Corn disease and insect pest intelligent identification and spatial distribution visualization system

Abstract

The invention relates to the technical field of agricultural information, in particular to a corn disease and insect pest intelligent identification and spatial distribution visualization system which comprises a data acquisition module, a multi-scale feature representation module, a feature space optimization module, a multi-modal feature fusion module, a disease and insect pest identification module, a spatial distribution analysis module and a visualization decision module. By introducing a differential geometry theory, a manifold learning-based multi-scale feature representation technology is constructed, and the problem of early-stage tiny scab recognition is solved; a feature space transformation and optimization strategy under geometric constraints is adopted, so that the recognition robustness in a complex field environment is improved; a multi-modal feature intelligent fusion mechanism is designed, and collaborative analysis of RGB, multispectral and hyperspectral data is realized; generating spatial distribution data of diseases and pests through spatial interpolation, spatial-temporal evolution, clustering analysis and propagation path analysis; according to the method, multi-layer visualization and risk assessment are combined, partition prevention and control suggestions are provided, and the early recognition capability of corn diseases and insect pests is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural information technology, in particular to a corn disease and pest intelligent identification and spatial distribution visualization system for realizing intelligent identification, spatial distribution analysis and visual decision support of corn diseases and pests. BACKGROUND

[0002] Corn is an important food crop in China, and its stable and high yield is of great significance to the national food security. However, corn diseases and pests cause serious losses to agricultural production every year. Traditional disease and pest monitoring mainly relies on manual inspection, which is inefficient and greatly affected by subjective factors.

[0003] With the development of artificial intelligence and information technology, intelligent disease and pest identification technology has been gradually applied to agricultural production. Existing technologies mainly use single modal image recognition methods, usually using deep learning algorithms based on RGB images for disease classification. However, these methods have low accuracy in dealing with early-stage corn diseases, especially for early-stage diseases with unclear symptoms. In addition, most existing technologies are limited to single disease identification and lack analysis of spatial distribution patterns and development trends of diseases and pests, making it difficult to provide effective decision support for precision prevention and control.

[0004] The main problems existing in the prior art include: 1) poor recognition effect for early-stage corn micro lesions; 2) insufficient recognition stability under complex field environmental conditions (such as light changes, angle differences, etc.); 3) unable to effectively integrate multi-source data for comprehensive analysis; 4) lack of prediction of spatial distribution and transmission trend of diseases and pests. Therefore, it is urgent to develop a system that can realize early-stage accurate identification of corn diseases and pests, spatial distribution analysis and visual decision support. SUMMARY

[0005] The purpose of the present application is to provide a corn disease and pest intelligent identification and spatial distribution visualization system, which aims to overcome the problems existing in the prior art and realize early-stage accurate identification, spatial distribution dynamic analysis and visual decision support of corn diseases and pests.

[0006] The present application proposes a big data driven dynamic pricing intelligent decision system, which includes:

[0007] A data acquisition module for acquiring multi-source image data of corn;

[0008] A multi-scale feature representation module in communication connection with the data acquisition module, for receiving the multi-source image data and mapping the multi-source image data to a high-dimensional manifold space to form a feature representation;

[0009] A feature space optimization module in communication connection with the multi-scale feature representation module, for receiving the feature representation, constructing an adaptive metric structure and generating deformation invariant features;

[0010] The multimodal feature fusion module is communicatively connected to the multi-scale feature representation module and the feature space optimization module, respectively, and is used to establish the mapping relationship between multimodal data, fuse different modal features, and generate fused features;

[0011] The pest and disease identification module is communicatively connected to the multimodal feature fusion module and is used to identify the type and location information of corn pests and diseases based on the fused features.

[0012] The spatial distribution analysis module is communicatively connected to the pest and disease identification module and is used to receive the corn pest and disease type and location information and generate pest and disease spatial distribution data.

[0013] The visualization decision-making module is communicatively connected to the spatial distribution analysis module and is used to receive the spatial distribution data of pests and diseases, and generate visualization results of pest and disease distribution and prevention and control decision-making suggestions.

[0014] Preferably, the data acquisition module includes:

[0015] The drone data acquisition unit is used to carry a multispectral camera and a positioning system to collect aerial image data of corn fields.

[0016] Ground-based mobile platform for acquiring high-resolution RGB images and close-range spectral data;

[0017] The data preprocessing unit is communicatively connected to the UAV acquisition unit and the ground mobile platform, and is used to preprocess the acquired image data to generate standardized image data.

[0018] The task scheduling unit is communicatively connected to the UAV data collection unit and the ground mobile platform. It is used to automatically plan the data collection path according to the field conditions and arrange close-range sampling for detected suspicious areas.

[0019] Preferably, the multi-scale feature representation module includes:

[0020] A multi-scale processing unit is used to process the input image data in different spatial resolutions at different levels.

[0021] The feature descriptor generation unit is communicatively connected to the multi-scale processing unit and is used to construct a set of feature descriptors for each resolution level.

[0022] The manifold mapping unit, which is communicatively connected to the feature descriptor generation unit, is used to map high-dimensional features to a low-dimensional manifold space while preserving the inherent structural relationships of the data.

[0023] A scale correlation unit, in communication connection with the manifold mapping unit, is configured to construct a scale correlation graph and automatically adjust weights of different scale features according to a disease type.

[0024] Preferably, the feature space optimization module comprises:

[0025] A metric construction unit is configured to construct a local adaptive metric matrix based on sample distribution characteristics;

[0026] A deformation invariant feature unit, in communication connection with the metric construction unit, is configured to establish a light change model and a view angle transformation set and extract deformation invariant features;

[0027] A manifold optimization unit, in communication connection with the deformation invariant feature unit, is configured to analyze local curvature characteristics of a manifold and implement a dense sampling strategy on a high curvature area;

[0028] A feature conversion unit, in communication connection with the manifold optimization unit, is configured to define a geodesic path in a feature space and implement optimal feature conversion between different observation conditions.

[0029] Preferably, the multi-modal feature fusion module comprises:

[0030] A modal mapping unit is configured to construct a mapping function between different modal feature spaces;

[0031] A semantic consistency unit, in communication connection with the modal mapping unit, is configured to construct a multi-modal shared semantic space and ensure semantic consistency of different modal features;

[0032] An entropy evaluation unit, in communication connection with the semantic consistency unit, is configured to evaluate information richness of different modal features through information entropy and dynamically adjust weights of the modal features;

[0033] A decision integration unit, in communication connection with the entropy evaluation unit, is configured to establish a multi-level decision framework and implement reliability evaluation and error correction on a fusion result.

