Tumorous stem mustard disease identification method and system based on artificial intelligence

Abstract

The invention discloses a tumorous stem mustard disease identification method and system based on artificial intelligence, and relates to the technical field of tumorous stem mustard disease identification, and the method comprises five steps of image sensing acquisition, multi-source data fusion, space-time disease perception, environmental pathogenesis association and disease identification. The method comprises the following steps: constructing multi-modal optimization data by collecting a visible light image, an infrared thermal image and environment sensing data; an infrared correction mechanism is introduced, and a multi-branch network is constructed to realize space-time perception of a disease area; a disease type driven environmental factor clustering analysis method is adopted to realize disease cause tracing and dominant factor extraction; and finally, disease spot features, spatio-temporal information and environmental data are fused to realize disease classification and identification. Compared with a traditional method, the method has higher recognition precision, higher interpretability and better environmental adaptability, and is suitable for intelligent disease monitoring and agricultural decision support of crops such as tumorous stem mustard.

Description

Technical Field

[0001] This invention relates to the field of stem mustard disease identification technology, specifically to a stem mustard disease identification method and system based on artificial intelligence. Background Technology

[0002] The AI-based method and system for identifying diseases in stem mustard refers to the use of AI technologies such as deep learning and computer vision to collect image data of leaves, stems, and other parts of stem mustard plants, construct disease feature recognition models, and achieve automated detection and classification of common diseases such as black rot and clubroot. This system typically consists of an image acquisition module, a data preprocessing module, an AI model analysis module, and a decision output module. Its core functions are: 1) improving disease identification efficiency and overcoming the subjectivity and lag of traditional manual inspections; 2) providing early and accurate diagnosis of disease types, offering a scientific basis for precise pesticide application; and 3) enabling real-time field monitoring via mobile devices or IoT devices, assisting in intelligent decision-making in agricultural production, ultimately reducing pesticide overuse and crop losses.

[0003] However, existing methods for identifying stem mustard diseases have technical problems such as low accuracy in identifying diseased areas, lack of systematic analysis of the causes of disease, and one-sided identification due to reliance on single-point image data.

[0004] Existing spatiotemporal location and perception methods for diseases have technical problems such as insufficient utilization of infrared information, lack of dynamic attention capability for abnormal lesion areas, and inability to distinguish the evolution trend of lesions in adjacent time or space.

[0005] Existing methods for analyzing the environmental causes of diseases suffer from technical problems such as the disconnect between disease types and environmental factors, the inability to extract key dominant environmental factors, and the lack of practicality and controllability in the analysis results.

[0006] Existing methods for comprehensive disease identification suffer from technical problems such as fragmented identification processes and a lack of joint modeling of time, space, and environment, ultimately resulting in identification results lacking global reference value. Summary of the Invention

[0007] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an artificial intelligence-based method and system for identifying stem mustard diseases. Addressing the problems of low accuracy in identifying diseased areas, lack of systematic analysis of pathogens, and one-sided identification due to reliance on single-point image data in existing methods, this solution creatively adopts a comprehensive disease identification approach that combines spatiotemporal disease perception with environmental pathogen correlation. Through multi-source fusion of main images, infrared thermal imaging, and environmental data, and by introducing spatiotemporal tags and pathogen clustering analysis mechanisms, it achieves accurate identification of disease occurrence areas, development trends, and dominant inducing factors. This approach, combined with causal analysis, improves the interpretability and comprehensive identification accuracy of stem mustard diseases. Addressing the technical limitations of existing spatiotemporal disease localization and sensing methods—such as insufficient utilization of infrared information, lack of dynamic attention to anomalous lesion regions, and inability to distinguish disease evolution trends in adjacent time or space—this solution creatively employs a multimodal spatiotemporal sensing method with infrared correction. Through infrared-guided channel attention mechanisms and local enhancement techniques using hotspot masks, combined with an improved texture attention module and multi-scale convolutional structure in the backbone network, spatiotemporal disease sensing is achieved, enabling the detection of lesion evolution paths, spatial aggregation trends, and thermal anomalies. High-precision regional identification effectively enhances the sensitivity and generalization ability of early disease identification. Addressing the technical problems of existing disease environmental causation analysis methods, such as the decoupling of disease type from environmental factors and the inability to extract key dominant environmental factors, resulting in a lack of practicality and controllability in the analysis results, this solution creatively adopts a disease type-driven improved clustering selection method for main environmental factors to identify environmental causal associations. By introducing lesion manifestation weights and a weighted F-value dimensionality reduction mechanism, combined with an improved K-Means clustering algorithm, environmental causal association identification is achieved, realizing automatic attribution and association of key inducing factors under different disease types, thus providing a basis for agricultural management. This provides reliable data support for precision prevention and control. Addressing the technical problem that existing integrated disease identification methods suffer from fragmented identification processes and a lack of joint modeling of spatiotemporal and environmental factors, ultimately resulting in a lack of global reference value in the identification results, this solution creatively adopts an integrated identification process that combines spatiotemporal location, environmental disease correlation, and infrared image analysis. It unifies the modeling and classification of image features, thermal imaging information, spatial trajectories, and environmental factors, ultimately achieving comprehensive identification and output of disease types, lesion masks, infrared response intensity, and environmental factor sets. This improves the practical adaptability of the intelligent identification system and its ability to guide precision agricultural prevention and control.

[0008] The technical solution adopted by this invention is as follows: The method for identifying stem mustard diseases based on artificial intelligence provided by this invention includes the following steps:

[0009] Step S1: Image sensing and acquisition;

[0010] Step S2: Data fusion processing;

[0011] Step S3: Spatiotemporal disease perception;

[0012] Step S4: Environmental etiological association;

[0013] Step S5: Identification of stem mustard diseases.

[0014] Further, in step S1, the image sensing acquisition is used to acquire visual image information of the stem mustard plant, specifically by acquiring raw stem mustard data through multi-source image data acquisition;

[0015] The multi-source image data acquisition includes main image source acquisition, auxiliary sensor information acquisition, and recognition-enhanced image acquisition; the raw data of *Strombyx mori* includes visible light images, basic environmental data, and infrared image data.

