data, calculating the spectral-sensitive vegetation index, performing multi-scale enhancement processing to obtain the spectral enhancement index, and then constructing a multi-channel input tensor by combining the spectral enhancement index with the RGB channels, and generating an attention mask based on the normalization result of the spectral enhancement index. In step S23, the attention-guided feature fusion is used to fuse structural semantic information and attention information to improve the response sensitivity of early crop disease areas. Specifically, the structural semantic features are weighted and amplified element by element based on the early disease attention weight map to construct a spectral response attention map. Then, the spectral response attention map is fused with multiple layers of features through a joint optimization block of channel attention and spatial attention to obtain a spectral spatial fusion feature map. Step S3: Spatiotemporal simulation of disease, used to dynamically simulate and predict the spread path and future risks of disease. Specifically, based on the early disease probability distribution map of crops, soil data, meteorological monitoring data and digital elevation model, spatiotemporal simulation of disease is carried out by combining propagation potential energy and spatiotemporal graph convolutional network to obtain spatiotemporal simulation information of disease. The steps include: Step S31 Disease environment fusion modeling, Step S32 Propagation graph construction, Step S33 Future disease risk prediction, Step S34 Risk visualization and Step S35 Generation of spatiotemporal simulation information of disease. In step S32, the propagation graph is constructed to reflect the potential energy and directional characteristics of disease propagation. Specifically, each pixel is treated as a node, and the propagation potential energy between any two nodes is calculated by constructing a propagation potential energy function. For each node, the k nodes with the largest propagation potential energy are selected as neighboring nodes, propagation edges are constructed, a disease propagation topology graph is established, and a disease propagation adjacency matrix is ​​obtained through normalization. The propagation potential energy function is specifically constructed by comprehensively considering vegetation activity, spatial distance, humidity difference, wind direction angle and early disease gradient; Step S4: Crop Management Recommendations.

2. The intelligent crop disease identification method based on UAV imagery according to claim 1, characterized in that: In step S22, the dual-branch feature extraction network is constructed to extract structural, spectral, and spatial information. Specifically, it performs parallel encoding through a backbone feature extraction branch and an attention-guided branch. The backbone feature extraction branch uses EfficientNet-V2 to encode the multi-channel input tensor to obtain a structural semantic feature map. The attention-guided branch performs convolution and sigmoid activation on the attention mask to generate an early disease attention weight map. Then, the output size of the early disease attention weight map is aligned with the structural semantic feature map.

3. The intelligent crop disease identification method based on UAV imagery according to claim 2, characterized in that: In step S24, the disease occurrence probability assessment specifically involves extracting global features from the spectral spatial fusion feature map through global average pooling, then calculating the early crop disease occurrence probability through a fully connected layer and Sigmoid activation, and generating an early crop disease probability distribution map.

4. The intelligent crop disease identification method based on UAV imagery according to claim 3, characterized in that: In step S31, the disease environment fusion modeling is used to construct a spatiotemporal input tensor that reflects the coupling effect between disease occurrence and environmental conditions. Specifically, it involves spatiotemporally registering the probability distribution map of early crop diseases at consecutive time points to obtain the disease time-series vector of each pixel, extracting the terrain feature vector and the meteorological comprehensive index, and then concatenating the disease time-series vector, the terrain feature vector and the meteorological comprehensive index to form a spatiotemporal input tensor. In step S33, the prediction of future disease risk specifically involves constructing a spatiotemporal attention graph convolutional network that combines graph convolution, temporal convolution, and an improved self-attention mechanism, based on the spatiotemporal input tensor and the disease propagation adjacency matrix, to perform spatial neighborhood aggregation and temporal evolution trend capture for each node, thereby obtaining the future disease risk of the node. The improved self-attention mechanism specifically calculates the query, key, and value mapping only within the node's neighborhood and adjusts the attention coefficient in conjunction with the propagation weight to achieve adaptive attention to high-impact nodes.

5. The intelligent crop disease identification method based on UAV imagery according to claim 4, characterized in that: In step S34, the risk visualization specifically involves mapping the future disease risk of each node to the farmland grid, generating a future disease risk prediction map, and calculating the disease propagation flux between nodes based on the disease propagation adjacency matrix. Only the top k edges with the largest disease propagation flux are retained to generate a future disease propagation vector map for visualizing the main propagation direction and intensity. In step S35, the disease spatiotemporal projection information is generated, specifically by executing steps S31 to S34 to obtain the disease spatiotemporal projection information, which includes a future disease risk prediction map and a future disease propagation vector map.

6. The intelligent crop disease identification method based on UAV imagery according to claim 5, characterized in that: In step S4, the crop management suggestion is used to generate an executable agricultural decision-making plan. Specifically, based on the early crop disease probability distribution map and disease spatiotemporal projection information, the early crop disease probability distribution map and the future disease risk prediction map are fused to obtain a comprehensive risk score matrix. Based on the preset risk score function, the farmland grid is divided into risk levels to generate high-risk, medium-risk and low-risk areas. Then, according to the risk level, the corresponding crop management measures are automatically matched. In step S1, the multi-dimensional data acquisition specifically involves collecting soil data and meteorological monitoring data, acquiring multi-source image data of farmland using drones, generating a digital elevation model by combining it with lidar, and then performing data preprocessing.

7. A crop disease intelligent identification system based on UAV imagery, used to implement the crop disease intelligent identification method based on UAV imagery as described in any one of claims 1-6, characterized in that: It includes a multi-dimensional data acquisition module, an early disease identification module, a disease spatiotemporal projection module, and a crop management suggestion module.

8. The intelligent crop disease identification system based on UAV imagery according to claim 7, characterized in that: The data multidimensional acquisition module is used for data multidimensional acquisition. Through data multidimensional acquisition, soil data, meteorological monitoring data, digital elevation model and multi-source image data are obtained, and the multi-source image data is sent to the early disease identification module. The soil data, meteorological monitoring data and digital elevation model are sent to the disease spatiotemporal inference module. The early disease identification module is used for early disease identification. Through early disease identification, it obtains a probability distribution map of early crop diseases and sends the probability distribution map of early crop diseases to the disease spatiotemporal inference module and the crop management suggestion module. The disease spatiotemporal projection module is used for disease spatiotemporal projection. Through disease spatiotemporal projection, disease spatiotemporal projection information is obtained, and the disease spatiotemporal projection information is sent to the crop management suggestion module. The crop management suggestion module is used to provide crop management suggestions and generate crop management measures.

Images