[0034] Preferably, the spatial distribution analysis module comprises:

[0035] A spatial interpolation unit is configured to construct a disease distribution discrete point set based on a detection result and generate a continuous distribution density map by using an adaptive spatial interpolation algorithm;

[0036] A spatio-temporal evolution unit, in communication connection with the spatial interpolation unit, is configured to construct a disease propagation model and predict a disease diffusion trend;

[0037] A clustering analysis unit, in communication connection with the spatial interpolation unit, is configured to implement multi-scale spatial clustering and identify a high disease incidence area;

[0038] The propagation path analysis unit is in communication connection with the spatio-temporal evolution unit and the clustering analysis unit, and is configured to analyze the main path of disease propagation in combination with environmental factors.

[0039] As preferred, the visualization decision module comprises:

[0040] The multi-layer visualization unit is configured to construct a multi-layer visualization interface based on a geographic information system, and support multi-layer superimposed display of a base map, a distribution map, a prediction map, etc.

[0041] The interactive control unit is in communication connection with the multi-layer visualization unit, and is configured to realize human-computer interaction functions such as region zooming, layer control, and condition query.

[0042] The risk assessment unit is in communication connection with the multi-layer visualization unit, and is configured to establish a disease impact assessment model to quantify the potential impact of diseases on yield.

[0043] The prevention and control suggestion generation unit is in communication connection with the risk assessment unit, and is configured to generate zoned prevention and control suggestions and optimal pesticide application paths based on disease distribution characteristics and severity.

[0044] As preferred, the scale correlation unit constructs scale correlation by the following ways:

[0045] An association graph composed of nodes representing different scale characteristics and edges representing correlation strength is constructed.

[0046] A scale adaptability weight function is designed to automatically adjust the importance of different scale characteristics according to disease types.

[0047] A scale space smooth transition mechanism is implemented to ensure continuous representation from coarse-grained to fine-grained characteristics.

[0048] For early disease characteristics, the weight of fine-grained characteristics is dynamically enhanced; for mid-late disease, the weights of multi-scale characteristics are balanced.

[0049] As preferred, the manifold optimization unit optimizes the manifold structure by the following ways:

[0050] The local curvature characteristics of the manifold are analyzed based on sample distribution to identify high-curvature regions.

[0051] A dense sampling strategy is implemented for high-curvature regions to enhance the learning of the model on decision boundaries.

[0052] Manifold smooth regularization is introduced to prevent overfitting and noise interference.

[0053] An adaptive learning rate scheduler is constructed to use different learning rates for high-curvature regions and low-curvature regions.

[0054] As preferred, the semantic consistency unit maintains semantic consistency of the multi-modal features by the following ways:

[0055] A multi-modal shared semantic space is constructed, and all modal features are finally mapped to this space;

[0056] A semantic consistency evaluation index is designed to quantify the semantic matching degree of different modal features;

[0057] A semantic alignment optimization is implemented to make the feature representations of the same disease under different modalities consistent;

[0058] A multi-level semantic hierarchy is established to support the identification needs of different granularities of disease categories, disease subtypes and disease severity.

[0059] The present application innovatively establishes a multi-scale feature representation technology based on manifold learning by introducing differential geometry theory, solves the problem of early small lesion identification; by constructing feature space transformation and optimization strategy under geometric constraints, the recognition robustness of the system in complex field environment is improved; by designing a multi-modal feature intelligent fusion mechanism, the collaborative analysis of RGB, multispectral and hyperspectral data is realized; through spatial analysis and visualization decision support, scientific basis is provided for disease and pest control.

[0060] The beneficial effects of the present application include:

[0061] 1. The early identification ability of corn diseases and pests is greatly improved, the detection rate when the lesion area is less than 1% reaches 85%-90%, which is 7-10 days earlier than traditional methods to find diseases, and valuable time is won for early prevention and control;

[0062] 2. The adaptability of the system in complex environment is significantly enhanced, and the recognition accuracy under different light and angle conditions is controlled within ±3%;

[0063] 3. Through intelligent fusion of multi-source data, the identification accuracy of diseases and pests is improved by 15-20 percentage points, especially the identification accuracy of corn borer, corn leaf roller and other pests is improved to 97%;

[0064] 4. Dynamic visualization and transmission trend prediction of disease and pest distribution are realized, the prediction accuracy is more than 85%, and scientific basis is provided for precision prevention and control;

[0065] 5. In practical application, the regional prevention and control measures under the guidance of the system reduce the pesticide use amount by 30%, improve the prevention and control effect by 25%, significantly reduce the environmental pollution and improve the economic benefit. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 It is the overall architecture diagram of the corn disease and pest intelligent identification and spatial distribution visualization system of the present application;

[0067] Figure 2 It is a structural schematic diagram of the data acquisition module of the application;

[0068] Figure 3 It is a workflow diagram of the multi-scale feature representation module of the application;

[0069] Figure 4 It is a structural schematic diagram of the feature space optimization module of the application;

[0070] Figure 5 It is a working principle diagram of the multi-modal feature fusion module of the application;

[0071] Figure 6 It is a workflow diagram of the spatial distribution analysis module of the application;

[0072] Figure 7 It is an interface schematic diagram of the visual decision module of the application. DETAILED DESCRIPTION

[0073] Reference Figures 1-7 The specific embodiments described herein are intended to be illustrative only and are not intended to limit the scope of the application. Unless otherwise defined, scientific and technical terms used in connection with the application shall have the meanings that are commonly understood by those of ordinary skill in the art.

[0074] Reference Figure 1 The corn pest intelligent identification and spatial distribution visualization system provided by the application comprises a data acquisition module 1, a multi-scale feature representation module 2, a feature space optimization module 3, a multi-modal feature fusion module 4, a pest identification module 5, a spatial distribution analysis module 6, and a visual decision module 7.

[0075] The data acquisition module 1 is used to acquire corn multi-source image data. The multi-scale feature representation module 2 is in communication connection with the data acquisition module 1, and is used to receive the multi-source image data and map the multi-source image data to a high-dimensional manifold space to form a feature representation. The feature space optimization module 3 is in communication connection with the multi-scale feature representation module 2, and is used to receive the feature representation, construct an adaptive metric structure, and generate a deformation-invariant feature. The multi-modal feature fusion module 4 is in communication connection with the multi-scale feature representation module 2 and the feature space optimization module 3, respectively, and is used to establish a mapping relationship between multi-modal data, fuse different modal features, and generate a fusion feature. The pest identification module 5 is in communication connection with the multi-modal feature fusion module 4, and is used to identify the type and location information of corn pests based on the fusion feature. The spatial distribution analysis module 6 is in communication connection with the pest identification module 5, and is used to receive the type and location information of corn pests and generate pest spatial distribution data. The visual decision module 7 is in communication connection with the spatial distribution analysis module 6, and is used to receive the pest spatial distribution data, generate pest distribution visualization results, and provide prevention and control decision suggestions.

[0076] In the embodiments of the present application, data transmission is carried out between modules through a wireless network or a wired connection, and the whole system adopts a distributed architecture to realize intelligentization of the whole process of data acquisition, processing, analysis and decision-making.