[0016] Further, in step S2, the data fusion processing is used to preprocess and fuse the acquired multi-source image data. Specifically, based on the original *Strombyx mori* data, optimized *Strombyx mori* data is obtained through image processing and sensor data encoding, including the following steps:

[0017] Step S21: Main image source preprocessing, specifically, performing image size unification, color space correction and brightness normalization on the visible light image in the original data of the stem mustard to obtain preprocessed image data, and performing plant region extraction operation based on the preprocessed image data, and obtaining the main image source features through image tensor construction;

[0018] Step S22: Auxiliary sensor data synchronization encoding, specifically, performing outlier removal, timestamp standardization, and data encoding vectorization operations on the basic environmental data in the original data of the stem mustard to obtain sensor feature data;

[0019] The sensor feature data specifically includes time features, latitude and longitude location features, air humidity features, air temperature features, soil features, weather features, wind speed features, risk features, and human intervention event features;

[0020] Step S23: Feature fusion encoding, specifically, based on the main image source features and the sensor feature data, feature concatenation is used to obtain fused feature encoding data;

[0021] Step S24: Infrared image archiving, specifically, the infrared image data in the original data of the stem mustard is archived and saved as structured data to obtain the infrared image archive data;

[0022] Step S25: Data fusion processing, specifically, the fused feature encoding data and the infrared image archive data are stored in a structured manner to obtain the stem mustard optimized data;

[0023] The optimized data for *Strombyx mori* specifically includes optimized main image source features, optimized sensor features, and infrared images.

[0024] Further, in step S3, the spatiotemporal disease perception is used to combine the temporal and spatial distribution information of the image and perceive the disease area. Specifically, based on the optimized data of the stem mustard, a multimodal spatiotemporal perception method with infrared correction is used to perform spatiotemporal disease perception and obtain reference data for the spatiotemporally perceived disease area, including the following steps:

[0025] Step S31: Infrared feature branch construction, specifically, based on the infrared image in the optimized data of the stem mustard, the original infrared image is normalized to obtain a normalized infrared image feature map, and based on the normalized infrared image feature map, a single-channel convolutional encoder is constructed to extract the potential region features of the infrared hot spot to obtain the potential feature map data of the hot spot.

[0026] Step S32: Spatiotemporal main map coding branch construction, specifically, constructing an improved backbone network and extracting features based on the optimized main map source features in the optimized stem mustard data to obtain map source feature data, and performing sinusoidal position coding based on the time features and latitude and longitude position features in the optimized sensing features to obtain spatiotemporal coding feature data, and fusing the map source feature data and the spatiotemporal coding feature data to obtain fused spatiotemporal feature data;

[0027] The sinusoidal position coding adopts the sinusoidal position coding form in the standard transformer structure for spatiotemporal coding.

[0028] The improved backbone network includes a fine-grained perception module, a multi-scale convolution extraction module, a texture attention module, and a feature aggregation module;

[0029] The fine-grained perception module is used to extract low-level local texture features and retain detailed information of the diseased area;

[0030] The multi-scale convolution extraction module is used to capture features of lesions of different sizes;

[0031] The texture attention module is used to focus on areas of abnormal texture in lesions and improve the perception of lesion edges and color gradient areas.

[0032] The feature aggregation module is used to aggregate disease trend information;

[0033] Step S33: Feature fusion improvement, specifically, based on the hot spot potential feature map data, through the construction of a feature fusion layer, the attention feature weight of the infrared guidance channel is calculated, and based on the attention feature weight of the infrared guidance channel, the spatiotemporal feature data and the hot spot potential feature map data are dynamically weighted and fused to obtain infrared fused feature data. Then, by introducing a hot spot mask, the infrared fused feature data is locally enhanced to obtain infrared corrected fused feature data.

[0034] Step S34: Disease area identification, specifically, based on the infrared correction and fusion feature data, a lightweight decoder is constructed to perform disease area identification decoding to obtain disease area mask data, and spatiotemporal disease area identification data is obtained by outputting disease spatiotemporal features.

[0035] Step S35: Spatiotemporal disease perception, specifically, a single-channel convolutional encoder is constructed through the infrared feature branch, and an improved backbone network and transformer coding structure are constructed through the spatiotemporal main graph coding branch. A feature fusion layer is obtained through feature fusion improvement, and a decoder structure is obtained through disease area identification. Based on the single-channel convolutional encoder, the improved backbone network and transformer coding structure, the feature fusion layer, and the lightweight decoder, a spatiotemporal disease perception model is trained to obtain a spatiotemporal disease perception model. Using the spatiotemporal disease perception model and the stem mustard optimized data, spatiotemporal disease perception is performed to obtain spatiotemporal perception disease area reference data.

[0036] The spatiotemporal perception disease area reference data specifically includes a binary disease mask, infrared guidance channel attention feature weights, spatiotemporal annotation point set, and spatiotemporal disease perception visualization reference data.

[0037] Further, in step S4, the environmental etiology association identification is used to perform correlation analysis between disease manifestations and environmental factors. Specifically, based on the spatiotemporally perceived disease area reference data and the optimized data of stem mustard, an improved clustering selection method driven by disease type is used to identify environmental etiology associations and obtain environmental etiology association feature data, including the following steps:

[0038] Step S41: Disease type determination, specifically, based on the spatiotemporal perception disease area reference data, visual texture features, infrared thermal features and spatial statistical features are extracted sequentially to obtain disease feature data, and a standard multilayer perceptron model is trained to determine the disease type based on the disease feature data to obtain disease type classification data;

[0039] The specific types of diseases mentioned include clubroot, white rust, black spot, downy mildew, and anthracnose.

[0040] Step S42: Disease environment data matching, specifically, from the disease type classification data, the predicted type, spatial information, timestamp information, and binary mask of the lesion area for each disease area are obtained, and sensor feature data is extracted from the optimized sensor features to obtain the environmental data index; based on the environmental data index, a joint distance metric method based on spatial distance and temporal difference is used to match each disease area with the environmental record that is closest to it in time and space, and the disease environment matching data is obtained by outputting structured disease-environment matching pairs;

[0041] The disease-environment matching data specifically includes disease type, spatial location, time, and matching environmental factor vectors;

[0042] Step S43: Improve clustering selection of dominant environmental factors. Based on the disease environment matching data, extract the sensor feature data corresponding to each disease area as an environmental factor vector, and extract the lesion mask area and infrared response intensity information. By calculating the lesion performance weight of each disease sample, obtain the lesion severity weight data. Based on the lesion performance weight, construct a standard weighted F value to calculate the weighted discrimination ability of each environmental factor vector for different disease types, obtain the dimensionality-reduced environmental factor parameters, and based on the dimensionality-reduced environmental factor parameters, use a clustering method that introduces lesion performance weight to perform cluster analysis on the disease samples to obtain the dominant environmental factor disease mapping data.

[0043] The clustering method that introduces lesion manifestation weights specifically refers to an improvement on the standard K-Means clustering algorithm based on the lesion manifestation weights.

[0044] In the formula, J is the clustering objective function value, N is the total number of disease samples, i is the disease sample index, and w i It is the weight of lesion manifestation, e i These are the dimension-reduced environmental factor parameters corresponding to the diseased samples. It is the zth i The cluster center vectors corresponding to each cluster category, where z i It is the cluster category number;

[0045] Step S44: Environmental etiology association identification, specifically, based on the dominant environmental factor disease mapping data, extracting the most significant environmental inducing factors for each disease type to obtain environmental etiology association feature data.

[0046] Further, in step S5, the stem mustard disease identification is used to identify and classify diseases by combining disease area and disease association information. Specifically, based on the environmental etiology association feature data and the spatiotemporal perception disease area reference data, the disease type classification results, lesion spatial distribution information and dominant environmental factor analysis results are integrated and classified to obtain disease identification reference data.

[0047] The disease identification reference data specifically includes disease type labels, disease area masks, disease spatiotemporal information, infrared response intensity, and a set of dominant environmental factors.

[0048] The stem mustard disease identification system based on artificial intelligence provided by the present invention includes a data collection module, a fusion processing module, a disease perception module, a disease cause association module, and a disease identification module;

[0049] The data collection module is used for image sensing acquisition. Through image sensing acquisition, it obtains raw data of stem mustard and sends the raw data of stem mustard to the fusion processing module.

[0050] The fusion processing module is used for data fusion processing. Through data fusion processing, optimized data of stem mustard is obtained, and the optimized data of stem mustard is sent to the disease perception module and the cause association module.

[0051] The disease perception module is used for spatiotemporal disease perception. Through spatiotemporal disease perception, it obtains spatiotemporal perception disease area reference data and sends the spatiotemporal perception disease area reference data to the disease cause association module and the disease identification module.

[0052] The etiology association module is used for environmental etiology association. Through environmental etiology association, environmental etiology association feature data is obtained, and the environmental etiology association feature data is sent to the disease identification module.

[0053] The disease identification module is used for identifying stem mustard diseases and obtaining disease identification reference data through stem mustard disease identification.

[0054] The beneficial effects achieved by the present invention using the above solution are as follows:

[0055] (1) In view of the technical problems in the existing stem mustard disease identification methods, such as low accuracy of disease area identification, lack of systematic analysis support for the cause, and one-sided identification due to reliance on single-time point image data, this solution creatively adopts the overall design of a comprehensive disease identification approach that combines spatiotemporal disease perception and environmental cause correlation. Through multi-source fusion of main image, infrared thermal imaging and environmental data, and the introduction of spatiotemporal tags and cause clustering analysis mechanism, the solution achieves accurate identification and cause analysis of disease occurrence area, development trend and dominant inducing factors, thereby improving the interpretability and comprehensive identification accuracy of stem mustard disease.

[0056] (2) In view of the technical problems in the existing spatiotemporal localization and perception methods for diseases, such as insufficient utilization of infrared information, lack of dynamic attention to abnormal lesion areas, and inability to distinguish the evolution trend of lesions in adjacent time or space, this solution creatively adopts a multimodal spatiotemporal perception method with infrared correction to perform spatiotemporal disease perception. Through the infrared-guided channel attention mechanism and hot spot mask local enhancement technology, combined with the improved texture attention module and multi-scale convolution structure of the backbone network, spatiotemporal disease perception is performed, realizing high-precision identification of lesion evolution path, spatial aggregation trend and thermal anomaly area, effectively enhancing the sensitivity and generalization ability of early disease identification.

[0057] (3) In view of the technical problems in the existing disease environmental cause analysis methods, such as the decoupling of disease type and environmental factors and the inability to extract key dominant environmental factors, the analysis results lack practicality and controllability, this scheme creatively adopts the disease type-driven main environmental factor improved cluster selection method to identify environmental cause associations. By introducing lesion manifestation weight and weighted F value dimensionality reduction mechanism, and combined with the improved K-Means clustering algorithm, environmental cause associations are identified, realizing automatic attribution and association of key causes under different disease types, providing reliable data support for agricultural management and precision prevention and control.

[0058] (4) In view of the technical problem that existing disease comprehensive identification methods have a fragmented identification process and lack joint modeling of time, space and environment, which ultimately leads to the lack of global reference value of the identification results, this solution creatively adopts a comprehensive identification process that combines time and space location, environmental disease correlation and infrared image analysis. It unifies the modeling and classification of image features, thermal imaging information, spatial trajectory and environmental factors, and finally realizes the comprehensive identification output of disease type, lesion mask, infrared response intensity and environmental factor set, which improves the practical adaptability of intelligent identification system and its ability to guide precise agricultural prevention and control. Attached Figure Description

[0059] Figure 1 A flowchart illustrating the artificial intelligence-based stem mustard disease identification method provided by this invention;

[0060] Figure 2 A schematic diagram of the stem mustard disease identification system based on artificial intelligence provided by the present invention;

[0061] Figure 3 This is a flowchart illustrating the data fusion process in step S2.