[0077] With reference to Figure 2 , the data acquisition module 1 comprises: a UAV acquisition unit 11, a ground mobile platform 12, a data preprocessing unit 13 and a task scheduling unit 14.

[0078] The UAV acquisition unit 11 is used to carry a multispectral camera and a positioning system to acquire aerial image data of corn fields. Preferably, the UAV acquisition unit 11 adopts a fixed-wing UAV, carries a multispectral camera of 5-12 bands, has a resolution of 1280x960 pixels, and is equipped with a high-precision GPS / IMU system for geographic positioning. The flight height is usually set between 60-120 meters to ensure that the image resolution reaches 3-8 centimeters / pixel.

[0079] The ground mobile platform 12 is used to acquire high-resolution RGB images and close-range spectral data. In an embodiment of the present application, the ground mobile platform 12 adopts a tracked or wheeled robot, is equipped with a 4K resolution RGB camera and a portable spectrometer (wavelength range 400-2500 nm), can enter the plant gap for close-range observation, and can acquire high-precision disease detail images.

[0080] The data preprocessing unit 13 is in communication connection with the UAV acquisition unit 11 and the ground mobile platform 12, and is used to preprocess the acquired image data to generate standardized image data. Specifically, the data preprocessing unit 13 performs geometric correction, radiation correction and atmospheric correction and the like on the images to eliminate the influences of lens distortion, uneven illumination and atmospheric scattering and the like. In addition, through an image registration algorithm, the images acquired at different times and by different sensors are aligned, and the registration accuracy is controlled within 2 pixels.

[0081] The task scheduling unit 14 is in communication connection with the UAV acquisition unit 11 and the ground mobile platform 12, and is used to automatically plan an acquisition path according to the field conditions, and arrange close-range sampling for the detected suspicious areas. The task scheduling unit 14 generates an optimal flight path by using a path planning algorithm according to the area and shape of the farmland, ensures that the forward overlap degree of image acquisition reaches 60%, the lateral overlap degree reaches 75%, and the accuracy of three-dimensional reconstruction and tiling is ensured. At the same time, when a suspicious area is found in the aerial image, the task scheduling unit 14 calculates its geographic coordinates and assigns the ground mobile platform 12 to go to the corresponding position for high-precision sampling.

[0082] With reference to Figure 3The multi-scale feature representation module 2 comprises a multi-scale processing unit 21, a feature descriptor generation unit 22, a manifold mapping unit 23 and a scale correlation unit 24.

[0083] The multi-scale processing unit 21 is configured to process the input image data at different spatial resolutions. Specifically, the multi-scale processing unit 21 decomposes the original image into 4-6 different resolution levels by methods such as Gaussian pyramid or wavelet transform, with the resolution ranging from the original resolution to 1 / 16 of the original resolution. For early corn diseases, particular attention is paid to the high resolution levels (original resolution and 1 / 2 resolution) because these levels can capture the detailed features of small disease spots.

[0084] The feature descriptor generation unit 22 is in communication connection with the multi-scale processing unit 21 and is configured to construct a feature descriptor set for each resolution level. Preferably, for RGB images, texture features (such as LBP, HOG, etc.) are extracted; for multispectral images, various vegetation indices (such as NDVI, GNDVI, etc.) are calculated; and for hyperspectral images, spectral feature curves and feature bands are extracted. For early symptoms of corn small spot disease and other diseases that are not obvious, the feature descriptor generation unit 22 focuses on extracting edge and texture change features, which have high sensitivity to early disease spots.

[0085] The manifold mapping unit 23 is in communication connection with the feature descriptor generation unit 22 and is configured to map high-dimensional features to a low-dimensional manifold space while maintaining the intrinsic structural relationship of the data. Here, the manifold learning method in differential geometry theory is introduced to map the data points in the high-dimensional feature space to a low-dimensional manifold, thereby better capturing the intrinsic structure of the data. The mapping process can be represented as:

[0086] ,

[0087] wherein, is a high-dimensional feature descriptor set, representing a set of feature vectors extracted from the original image, usually with a dimension of hundreds to thousands of dimensions; is a manifold space, representing a low-dimensional representation space after mapping, usually with a dimension of 15-60; is a mapping function, representing a non-linear mapping from a high-dimensional feature space to a low-dimensional manifold space. In actual implementation, algorithms such as Local Linear Embedding (LLE) or Laplacian Eigenmap can be used to realize this mapping. The manifold dimension is usually set to 15-60, which is adjusted adaptively according to the complexity of the disease.

[0088] The scale correlation unit 24 is in communication connection with the manifold mapping unit 23 and is configured to construct a scale correlation graph and automatically adjust the weights of different scale features according to the disease type. The scale correlation unit 24 constructs the scale correlation in the following way:

[0089] 1) Construct a correlation graph composed of nodes representing different scale features and edges representing correlation strength. The correlation graph can be represented as where the node set represents features of different scales, each node corresponds to a feature representation of a scale level; the edge set represents correlation strength, representing the degree of correlation between different scale features, with a value range of [0, 1]. The construction of the correlation graph is based on the mutual information or correlation measure between different scale features, ensuring that the scale features with strong correlation have higher edge weights.

[0090] 2) Design a scale adaptive weight function to automatically adjust the importance of different scale features according to the disease type. The weight function can be represented as:

[0091]

[0092] where, represents the disease type the weight of the lower scale , with a value range of [0, 1]; represents the current scale level, an integer value, ranging from 1 to 6, corresponding to different resolution levels; represents the disease type, such as corn small spot, corn rust, etc.; represents the optimal recognition scale of the disease type , an integer value, ranging from 1 to 6; is the weight adjustment parameter, with a value range of [0, 1]; is the scale decay parameter, with a value range of [0, 5], controlling the rate of weight change with scale; is the base of natural logarithm. For example, for corn small spot, set to high resolution level (such as 1, representing the original resolution), set to 0.8, set to 2.5; while for corn rust and other diseases with obvious symptoms in the middle and late stages, can be set to medium resolution level (such as 3, representing 1 / 4 of the original resolution), set to 0.6, set to 2.5; while for corn rust and other diseases with obvious symptoms in the middle and late stages, can be set to medium resolution level (such as 3, representing 1 / 4 of the original resolution), set to 0.6, set to 1.2.

[0093] 3) Implement a scale space smooth transition mechanism to ensure continuous representation from coarse to fine-grained features. By constructing a scale space continuity constraint, the feature representation between adjacent scales has a smooth transition, avoiding feature discontinuity between scales.

[0094] 4) Dynamically enhance fine-grained feature weights for early disease features; balance multi-scale feature weights for mid-late disease. Specifically, for early disease with lesion area less than 1%, fine-grained feature (high resolution) weight is set to 0.6-0.8, while for mid-late disease with lesion area greater than 5%, multi-scale feature weight is balanced to 0.4-0.6.