[0062] Figure 4 This is a flowchart illustrating the process of spatiotemporal disease perception in step S3.

[0063] Figure 5 This is a flowchart illustrating the process of linking environmental causes in step S4.

[0064] 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

[0065] 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.

[0066] 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.

[0067] Example 1, see Figure 1 The present invention provides an artificial intelligence-based method for identifying stem mustard diseases, which includes the following steps:

[0068] Step S1: Image sensing and acquisition;

[0069] Step S2: Data fusion processing;

[0070] Step S3: Spatiotemporal disease perception;

[0071] Step S4: Environmental etiological association;

[0072] Step S5: Identification of stem mustard diseases.

[0073] By performing the above operations, this solution addresses the technical problems of low accuracy in identifying diseased areas, lack of systematic analysis of causes, and one-sided identification due to reliance on single-point image data in existing stem mustard disease identification methods. It creatively adopts a comprehensive disease identification approach that combines spatiotemporal disease perception and environmental cause correlation. Through multi-source fusion of main images, infrared thermal imaging, and environmental data, and by introducing spatiotemporal tags and cause clustering analysis mechanisms, it achieves accurate identification and causal analysis of disease occurrence areas, development trends, and dominant inducing factors, thereby improving the interpretability and comprehensive identification accuracy of stem mustard diseases.

[0074] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the image sensing acquisition is used to acquire visual image information of the stem mustard plant, specifically by acquiring raw stem mustard data through multi-source image data acquisition.

[0075] The multi-source image data acquisition includes main image source acquisition, auxiliary sensor information acquisition, and recognition-enhanced image acquisition; the raw data of the stem mustard includes visible light images, basic environmental data, and infrared image data.

[0076] Preferably, the visible light images include overall plant images, top-view images, and close-up images of lesions; the basic environmental data includes shooting time, coordinate records, soil data, climate data, temporal variation data, and human intervention data; the infrared image data includes infrared thermal imaging images, average plant temperature, and collection timestamp data.

[0077] The soil data specifically includes soil temperature, soil pH value, and soil conductivity parameters; the climate data specifically includes light intensity, rainfall, and wind speed and direction data; the time-series variation data specifically refers to the temperature and humidity variation curve data within seven days; and the human intervention data specifically includes pesticide application records and irrigation time data.

[0078] Preferably, Table 1 is an example table of implementation parameters for the acquisition of the main image source. As shown in the table, the acquisition of the main image source is used to acquire visible light images, and the implementation parameters include image type, resolution, image angle, shooting distance, shooting lighting conditions, and close-up of suspected lesions; Table 2 is an example table of implementation parameters for the acquisition of the recognition-enhanced image. As shown in the table, the acquisition of the recognition-enhanced image is used to acquire infrared image data, and the implementation parameters include image type, resolution, image angle, shooting distance, temperature sensitivity, temperature measurement range, and acquisition time period.

[0079] Table 1. Example of implementation parameters for acquiring the main image source.

[0080]

[0081] Table 2 Example of implementation parameters for recognition-enhanced image acquisition

[0082]

[0083] Example 3, see Figure 1 , Figure 2 and Figure 3This embodiment is based on the above embodiment. In step S2, the data fusion processing is used to preprocess and fuse the acquired multi-source image data. Specifically, based on the original data of *Strombyx mori*, optimized data of *Strombyx mori* is obtained through image processing and sensor data encoding, including the following steps:

[0084] Step S21: Main image source preprocessing, specifically, performing image size unification, color space correction and brightness normalization on the visible light image in the original data of the stem mustard to obtain preprocessed image data, and performing plant region extraction operation based on the preprocessed image data, and obtaining the main image source features through image tensor construction;

[0085] Preferably, the plant region extraction operation specifically uses a lightweight SegNet to extract the plant region based on the preprocessed image data to obtain plant region image data.

[0086] The image tensor construction is specifically based on the plant region image data and the preprocessed image data, converting pixel values ​​and plant region binary masks into four-dimensional channel tensors to obtain the main image source features;

[0087] The formula for calculating the main image source features is as follows:

[0088] X rgbmask =[R,G,B,Mask];

[0089] In the formula, X rgbmask R is the normalized pixel value of the red channel, G is the normalized pixel value of the green channel, B is the normalized pixel value of the blue channel, and Mask is the binary mask of the plant region.

[0090] Step S22: Auxiliary sensor data synchronization encoding, specifically, performing outlier removal, timestamp standardization, and data encoding vectorization operations on the basic environmental data in the original data of the stem mustard to obtain sensor feature data;

[0091] The sensor feature data specifically includes time features, latitude and longitude location features, air humidity features, air temperature features, soil features, weather features, wind speed features, risk features, and human intervention event features;

[0092] Step S23: Feature fusion encoding, specifically, based on the main image source features and the sensor feature data, feature concatenation is used to obtain fused feature encoding data;

[0093] Step S24: Infrared image archiving, specifically, the infrared image data in the original data of the stem mustard is archived and saved as structured data to obtain the infrared image archive data;

[0094] Step S25: Data fusion processing, specifically, the fused feature encoding data and the infrared image archive data are stored in a structured manner to obtain the stem mustard optimized data;

[0095] The optimized data for *Strombyx mori* specifically includes optimized main image source features, optimized sensor features, and infrared images.

[0096] Example 4, see Figure 1 , Figure 2 and Figure 4 This embodiment is based on the above embodiment. In step S3, the spatiotemporal disease perception is used to combine the temporal and spatial distribution information of the image and perceive the disease area. Specifically, based on the optimized data of the stem mustard, a multimodal spatiotemporal perception method with infrared correction is used to perform spatiotemporal disease perception and obtain reference data of the spatiotemporally perceived disease area, including the following steps:

[0097] Step S31: Infrared feature branch construction, specifically, based on the infrared image in the optimized data of the stem mustard, the original infrared image is normalized to obtain a normalized infrared image feature map, and based on the normalized infrared image feature map, a single-channel convolutional encoder is constructed to extract the potential region features of the infrared hot spot to obtain the potential feature map data of the hot spot.

[0098] The formula for calculating the potential feature map data of the hot spot is:

[0099] ;

[0100] In the formula, F IR This is latent feature map data for hotspots. Conv1D is a one-dimensional convolution operator, ReLU is a non-linear activation function, BN is a batch normalization operator, and I... IR It is the original infrared image. It is the pixel average of the infrared image. It is the standard deviation of pixels in an infrared image;

[0101] Step S32: Spatiotemporal main map coding branch construction, specifically, constructing an improved backbone network and extracting features based on the optimized main map source features in the optimized stem mustard data to obtain map source feature data, and performing sinusoidal position coding based on the time features and latitude and longitude position features in the optimized sensing features to obtain spatiotemporal coding feature data, and fusing the map source feature data and the spatiotemporal coding feature data to obtain fused spatiotemporal feature data;

[0102] The sinusoidal position coding adopts the sinusoidal position coding form in the standard transformer structure for spatiotemporal coding.

[0103] The improved backbone network includes a fine-grained perception module, a multi-scale convolution extraction module, a texture attention module, and a feature aggregation module;

[0104] The fine-grained perception module is used to extract low-level local texture features and retain detailed information of the diseased area. Specifically, it uses 3×3 convolution, preferably with a stride of 2, an edge padding of 1, and an output channel number of 64.

[0105] The multi-scale convolution extraction module is used to capture features of lesions of different sizes. Specifically, it adopts a three-branch parallel structure. The first branch uses a 3×3 convolution with an edge padding of 1, the second branch uses a 5×5 convolution with an edge padding of 2, the third branch uses a 7×7 convolution with an edge padding of 3, and finally, it is summarized by a 1×1 convolution to form 128 channels.

[0106] The texture attention module is used to focus on abnormal texture areas of lesions and improve the perception of lesion edges and color gradient areas. Specifically, it introduces a channel attention mechanism, first using 1×1 convolution to compress the channel dimension, then performing global average pooling, and finally generating attention weights through the Sigmoid function and applying them to the original feature map. The number of input and output channels is 128.

[0107] The feature aggregation module is used to aggregate disease trend information, specifically including two dilated convolution layers: the first dilated convolution layer uses a 3×3 dilated convolution with an inflation rate of 2, and the second dilated convolution layer uses a 3×3 dilated convolution with an inflation rate of 4. Preferably, after the feature aggregation module, a global average pooling is performed again, and the feature is added to the original feature through channel fusion to form a residual connection structure.

[0108] Step S33: Feature fusion improvement, specifically, based on the hot spot potential feature map data, through the construction of a feature fusion layer, the attention feature weight of the infrared guidance channel is calculated, and based on the attention feature weight of the infrared guidance channel, the spatiotemporal feature data and the hot spot potential feature map data are dynamically weighted and fused to obtain infrared fused feature data. Then, by introducing a hot spot mask, the infrared fused feature data is locally enhanced to obtain infrared corrected fused feature data.

[0109] The formula for calculating the attention feature weight of the infrared guidance channel is as follows:

[0110] ;

[0111] In the formula, The infrared guidance channel attention feature weights, Softmax is the classifier operator, and W... q It is the feature weight of hotspots, F IR It is the potential feature map data of hot spots, W kIt is a fusion of spatiotemporal feature weights, F RF It integrates spatiotemporal feature data;

[0112] The formula for calculating the infrared fusion feature data is:

[0113] ;

[0114] In the formula, F fusion It is infrared fusion feature data. It is the attention feature weight of the infrared guidance channel, F RF It is a fusion of spatiotemporal feature data, F IR It is potential feature map data of hot spots;

[0115] The calculation formula for the infrared corrected fusion feature data is as follows:

[0116] ;

[0117] In the formula, F FU It is infrared corrected and fused feature data. This is the infrared correction parameter, with a default value of 0.5, M. hot It is a hotspot mask, and the specific calculation formula is as follows: ;

[0118] Step S34: Disease area identification, specifically, based on the infrared correction and fusion feature data, a lightweight decoder is constructed to perform disease area identification decoding to obtain disease area mask data, and spatiotemporal disease area identification data is obtained by outputting disease spatiotemporal features.

[0119] Step S35: Spatiotemporal disease perception, specifically, a single-channel convolutional encoder is constructed through the infrared feature branch, and an improved backbone network and transformer coding structure are constructed through the spatiotemporal main graph coding branch. A feature fusion layer is obtained through feature fusion improvement, and a decoder structure is obtained through disease area identification. Based on the single-channel convolutional encoder, the improved backbone network and transformer coding structure, the feature fusion layer, and the lightweight decoder, a spatiotemporal disease perception model is trained to obtain a spatiotemporal disease perception model. Using the spatiotemporal disease perception model and the stem mustard optimized data, spatiotemporal disease perception is performed to obtain spatiotemporal perception disease area reference data.

[0120] The spatiotemporal perception disease area reference data specifically includes a binary disease mask, infrared guidance channel attention feature weights, spatiotemporal annotation point set, and spatiotemporal disease perception visualization reference data.

[0121] By performing the above operations, this solution addresses the technical problems of insufficient utilization of infrared information, lack of dynamic attention capability for abnormal lesion regions, and inability to distinguish the evolution trend of lesions in adjacent time or space in existing spatiotemporal disease localization and perception methods. It creatively adopts a multimodal spatiotemporal perception method with infrared correction for spatiotemporal disease perception. Through infrared-guided channel attention mechanism and hot spot mask local enhancement technology, combined with the improved texture attention module and multi-scale convolutional structure of the backbone network, spatiotemporal disease perception is achieved, realizing high-precision identification of lesion evolution path, spatial aggregation trend and thermal anomaly region, effectively enhancing the sensitivity and generalization ability of early disease identification.