[0095] Through the above method, the multi-scale feature representation module 2 can effectively capture the characteristics of corn diseases at different development stages, especially improving the recognition sensitivity to early small lesions.

[0096] Reference Figure 4 , the feature space optimization module 3 includes: metric construction unit 31, deformation invariant feature unit 32, manifold optimization unit 33 and feature conversion unit 34.

[0097] The metric construction unit 31 is used to construct a local adaptive metric matrix based on the sample distribution characteristics. Specifically, the metric construction unit 31 constructs an adaptive metric matrix for each local region in the manifold space , minimizing the distance between similar samples and maximizing the distance between different class samples. The metric matrix is constructed in the form of a region block (block size from 8x8 to 32x32), which adapts to image regions of different complexity. For each block, the metric matrix can be represented as:

[0098] ,

[0099] where is the metric matrix at position , which is a symmetric positive definite matrix with the same dimension as the feature dimension; is the eigenvalue, indicating the importance weight of the corresponding feature direction, with a positive real value range; is the eigenvector, indicating the main variation direction in the feature space, which is a unit vector; is the transpose of ; is the rank of the metric matrix, an integer value, usually set to 1 / 3 to 1 / 2 of the feature dimension; represents the summation operation. For high complexity regions near the classification boundary, increase the value of to provide a more fine-grained metric structure; for low complexity regions within the class, reduce the value of to reduce the risk of overfitting.

[0100] The deformation-invariant feature unit 32 is in communication connection with the metric construction unit 31, for establishing a light change model and a view transformation set, and extracting deformation-invariant features. Specifically, the deformation-invariant feature unit 32 first constructs a light change model , simulating light conditions at different times such as morning, noon, and evening, is a light parameter vector, including light intensity, direction, and color temperature, etc. At the same time, a view transformation set is constructed, covering multi-angle observation conditions such as top view and side view. Based on these models, deformation-invariant features that remain stable under different observation conditions are extracted, which can be expressed as:

[0101] ,

[0102] Among them, is a deformation-invariant feature extractor, representing a function capable of extracting features insensitive to deformation; is a feature extraction function, representing a mapping from an image to a feature; is an input image; is a light change model, mapping an image and a light parameter to a transformed image; is the view transformation, mapping an image to an image under different view angles; is the light parameter vector; is a feature distance metric function, calculating the distance between two feature vectors; is the number of view transformations, an integer value, usually 5-10; is the number of light changes, an integer value, usually 3-5; represents a function that finds the minimum of the following expression In actual implementation, the deformation-invariant feature extractor can be obtained through data augmentation and adversarial training, etc.

[0103] The manifold optimization unit 33 is in communication connection with the deformation-invariant feature unit 32, for analyzing the local curvature characteristics of the manifold and implementing a dense sampling strategy on high-curvature areas. The manifold optimization unit 33 optimizes the manifold structure in the following ways:

[0104] 1) Based on sample distribution, the local curvature characteristics of the manifold are analyzed, and high-curvature areas are identified. The local curvature can be calculated through the Riemann curvature tensor on the manifold:

[0105] ,

[0106] Among them, represents the Riemann curvature tensor, which is a fourth-order tensor, used to measure the degree of bending of the manifold; , and is a vector field on the manifold, representing the tangent vector on the manifold; is a connection on the manifold, a differential operator that computes the derivative of a vector field; is the Lie bracket of vector fields, computing the commutator of two vector fields. High curvature regions (curvature value > 0.8) usually correspond to decision boundaries of different classes of diseases, requiring denser sampling and more refined modeling.

[0107] 2) Implement a dense sampling strategy in high curvature regions to enhance the model's learning of decision boundaries. Specifically, the sampling density in high curvature regions is set to 2-3 times that of normal regions, ensuring that the model can accurately capture class boundary features.

[0108] 3) Introduce manifold smoothing regularization to prevent overfitting and noise interference. Manifold smoothing regularization can be represented as:

[0109] ,

[0110] where, is a non-negative real number; is a function defined on the manifold , i.e., the feature representation; is the gradient operator on the manifold, computing the directional derivative of a function on the manifold; is the norm of a vector, usually the Euclidean norm; is the volume element on the manifold; is the integral over the manifold . By minimizing this regularization term, the feature representation can be made smoother on the manifold, reducing noise interference.

[0111] 4) Construct an adaptive learning rate scheduler, using different learning rates for high curvature regions and low curvature regions. Specifically, high curvature regions (curvature value > 0.8) use a smaller learning rate (0.001-0.005) to ensure accurate convergence; low curvature regions (curvature value < 0.3) use a larger learning rate (0.01-0.05) to speed up the training process.

[0112] The feature conversion unit 34 is communicatively connected to the manifold optimization unit 33, used to define the geodesic path in the feature space, realizing optimal feature conversion between different observation conditions. Specifically, the geodesic path can be represented as the shortest path between two points on the manifold, with the calculation formula as follows:

[0113] ,

[0114] where, is the path from point to point , This is a path parameter, and its value range is [0, 1]. It represents a metric on a manifold, which is a second-order covariant tensor used to measure distances on the manifold; Representing a path The metric tensor at that location; The tangent vector of the path is represented, which is the derivative of the path. Indicates the length of the path tangent vector; Indicates to Integrals from 0 to 1; This represents the search for the path γ that minimizes the integral. Using geodesic paths, smooth feature transformations can be achieved between different observation conditions, maintaining semantic consistency of the features.

[0115] Through the above methods, the feature space optimization module 3 significantly improves the system's adaptability to complex field environments, and keeps the fluctuation of recognition accuracy under different lighting and angle conditions within ±3%.

[0116] Reference Figure 5 The multimodal feature fusion module 4 includes: a modality mapping unit 41, a semantic consistency unit 42, an entropy evaluation unit 43, and a decision integration unit 44.

[0117] Modality mapping unit 41 is used to construct mapping functions between different modality feature spaces. Specifically, for different modality data such as RGB images, multispectral images, and hyperspectral images, modality mapping unit 41 establishes the mapping relationship between their respective feature spaces, which can be expressed as:

[0118] ,

[0119] in, Indicates from modality To mode The mapping function is a nonlinear transformation; and Representing modes and modality The feature manifolds, corresponding to the feature representation spaces of different modalities, are obtained by minimizing the geodesic distance between corresponding points:

[0120] ,

[0121] in, This indicates a minimize operation; It is a set of corresponding points, containing point pairs with known intermodal correspondences; Represents a pair of corresponding points. From modality , From modality ; denotes the geodesic distance, which calculates the shortest path length between two points on a manifold; denotes the mapping of point from modality to modality after the transformation; denotes the summation over all corresponding point pairs. For example, the lesion features in the RGB image should be able to map to the corresponding feature locations in the multispectral image for the same lesion sample.