[0122] Example 5, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S4, the environmental cause association identification is used to analyze the association between disease manifestations and environmental factors. Specifically, based on the spatiotemporally perceived disease area reference data and the stem mustard optimized data, an improved clustering selection method driven by disease type is used to identify environmental cause associations and obtain environmental cause association feature data. This includes the following steps:

[0123] Step S41: Disease type determination, specifically, based on the spatiotemporal perception disease area reference data, visual texture features, infrared thermal features and spatial statistical features are extracted sequentially to obtain disease feature data, and a standard multilayer perceptron model is trained to determine the disease type based on the disease feature data to obtain disease type classification data;

[0124] The specific types of diseases mentioned include clubroot, white rust, black spot, downy mildew, and anthracnose.

[0125] The visual texture features include color histogram, grayscale features, and edge density features;

[0126] The infrared thermal features include average temperature features and maximum hot spot region features;

[0127] The spatial statistical features include lesion features, lesion shape features, and lesion aspect ratio features;

[0128] The standard multilayer perceptron model uses the disease feature data as input, employs ReLU and Softmax activation functions, and uses cross-entropy loss as the loss function to perform lightweight model classification.

[0129] Step S42: Disease environment data matching, specifically, from the disease type classification data, the predicted type, spatial information, timestamp information, and binary mask of the lesion area for each disease area are obtained, and sensor feature data is extracted from the optimized sensor features to obtain the environmental data index; based on the environmental data index, a joint distance metric method based on spatial distance and temporal difference is used to match each disease area with the environmental record that is closest to it in time and space, and the disease environment matching data is obtained by outputting structured disease-environment matching pairs;

[0130] The disease-environment matching data specifically includes disease type, spatial location, time, and matching environmental factor vectors;

[0131] Preferably, in the joint distance measurement method based on spatial distance and time difference, the spatial distance is calculated using the Haversine formula to determine the difference in latitude and longitude, and the time difference is calculated using hours as the unified unit.

[0132] Step S43: Improve clustering selection of dominant environmental factors. Based on the disease environment matching data, extract the sensor feature data corresponding to each disease area as an environmental factor vector, and extract the lesion mask area and infrared response intensity information. By calculating the lesion performance weight of each disease sample, obtain the lesion severity weight data. Based on the lesion performance weight, construct a standard weighted F value to calculate the weighted discrimination ability of each environmental factor vector for different disease types, obtain the dimensionality-reduced environmental factor parameters, and based on the dimensionality-reduced environmental factor parameters, use a clustering method that introduces lesion performance weight to perform cluster analysis on the disease samples to obtain the dominant environmental factor disease mapping data.

[0133] The formula for calculating the weight of the lesion manifestation is as follows:

[0134] ;

[0135] In the formula, w i This represents the weighting of lesion manifestations, where i is the lesion sample index, a0 is the adjustment coefficient (default value is 0.6), and A... i Here, A is the lesion mask area, j is the normalized lesion sample index, and A is the lesion mask area. j H is the lesion area corresponding to the normalized disease sample. i It is infrared response intensity information, specifically using the average infrared pixel value within the lesion mask area as the response intensity value, H j It is the infrared response intensity information corresponding to the normalized disease sample;

[0136] The reduced-dimensional environmental factor parameters are a set of parameters consisting of the top 6 environmental factors with the largest weighted F values;

[0137] The clustering method that introduces lesion manifestation weights specifically refers to an improvement on the standard K-Means clustering algorithm based on the lesion manifestation weights. The calculation formula is as follows:

[0138] ;

[0139] In the formula, J is the clustering objective function value, N is the total number of disease samples, i is the disease sample index, and w i It is the weight of lesion manifestation, e i These are the dimension-reduced environmental factor parameters corresponding to the diseased samples. It is the zth i The cluster center vectors corresponding to each cluster category, where z i It is the cluster category number;

[0140] Step S44: Environmental etiology association identification, specifically, based on the dominant environmental factor disease mapping data, extracting the most significant environmental inducing factors for each disease type to obtain environmental etiology association feature data.

[0141] By performing the above operations, this solution addresses the technical problems in existing disease environmental cause analysis methods, such as the decoupling of disease types from environmental factors and the inability to extract key dominant environmental factors, resulting in a lack of practicality and controllability in the analysis results. It creatively adopts a disease type-driven improved clustering selection method for major environmental factors to identify environmental causal associations. By introducing lesion manifestation weights and a weighted F-value dimensionality reduction mechanism, combined with an improved K-Means clustering algorithm, it achieves automatic attribution and association of key inducing factors under different disease types, providing reliable data support for agricultural management and precision prevention.

[0142] Example 6, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S5, the stem mustard disease identification is used to identify and classify diseases by combining disease area and disease association information. Specifically, based on the environmental etiology association feature data and the spatiotemporal perception disease area reference data, the disease type classification results, lesion spatial distribution information and dominant environmental factor analysis results are integrated to perform comprehensive classification and output to obtain disease identification reference data.

[0143] The disease identification reference data specifically includes disease type labels, disease area masks, disease spatiotemporal information, infrared response intensity, and a set of dominant environmental factors.

[0144] By performing the above operations, this solution addresses the technical problem in existing integrated disease identification methods that suffer from fragmented identification processes and a lack of joint modeling of spatiotemporal and environmental factors, ultimately resulting in identification results lacking global reference value. This solution creatively adopts an integrated identification process that combines spatiotemporal location, environmental disease correlation, and infrared image analysis. It unifies the modeling and classification of image features, thermal imaging information, spatial trajectories, and environmental factors, ultimately achieving comprehensive identification and output of disease types, lesion masks, infrared response intensity, and environmental factor sets. This improves the practical adaptability of the intelligent identification system and its ability to guide precise agricultural pest control.

[0145] Example 7, see Figure 1 and Figure 2 Based on the above embodiments, the artificial intelligence-based stem mustard disease identification system provided by the present invention includes a data collection module, a fusion processing module, a disease perception module, a disease cause association module, and a disease identification module.

[0146] The data collection module is used for image sensing acquisition. Through image sensing acquisition, it obtains raw data of stem mustard and sends the raw data of stem mustard to the fusion processing module.

[0147] The fusion processing module is used for data fusion processing. Through data fusion processing, optimized data of stem mustard is obtained, and the optimized data of stem mustard is sent to the disease perception module and the cause association module.

[0148] The disease perception module is used for spatiotemporal disease perception. Through spatiotemporal disease perception, it obtains spatiotemporal perception disease area reference data and sends the spatiotemporal perception disease area reference data to the disease cause association module and the disease identification module.

[0149] The etiology association module is used for environmental etiology association. Through environmental etiology association, environmental etiology association feature data is obtained, and the environmental etiology association feature data is sent to the disease identification module.

[0150] The disease identification module is used for identifying stem mustard diseases and obtaining disease identification reference data through stem mustard disease identification.

[0151] 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.

[0152] 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.

[0153] 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 identifying stem mustard diseases based on artificial intelligence, characterized in that: The method includes the following steps: Step S1: Image sensing acquisition, obtaining raw data of stem mustard through multi-source image data acquisition; Step S2: Data fusion processing: Based on the original data of the stem mustard, optimized data of the stem mustard is obtained through image processing and sensor data encoding; Step S3: Spatiotemporal disease perception. Based on the optimized data of stem mustard, a multimodal spatiotemporal perception method with infrared correction is adopted to perform spatiotemporal disease perception and obtain reference data of spatiotemporal disease area. The proposed multimodal spatiotemporal sensing method with infrared correction involves constructing a single-channel convolutional encoder to extract features from potential infrared hotspot regions, building an improved backbone network to extract main image source features, performing feature fusion through sinusoidal position encoding, introducing hotspot masks for local enhancement, and constructing a lightweight decoder for disease area identification and decoding. The improved backbone network includes a fine-grained sensing module, a multi-scale convolutional extraction module, a texture attention module, and a feature aggregation module. Step S4: Environmental etiology association. Based on the spatiotemporally perceived disease area reference data and the stem mustard optimized data, the main environmental factor improved clustering selection method driven by disease type is used to identify environmental etiology associations and obtain environmental etiology association feature data. Step S5: Identify stem mustard diseases, perform comprehensive classification and output, and obtain disease identification reference data.

2. The method for identifying stem mustard diseases based on artificial intelligence according to claim 1, characterized in that: In step S1, the image sensing acquisition is used to acquire visual image information of the stem mustard plant, specifically by acquiring raw stem mustard data through multi-source image data acquisition. The multi-source image data acquisition includes main image source acquisition, auxiliary sensor information acquisition, and recognition-enhanced image acquisition; the raw data of *Strombyx mori* includes visible light images, basic environmental data, and infrared image data.

3. The method for identifying stem mustard disease based on artificial intelligence according to claim 2, characterized in that: In step S2, the data fusion processing is used to preprocess and fuse the acquired multi-source image data. Specifically, based on the original data of *Strombyx mori*, optimized data of *Strombyx mori* is obtained through image processing and sensor data encoding. The process includes the following steps: Step S21: Main image source preprocessing, specifically, the visible light image in the original data of *Strombyx mori* is subjected to image size unification, color space correction and brightness normalization to obtain preprocessed image data. Based on the preprocessed image data, plant region extraction is performed, and the main image source features are obtained through image tensor construction. Step S22: Auxiliary sensor data synchronization encoding, specifically, performing outlier removal, timestamp standardization, and data encoding vectorization operations on the basic environmental data in the original data of the stem mustard to obtain sensor feature data; The sensor feature data specifically includes time features, latitude and longitude location features, air humidity features, air temperature features, soil features, weather features, wind speed features, risk features, and human intervention event features; Step S23: Feature fusion encoding, specifically, based on the main image source features and the sensor feature data, feature concatenation is used to obtain fused feature encoding data; Step S24: Infrared image archiving, specifically, the infrared image data in the original data of the stem mustard is archived and saved as structured data to obtain the infrared image archive data; Step S25: Data fusion processing, specifically, the fused feature encoding data and the infrared image archive data are structured and stored to obtain stem mustard optimized data.

4. The method for identifying stem mustard disease based on artificial intelligence according to claim 3, characterized in that: In step S2, the optimized data for *Strombyx mori* specifically includes optimized main image source features, optimized sensor features, and infrared images.

5. The method for identifying stem mustard disease based on artificial intelligence according to claim 4, characterized in that: In step S3, the spatiotemporal disease perception is used to combine the temporal and spatial distribution information of the image and perceive the disease area. Specifically, based on the optimized data of the stem mustard, a multimodal spatiotemporal perception method with infrared correction is used to perform spatiotemporal disease perception and obtain reference data for the spatiotemporally perceived disease area, including the following steps: Step S31: Infrared feature branch construction, specifically, based on the infrared image in the optimized data of the stem mustard, the original infrared image is normalized to obtain a normalized infrared image feature map, and based on the normalized infrared image feature map, a single-channel convolutional encoder is constructed to extract the potential region features of the infrared hot spot to obtain the potential feature map data of the hot spot. Step S32: Spatiotemporal main map coding branch construction, specifically, constructing an improved backbone network and extracting features based on the optimized main map source features in the optimized stem mustard data to obtain map source feature data, and performing sinusoidal position coding based on the time features and latitude and longitude position features in the optimized sensing features to obtain spatiotemporal coding feature data, and fusing the map source feature data and the spatiotemporal coding feature data to obtain fused spatiotemporal feature data; The sinusoidal position encoding adopts the sinusoidal position encoding form in the standard transformer structure for spatiotemporal encoding; the improved backbone network includes a fine-grained perception module, a multi-scale convolution extraction module, a texture attention module, and a feature aggregation module; The fine-grained perception module is used to extract low-level local texture features and retain detailed information of the diseased area; the multi-scale convolution extraction module is used to capture features of lesions of different sizes; the texture attention module is used to focus on abnormal texture areas of lesions and improve the perception ability of lesion edges and color gradient areas; the feature aggregation module is used to aggregate disease trend information.