[0122] The semantic consistency unit 42 is communicatively connected with the modality mapping unit 41, and is configured to construct a multi-modal shared semantic space and ensure semantic consistency of features of different modalities. The semantic consistency unit 42 maintains semantic consistency of multi-modal features in the following ways:

[0123] 1) Construct a multi-modal shared semantic space, and all modal features are finally mapped to this space. The shared semantic space is denoted as , and each modal feature is mapped to the space by a mapping function , where is the shared semantic space, is the mapping function of modality to the shared semantic space.

[0124] 2) Design a semantic consistency evaluation index to quantify the semantic matching degree of features of different modalities. The evaluation index can be expressed as:

[0125] ,

[0126] wherein, denotes the semantic consistency between the feature of modality and the feature of modality , and the value range is [-1, 1]; and are the feature vectors of modality and modality , respectively; and are the functions of mapping modality and modality to the shared semantic space; denotes the inner product operation, which calculates the dot product of two vectors; denotes the norm of a vector, which is usually the Euclidean norm; the numerator represents the inner product of the two features in the shared semantic space, and the denominator represents the product of their norms. Essentially, this index is the cosine similarity of the feature vectors in the shared semantic space, and the greater the value, the higher the semantic consistency. In practical applications, when When two modal features are considered to have high semantic consistency, it is assumed that they are highly semantically consistent. At that time, it was determined that the semantics of the two modal features were inconsistent, and the mapping function needed to be further adjusted.

[0127] 3) Implement semantic alignment optimization to make the feature representations of the same disease in different modalities more consistent. Semantic alignment is achieved by minimizing the following objective function:

[0128] ,

[0129] in, This indicates a minimize operation; This is a set of positive sample pairs, representing different modal feature pairs of the same disease; This represents a pair of corresponding feature vectors; This represents the degree of semantic inconsistency, with a value range of [0, 2]. This represents the summation over all positive sample pairs. Through this optimization, the feature representations of the same disease under different modalities will tend to be consistent, improving the fusion effect.

[0130] 4) Establish a multi-level semantic hierarchy to support the identification needs of different granularities of disease categories, subtypes, and severity. Specifically, a three-level semantic hierarchy is constructed: the first level represents the major disease categories (e.g., fungal diseases, bacterial diseases, insect pests, etc.); the second level represents specific disease types (e.g., corn leaf blight, corn rust, etc.); and the third level represents disease severity (mild, moderate, severe). Different modalities of data are semantically aligned at each level to form a multi-granular semantically consistent representation.

[0131] Entropy evaluation unit 43 is communicatively connected to semantic consistency unit 42, and is used to evaluate the information richness of different modal features through information entropy, and dynamically adjust the weights of each modality. Specifically, for the features of modality $i$... Its information entropy can be expressed as:

[0132] ,

[0133] in, Representation of features The information entropy is a non-negative real number, with the unit being bits; The feature represents the first The probability distribution of each interval is a real number in the range [0, 1] that satisfies... ; Represents the logarithm to the base 2; The range number is a positive integer, usually set to 10-20; This indicates a summation operation.

[0134] Based on information entropy, dynamically adjust the weights of each modality:

[0135] ,

[0136] wherein, is the weight of the modality is a real number in the range of [0, 1] satisfying ; denotes the exponential function; is the adjustment parameter, which is a positive real number, usually set to 1.0-2.5, controlling the degree of influence of entropy on the weight; is the information entropy of the modality feature; is the total number of modalities, which is a positive integer, usually 2-4; the denominator is a normalization factor to ensure that the sum of all weights is 1. The higher the information entropy, the richer the information contained in the feature, and the greater the corresponding weight. For example, for samples under good lighting conditions, the information entropy of the RGB image feature is usually high, and the weight will increase accordingly; while for samples under weak light conditions, the multispectral image feature may have higher information entropy and weight.

[0137] The decision integration unit 44 is in communication connection with the entropy evaluation unit 43, and is used to establish a multi-level decision framework to realize reliability evaluation and error correction of the fusion result. Specifically, the decision integration unit 44 constructs a three-stage decision framework:

[0138] The first stage is preliminary screening, which quickly screens potential disease areas, and sets a lower decision threshold (0.4-0.5) to ensure high recall rate;

[0139] The second stage is accurate classification, which classifies the preliminary screening results in fine granularity, and sets a higher decision threshold (0.65-0.8) to ensure high accuracy;

[0140] The third stage is verification, which re-detects and cross- verifies samples with low confidence (usually confidence between 0.65-0.85) from multiple angles to further improve the classification reliability.

[0141] At the same time, the decision integration unit 44 realizes an error correction feedback loop, which includes misjudged samples into the incremental learning process, and continuously optimizes the model performance. For misjudged samples, especially false negative samples (missed detection), a higher learning weight (usually 2-3 times that of normal samples) is given to improve the recognition ability of difficult classification samples.

[0142] Through the above method, the multi-modal feature fusion module 4 realizes the intelligent cooperation of RGB, multispectral and hyperspectral data, and improves the corn disease and pest recognition accuracy by 15-20 percentage points, especially the recognition accuracy of corn borer, corn armyworm and other pests to 97%.

[0143] Referring toFigure 6 The spatial distribution analysis module 6 includes a spatial interpolation unit 61, a spatio-temporal evolution unit 62, a clustering analysis unit 63, and a transmission path analysis unit 64.

[0144] The spatial interpolation unit 61 is configured to construct a set of discrete points of disease distribution based on the detection results, and apply an adaptive spatial interpolation algorithm to generate a continuous distribution density map. Specifically, first, the pest identification results are combined with the GPS positioning data to generate a set of discrete points containing geographic coordinates wherein is the geographic coordinates, in meters or latitude and longitude; is the disease severity value, ranging from [0, 1]; is the number of sampling points; represents the set of all sampling points. Then, according to the sampling point density and distribution characteristics, the most suitable spatial interpolation method is automatically selected: for uniformly distributed high-density sampling points, Kriging interpolation is used; for non-uniformly distributed medium-density sampling points, IDW (Inverse Distance Weighted) interpolation is used; for sparse sampling points, Spline interpolation is used. The interpolation results form a continuous disease distribution density map, with a spatial resolution of typically 5-10 meters.

[0145] The spatio-temporal evolution unit 62 is communicatively connected to the spatial interpolation unit 61, and is configured to construct a disease transmission model to predict the disease spread trend. Specifically, based on multi-temporal disease distribution data, a spatio-temporal evolution model is established:

[0146] ,

[0147] wherein, represents the position at time the rate of change of disease severity, in ; represents the position at time disease severity, ranging from [0, 1]; is the diffusion coefficient, in , typically between 0.1-1.0, controlling the spatial diffusion rate of the disease; is the Laplacian operator, representing the second-order spatial derivative, calculating the spatial diffusion term; is the growth rate, in , typically between 0.1-0.5, controlling the natural growth rate of the disease; is the environmental capacity, dimensionless, typically set to 1.0, representing the upper limit of disease severity; This represents the ratio of disease severity to environmental carrying capacity. The model is essentially a diffusion equation with a Logistic growth term, capable of simulating the spread and growth process of diseases. Diffusion coefficient. The diffusion coefficient of fungal diseases will increase under high humidity conditions and environmental conditions (such as wind speed and humidity). Under windy conditions, the diffusion coefficient of insect pests will change with the wind direction.

[0148] Clustering analysis unit 63 is communicatively connected to spatial interpolation unit 61 to achieve multi-scale spatial clustering and identify high-incidence areas of diseases. Specifically, a density-based spatial clustering algorithm (such as DBSCAN) is used to analyze disease distribution data and identify high-density areas. Clustering parameters are adaptively adjusted according to different spatial scales: for field scale (1-10 hectares), a smaller neighborhood radius (10-30 meters) and a higher density threshold are set; for region scale (10-100 hectares), a larger neighborhood radius (50-100 meters) and a lower density threshold are set. Through multi-scale clustering analysis, high-incidence areas of diseases at different scales can be identified, providing a basis for precise prevention and control.

[0149] The propagation path analysis unit 64 is communicatively connected to the spatiotemporal evolution unit 62 and the clustering analysis unit 63, and is used to analyze the main pathways of disease propagation in conjunction with environmental factors. Specifically, it integrates environmental factor data such as topography, wind direction, temperature, and humidity to construct a propagation path analysis model.

[0150] ,

[0151] in, Indicates from position propagation to location The probability of [0, 1] is given; Represents the propagation probability function; The distance between two points is expressed in meters. The angle between the line connecting the two points and the wind direction is expressed in degrees and ranges from [0, 180]. Wind speed, in m / s; Humidity, expressed as a percentage (%). Temperature, in °C. When the wind direction is the same as the direction of propagation ( When the wind speed is close to zero, the probability of transmission increases; when the wind speed is high (e.g., >3 m / s), the probability of long-distance transmission increases; when the temperature and humidity are suitable (e.g., temperature 20-30°C, humidity >70%), the pathogen activity is enhanced, and the probability of transmission increases. This model can identify the main transmission routes and potential risk areas of pests and diseases, providing guidance for preventative control.

[0152] Through the above method, the spatial distribution analysis module 6 realizes dynamic analysis and propagation trend prediction of the distribution of pests and diseases, and the prediction accuracy reaches more than 85%, providing a scientific basis for precise prevention and control.

[0153] With reference to Figure 7 The visual decision module 7 includes a multi-layer visualization unit 71, an interactive control unit 72, a risk assessment unit 73, and a prevention and control suggestion generation unit 74.

[0154] The multi-layer visualization unit 71 is used to construct a multi-layer visualization interface based on a geographic information system, supporting multi-layer superimposed display of base maps, distribution maps, prediction maps, etc. Specifically, the bottom layer is a geographic base map, including satellite images and topographic maps; the middle layer is a pest and disease distribution map, which visually displays the disease distribution density in the form of a heat map, with the color gradually changing from green (no disease) to yellow (mild) to red (severe); the upper layer is a prediction map, which displays the disease transmission trend and potential risk areas. All layers are based on a unified geographic coordinate system to ensure accurate spatial correspondence. The multi-layer visualization unit 71 also supports time-series dynamic display, which can be controlled through a timeline to intuitively present the disease development process and help users understand the disease transmission pattern.

[0155] The interactive control unit 72 is in communication connection with the multi-layer visualization unit 71, and is used to realize the human-computer interaction functions of regional enlargement, layer control, condition query, etc. Specifically, the user can realize map zooming, panning and region selection through mouse or touch operation; adjust the display state and transparency of different layers through the layer control panel; and filter according to the conditions of disease type, time, severity, etc. In addition, the interactive control unit 72 also provides marking and annotation functions, allowing expert users to add supplementary information and prevention and control suggestions, promoting knowledge sharing and collaborative decision-making.

[0156] The risk assessment unit 73 is in communication connection with the multi-layer visualization unit 71, and is used to establish a disease impact assessment model to quantify the potential impact of diseases on yield. Specifically, based on the disease type and severity, combined with the corn growth stage and variety characteristics, the yield loss rate is predicted as follows:

[0157] ,

[0158] wherein, represents the predicted yield loss rate, ranging from [0, 1]; is the basic loss coefficient of disease type , ranging from [0, 1], indicating the maximum impact of the disease type on yield; is the infection area ratio of disease type , ranging from [0, 1], indicating the proportion of the infected area to the total area; is a time function related to the corn growth stage, ranging from [0, 1], represents the growth stage of corn; is a loss function related to disease severity, ranging from [0, 1], represents the disease severity; is the number of disease types; represents the summation operation. For example, for corn southern leaf blight, if infected at the seedling stage (growth stage coefficient about 0.3), mild infection (severity coefficient about 0.2), and the proportion of infected area is 5%, the base loss coefficient is 0.5, then the expected yield loss rate is about 0.3x0.2x5% x0.5=0.15%; if infected at the grain filling stage (growth stage coefficient about 0.8), severe infection (severity coefficient about 0.7), and the proportion of infected area is 30%, then the expected yield loss rate is about 0.8x0.7x30% x0.5=8.4%.

[0159] The risk assessment unit 73 also implements economic threshold analysis to assess the cost and benefit of prevention and control. The economic threshold calculation formula is:

[0160] ,

[0161] wherein, is the economic threshold, representing the disease occurrence degree, when the actual disease occurrence degree exceeds this value, it is recommended to carry out prevention and control; is the prevention and control cost, in yuan / ha, including pesticide cost, labor cost and mechanical cost, etc.; is the crop price, in yuan / kg; is the expected yield, in kg / ha; is the yield loss proportion caused by unit disease index, dimensionless, ranging from [0, 1]; is the prevention and control effect, dimensionless, ranging from [0, 1], representing the proportion of disease that can be reduced by prevention and control measures. When the actual disease occurrence degree exceeds the economic threshold, it is recommended to carry out prevention and control; otherwise, the prevention and control cost may be greater than the loss recovered. Preferably, the economic threshold reference values of common corn diseases and pests are as follows: the economic threshold of corn borer is 3-5 heads per hundred plants, the economic threshold of corn southern leaf blight is 5%-10% of the incidence rate, and the economic threshold of corn rust is 10%-15% of the incidence rate.