6. The method for identifying stem mustard disease based on artificial intelligence according to claim 5, characterized in that: In step S3, step S33: feature fusion improvement, specifically, based on the hot spot potential feature map data, through the construction of a feature fusion layer, the attention feature weight of the infrared guidance channel is calculated, and based on the attention feature weight of the infrared guidance channel, the spatiotemporal feature data and the hot spot potential feature map data are dynamically weighted and fused to obtain infrared fused feature data. Then, by introducing a hot spot mask, the infrared fused feature data is locally enhanced to obtain infrared corrected fused feature data. Step S34: Disease area identification, specifically, based on the infrared correction and fusion feature data, a lightweight decoder is constructed to perform disease area identification decoding to obtain disease area mask data, and spatiotemporal disease area identification data is obtained by outputting disease spatiotemporal features. Step S35: Spatiotemporal disease perception, specifically, a single-channel convolutional encoder is constructed through the infrared feature branch, and an improved backbone network and transformer coding structure are constructed through the spatiotemporal main graph coding branch. A feature fusion layer is obtained through feature fusion improvement, and a decoder structure is obtained through disease area identification. Based on the single-channel convolutional encoder, the improved backbone network and transformer coding structure, the feature fusion layer, and the lightweight decoder, a spatiotemporal disease perception model is trained to obtain a spatiotemporal disease perception model. Using the spatiotemporal disease perception model and the stem mustard optimized data, spatiotemporal disease perception is performed to obtain spatiotemporal perception disease area reference data. The spatiotemporal perception disease area reference data specifically includes a binary disease mask, infrared guidance channel attention feature weights, spatiotemporal annotation point set, and spatiotemporal disease perception visualization reference data.

7. The method for identifying stem mustard disease based on artificial intelligence according to claim 6, characterized in that: In step S4, the environmental etiology association identification is used to analyze the correlation between disease manifestations and environmental factors. Specifically, based on the spatiotemporally perceived disease area reference data and the optimized data of stem mustard, an improved clustering selection method driven by disease type is used to identify environmental etiology associations and obtain environmental etiology association feature data, including the following steps: Step S41: Disease type determination, specifically, based on the spatiotemporal perception disease area reference data, visual texture features, infrared thermal features and spatial statistical features are extracted sequentially to obtain disease feature data, and a standard multilayer perceptron model is trained to determine the disease type based on the disease feature data to obtain disease type classification data; The specific types of diseases mentioned include clubroot, white rust, black spot, downy mildew, and anthracnose. Step S42: Disease environment data matching, specifically, from the disease type classification data, the predicted type, spatial information, timestamp information, and binary mask of the lesion area for each disease area are obtained, and sensor feature data is extracted from the optimized sensor features to obtain the environmental data index; based on the environmental data index, a joint distance metric method based on spatial distance and temporal difference is used to match each disease area with the environmental record that is closest to it in time and space, and the disease environment matching data is obtained by outputting structured disease-environment matching pairs; Step S43: Improve clustering selection for dominant environmental factors. Based on the disease environment matching data, extract the sensor feature data corresponding to each disease area as an environmental factor vector, and extract the lesion mask area and infrared response intensity information. By calculating the lesion performance weight of each disease sample, obtain the lesion severity weight data. Based on the lesion performance weight, construct a standard weighted F value to calculate the weighted discrimination ability of each environmental factor vector for different disease types, obtain the dimensionality-reduced environmental factor parameters, and based on the dimensionality-reduced environmental factor parameters, use a clustering method that introduces lesion performance weight to perform cluster analysis on the disease samples to obtain the dominant environmental factor disease mapping data. The clustering method that introduces lesion performance weight specifically refers to improving the standard K-Means clustering algorithm based on the lesion performance weight. Step S44: Environmental etiology association identification, specifically, based on the dominant environmental factor disease mapping data, extracting the most significant environmental inducing factors for each disease type to obtain environmental etiology association feature data.

8. The method for identifying stem mustard disease based on artificial intelligence according to claim 7, characterized in that: In step S5, the stem mustard disease identification is used to identify and classify diseases by combining disease area and disease association information. Specifically, based on the environmental etiology association feature data and the spatiotemporal perception disease area reference data, the disease type classification results, lesion spatial distribution information and dominant environmental factor analysis results are integrated and classified to obtain disease identification reference data. The disease identification reference data specifically includes disease type labels, disease area masks, disease spatiotemporal information, infrared response intensity, and a set of dominant environmental factors.

9. An artificial intelligence-based stem mustard disease identification system, used to implement the artificial intelligence-based stem mustard disease identification method as described in any one of claims 1-8, characterized in that: It includes a data collection module, a fusion processing module, a disease perception module, a disease cause association module, and a disease identification module.

10. The artificial intelligence-based stem mustard disease identification system according to claim 9, characterized in that: The data collection module is used for image sensing acquisition. Through image sensing acquisition, it obtains raw data of stem mustard and sends the raw data of stem mustard to the fusion processing module. The fusion processing module is used for data fusion processing. Through data fusion processing, optimized data of stem mustard is obtained, and the optimized data of stem mustard is sent to the disease perception module and the cause association module. The disease perception module is used for spatiotemporal disease perception. Through spatiotemporal disease perception, it obtains spatiotemporal perception disease area reference data and sends the spatiotemporal perception disease area reference data to the disease cause association module and the disease identification module. The etiology association module is used for environmental etiology association. Through environmental etiology association, environmental etiology association feature data is obtained, and the environmental etiology association feature data is sent to the disease identification module. The disease identification module is used for identifying stem mustard diseases and obtaining disease identification reference data through stem mustard disease identification.

Images