[0162] The prevention and control suggestion generation unit 74 is in communication connection with the risk assessment unit 73, and is used for generating partition prevention and control suggestions and optimal pesticide application paths based on disease distribution characteristics and severity. Specifically, according to the disease distribution density map and the risk assessment result, the farmland is divided into a high-risk area, a medium-risk area and a low-risk area, and different prevention and control strategies are adopted respectively: the high-risk area (disease occurrence degree exceeding 150% of the economic threshold) adopts an immediate prevention and control strategy, and 1.2-1.5 times of the conventional dose is used; the medium-risk area (disease occurrence degree of 80%-150% of the economic threshold) adopts a conventional prevention and control strategy, and the standard dose is used; and the low-risk area (disease occurrence degree less than 80% of the economic threshold) adopts a monitoring-based strategy, and does not prevent and control for the time being but closely monitors.

[0163] The prevention and control suggestion generation unit 74 also implements a pesticide application path optimization algorithm to provide the most efficient operation route. Specifically, the farmland is modeled as a grid map, and each grid point represents a position, and the edge represents a feasible travel path. The pesticide application path optimization problem is converted into a path planning problem on the grid map, and the goal is to find the shortest path that can cover all the areas that need to be applied. The improved Christofides algorithm is used to solve the problem, first a minimum spanning tree is constructed, then a minimum perfect matching is added to the odd-degree vertex, and finally an Euler circuit is constructed and a shortcut is optimized. The optimized path can usually reduce the travel distance by 10%-20%, improving the operation efficiency.

[0164] Through the above method, the visual decision module 7 realizes intuitive visualization of the distribution of diseases and pests and scientific decision support, reduces the pesticide use amount by 30%, improves the prevention and control effect by 25%, and significantly improves the economic benefit and environmental benefit.

[0165] The overall working process of the corn disease and pest intelligent identification and spatial distribution visualization system provided by the application is as follows:

[0166] 1) The data acquisition module 1 collects aerial images through the unmanned aerial vehicle acquisition unit 11 and collects high-precision close-range images through the ground mobile platform 12, and the data preprocessing unit 13 is used for preprocessing to generate standardized image data;

[0167] 2) The multi-scale feature representation module 2 receives the standardized image data, and through the multi-scale processing unit 21, the feature descriptor generation unit 22, the manifold mapping unit 23 and the scale association unit 24, the image data is mapped to a high-dimensional manifold space to form a feature representation;

[0168] 3) The feature space optimization module 3 receives the feature representation, and through the metric construction unit 31, the deformation invariant feature unit 32, the manifold optimization unit 33 and the feature conversion unit 34, an adaptive metric structure is constructed to generate deformation invariant features;

[0169] 4) The multi-modal feature fusion module 4 receives the multi-scale feature representation and the deformation-invariant feature, establishes the mapping relationship between the multi-modal data through the modal mapping unit 41, the semantic consistency unit 42, the entropy evaluation unit 43 and the decision integration unit 44, fuses different modal features, and generates a fusion feature;

[0170] 5) The pest and disease identification module 5 identifies the corn pest and disease type and position information based on the fusion feature;

[0171] 6) The spatial distribution analysis module 6 receives the corn pest and disease type and position information, generates pest and disease spatial distribution data through the spatial interpolation unit 61, the spatio-temporal evolution unit 62, the clustering analysis unit 63 and the propagation path analysis unit 64;

[0172] 7) The visual decision module 7 receives the pest and disease spatial distribution data, generates pest and disease distribution visualization results and prevention and control decision suggestions through the multi-layer visualization unit 71, the interactive control unit 72, the risk assessment unit 73 and the prevention and control suggestion generation unit 74.

[0173] In one specific application scenario, the system monitors a certain corn planting area and finds early symptoms of corn small spot disease. Through multi-scale feature representation and feature space optimization, the system successfully identifies early disease with a lesion area of less than 1%, with a disease detection rate of 88%. The spatial distribution analysis module constructs a disease distribution map based on the identification results and predicts the diffusion trend in the next 7 days. The visual decision module generates zoned prevention and control suggestions, recommending immediate prevention and control in high-risk areas with 1.2 times the regular dose; prevention and control in medium-risk areas within 3 days with the standard dose; and no prevention and control in low-risk areas but enhanced monitoring. Field test results show that the area that follows the system's recommendations for prevention and control uses 35% less pesticide than the area that uses traditional uniform pesticide application, has a 22% improvement in prevention and control effect, and has a significant increase in economic benefit.

[0174] The corn pest and disease intelligent identification and spatial distribution visualization system provided by the application has been verified in multiple corn producing areas, and the main results are as follows:

[0175] 1) Early identification capability: In the early stage with a lesion area of less than 1%, the system has an average detection rate of 87% for corn small spot disease, rust and leaf spot disease, which is 45 percentage points higher than that of traditional methods. Through the multi-scale feature representation technology of manifold learning, the system can effectively capture the characteristics of small lesions and detect diseases 7-10 days in advance.

[0176] 2) Environmental adaptability: The system's recognition accuracy fluctuates within ±3% under different lighting conditions (overcast, sunny, evening) and different angles (vertical, 30° inclined, 60° inclined), demonstrating excellent environmental adaptability. This is mainly due to the deformation-invariant feature extraction and geometric constraint optimization strategies in the feature space optimization module.

[0177] 3) Multi-modal fusion effect: The system enhances the overall recognition accuracy of corn diseases and pests from 78% in single modal to 95% by intelligently fusing RGB, multispectral, and hyperspectral data, especially for corn borer and corn leaf roller, with an accuracy of 97%. The semantic consistency preservation and information entropy-guided adaptive fusion mechanism in the multi-modal feature fusion module are key.

[0178] 4) Spatial analysis accuracy: The system's prediction accuracy for spatial distribution of diseases and pests is 86%, and the prediction accuracy for transmission trend is 83%, effectively guiding prevention and control decisions. The adaptive spatial interpolation and transmission path analysis combined with environmental factors in the spatial distribution analysis module are core advantages.

[0179] 5) Prevention and control benefits: In practical applications, the regional prevention and control measures guided by the system reduce pesticide use by an average of 32%, improve prevention and control effects by 24%, and increase economic benefits by about 15% compared to traditional prevention and control methods. The precise prevention and control scheme and optimized pesticide application path generated by the visual decision module are key factors in improving prevention and control benefits.

[0180] In summary, the corn disease and pest intelligent recognition and spatial distribution visualization system provided by the present application solves the problems of early recognition difficulty, poor environmental adaptability, insufficient multi-source data fusion, and lack of spatial distribution analysis by innovatively applying differential geometry theory to construct a multi-scale multi-modal feature fusion architecture, achieving full-process intelligentization from intelligent recognition to spatial distribution analysis to visual decision support for corn diseases and pests, with significant technical innovation and practical value.

[0181] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A system for intelligent identification and spatial distribution visualization of corn diseases and pests, characterized in that, include: The data acquisition module is used to acquire multi-source image data of corn. A multi-scale feature representation module, which is communicatively connected to the data acquisition module, is used to receive the multi-source image data and map the multi-source image data to a high-dimensional manifold space to form a feature representation. The feature space optimization module is communicatively connected to the multi-scale feature representation module, and is used to receive the feature representation, construct an adaptive metric structure, and generate deformation-invariant features. The multimodal feature fusion module is communicatively connected to the multi-scale feature representation module and the feature space optimization module, respectively, and is used to establish the mapping relationship between multimodal data, fuse different modal features, and generate fused features; The pest and disease identification module is communicatively connected to the multimodal feature fusion module and is used to identify the type and location information of corn pests and diseases based on the fused features. The spatial distribution analysis module is communicatively connected to the pest and disease identification module and is used to receive the corn pest and disease type and location information and generate pest and disease spatial distribution data. The visualization decision-making module is communicatively connected to the spatial distribution analysis module and is used to receive the spatial distribution data of pests and diseases, and generate visualization results of pest and disease distribution and prevention and control decision-making suggestions.

2. The intelligent identification and spatial distribution visualization system for maize diseases and pests according to claim 1, characterized in that, The data acquisition module includes: The drone data acquisition unit is used to carry a multispectral camera and a positioning system to collect aerial image data of corn fields. Ground-based mobile platform for acquiring high-resolution RGB images and close-range spectral data; The data preprocessing unit is communicatively connected to the UAV acquisition unit and the ground mobile platform, and is used to preprocess the acquired image data to generate standardized image data. The task scheduling unit is communicatively connected to the UAV data collection unit and the ground mobile platform. It is used to automatically plan the data collection path according to the field conditions and arrange close-range sampling for detected suspicious areas.

3. The intelligent identification and spatial distribution visualization system for maize diseases and pests according to claim 1, characterized in that, The multi-scale feature representation module includes: A multi-scale processing unit is used to process the input image data in different spatial resolutions at different levels. The feature descriptor generation unit is communicatively connected to the multi-scale processing unit and is used to construct a set of feature descriptors for each resolution level. The manifold mapping unit, which is communicatively connected to the feature descriptor generation unit, is used to map high-dimensional features to a low-dimensional manifold space while preserving the inherent structural relationships of the data. The scale association unit is communicatively connected to the manifold mapping unit and is used to construct an inter-scale association graph and automatically adjust the weights of features at different scales according to the disease type.

4. The intelligent identification and spatial distribution visualization system for maize diseases and pests according to claim 1, characterized in that, The feature space optimization module includes: A metric building unit, used to construct a local adaptive metric matrix based on sample distribution characteristics; The deformation-invariant feature unit is communicatively connected to the metric construction unit and is used to establish an illumination change model and a viewpoint transformation set to extract deformation-invariant features. The manifold optimization unit, which is communicatively connected to the deformation-invariant feature unit, is used to analyze the local curvature characteristics of the manifold and implement a dense sampling strategy for high curvature regions. The feature transformation unit, which is communicatively connected to the manifold optimization unit, is used to define geodesic paths in the feature space and achieve optimal feature transformation between different observation conditions.

5. The intelligent identification and spatial distribution visualization system for maize diseases and pests according to claim 1, characterized in that, The multimodal feature fusion module includes: Modality mapping unit, used to construct mapping functions between different modal feature spaces; The semantic consistency unit, which is communicatively connected to the modality mapping unit, is used to construct a multimodal shared semantic space to ensure the semantic consistency of features of different modalities. The entropy evaluation unit, which is communicatively connected to the semantic consistency unit, is used to evaluate the information richness of different modal features through information entropy and dynamically adjust the weights of each modality. The decision integration unit is communicatively connected to the entropy evaluation unit and is used to establish a multi-level decision framework to achieve reliability evaluation and error correction of the fusion results.

6. The intelligent identification and spatial distribution visualization system for maize diseases and pests according to claim 1, characterized in that, The spatial distribution analysis module includes: Spatial interpolation unit is used to construct a discrete point set of disease distribution based on the detection results, and to generate a continuous distribution density map using an adaptive spatial interpolation algorithm; The spatiotemporal evolution unit is communicatively connected to the spatial interpolation unit and is used to construct a disease transmission model and predict the disease spread trend. The clustering analysis unit is communicatively connected to the spatial interpolation unit and is used to achieve multi-scale spatial clustering to identify high-incidence areas of diseases. The propagation path analysis unit is communicatively connected to the spatiotemporal evolution unit and the clustering analysis unit, and is used to analyze the main pathways of disease propagation in conjunction with environmental factors.

7. The intelligent identification and spatial distribution visualization system for maize diseases and pests according to claim 1, characterized in that, The visualization decision-making module includes: Multi-layer visualization unit is used to build a multi-layer visualization interface based on geographic information system, supporting the overlay display of multiple layers such as base map, distribution map, and prediction map; An interactive control unit is communicatively connected to the multi-layer visualization unit and is used to realize human-computer interaction functions such as area zooming, layer control, and conditional query. The risk assessment unit, which is communicatively connected to the multi-layer visualization unit, is used to establish a disease impact assessment model and quantify the potential impact of diseases on yield. The prevention and control suggestion generation unit is communicatively connected to the risk assessment unit and is used to generate regional prevention and control suggestions and optimal application routes based on the disease distribution characteristics and severity.

8. The intelligent identification and spatial distribution visualization system for maize diseases and pests according to claim 3, characterized in that, The scale association unit constructs inter-scale associations in the following manner: Construct an association graph consisting of nodes representing features at different scales and edges representing the strength of association; Design a scale-adaptive weighting function to automatically adjust the importance of features at different scales based on the type of disease; A mechanism for smooth transition in scale space is implemented to ensure continuous representation of features from coarse-grained to fine-grained. For early disease characteristics, dynamically increase the weight of fine-grained features; For diseases in the middle and late stages, balance the weights of multi-scale features.

9. The intelligent identification and spatial distribution visualization system for maize diseases and pests according to claim 4, characterized in that, The manifold optimization unit optimizes the manifold structure in the following ways: Based on the analysis of sample distribution, the local curvature characteristics of the manifold are analyzed to identify high curvature regions; A dense sampling strategy is implemented for high curvature regions to enhance the model's learning of decision boundaries; Manifold smoothing regularization is introduced to prevent overfitting and noise interference; Construct an adaptive learning rate scheduler that uses different learning rates for high curvature regions and low curvature regions respectively.

10. The intelligent identification and spatial distribution visualization system for maize diseases and pests according to claim 5, characterized in that, The semantic consistency unit maintains the semantic consistency of multimodal features in the following ways: Construct a multimodal shared semantic space, to which all modal features are ultimately mapped; Design a semantic consistency evaluation index to quantify the degree of semantic matching of different modal features; Implement semantic alignment optimization to make the feature representations of the same disease in different modalities tend to be consistent; Establish a multi-level semantic hierarchy to support the identification needs of different granularities of disease categories, disease subtypes, and disease severity.

Images