Crop disease and pest intelligent monitoring system based on multispectral imaging and method of using the same

Abstract

The present application relates to the field of agricultural information technology, in particular to a crop disease and pest intelligent monitoring system based on multispectral imaging and a use method thereof, comprising a multispectral data acquisition and preprocessing module, a visual intelligent analysis module, an intelligent diagnosis module and a decision support module.The present application realizes adaptive fusion and noise separation of multispectral features to generate sparse feature representation of focused disease and pest sensitive information by dynamically screening high contribution spectral bands through a cross-band attention mechanism and combining a 3D convolution network and a feature decoupling reconstruction algorithm constrained by mutual information, and based on a multi-model fusion algorithm of Inception-v3 and LightGBM cascade, combined with Monte Carlo Dropout confidence evaluation, Grad-CAM interpretability analysis and TCN time convolution network time series prediction, an intelligent diagnosis system with reliability quantification, decision visualization and trend prediction capability is constructed to realize accurate identification of disease and pest type and severity and dynamic prediction of future 5-day spread trend.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural information technology, in particular to a crop disease and pest intelligent monitoring system based on multi-spectral imaging and a use method thereof. BACKGROUND

[0002] Crop disease and pest refers to the abnormal state that crop physiological function, tissue structure and external morphology are damaged due to the invasion of harmful organisms (such as fungi, bacteria, viruses, pests, weeds, nematodes, etc.) or the influence of environmental factors during crop growth and development, and further affects crop yield, quality and even causes plant death. In modern agricultural production, early and accurate monitoring of crop disease and pest is the core link to ensure yield and quality. With the development of spectral detection technology, multi-spectral imaging has become an important means for disease and pest monitoring because it can capture the reflectivity difference of crops at different wavebands.

[0003] However, the traditional multi-spectral disease and pest monitoring technology has significant shortcomings in practical application: first, the waveband selection relies on manual experience, which is difficult to adapt to the spectral sensitivity characteristics of different diseases and pests, resulting in low feature extraction efficiency; second, the coupling problem of disease and pest features and background noise in multi-spectral data is prominent, and traditional algorithms cannot achieve effective separation, which seriously affects the recognition accuracy; third, the diagnosis model lacks visual decision basis, which is difficult to meet the demand of agricultural production for the explainability of monitoring results; fourth, it can only identify the current state, and cannot predict the disease development trend based on time series analysis, resulting in lagging prevention measures.

[0004] Therefore, the present application provides a crop disease and pest intelligent monitoring system based on multi-spectral imaging and a use method thereof to solve the above technical problems. SUMMARY

[0005] The present application aims to provide a crop disease and pest intelligent monitoring system based on multi-spectral imaging and a use method thereof. The present application dynamically selects high-contribution spectral wavebands based on cross-waveband attention mechanism, and combines 3D convolution network and feature decoupling reconstruction algorithm with mutual information constraint to realize adaptive fusion and noise separation of multi-spectral features to generate sparse feature representation focusing on disease and pest sensitive information. Based on the multi-model fusion algorithm of Inception-v3 and LightGBM cascade, combined with Monte Carlo Dropout confidence evaluation, Grad-CAM explainability analysis and TCN time convolution network time series prediction, an intelligent diagnosis system with reliability quantification, decision visualization and trend prediction ability is constructed to realize accurate identification of disease and pest type and severity and dynamic prediction of future 5-day spread trend.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] The application provides a crop disease and pest intelligent monitoring system based on multispectral imaging, comprising a multispectral data acquisition and preprocessing module, a visual intelligent analysis module, an intelligent diagnosis module, and a decision support module, wherein:

[0008] The multispectral data acquisition and preprocessing module is used for acquiring multispectral images of crops through a multispectral imaging device and preprocessing the acquired original images.

[0009] The visual intelligent analysis module is used for dynamically screening waveband contribution degrees through a cross-waveband adaptive feature fusion network, separating disease and pest sensitive features from background noise by using a feature decoupling reconstruction algorithm, and generating sparse feature representation.

[0010] The intelligent diagnosis module is based on a multi-model fusion algorithm with confidence evaluation and interpretability, accurately identifies the type and severity of diseases and pests, and predicts the development trajectory of diseases in combination with time series analysis.

[0011] The decision support module is used for generating an executable scheme containing pesticide dosage, operation time, and recommended pesticide in combination with diagnosis results and an agronomic knowledge base, and visualizing disease spatial distribution through a GIS map.

[0012] The multispectral data acquisition and preprocessing module comprises a spectral acquisition unit, a radiation correction unit, a geometric registration unit, and a noise suppression unit, wherein:

[0013] The spectral acquisition unit is used for synchronously acquiring image data of crops in multiple wavebands through a multispectral camera array, covering a disease and pest sensitive spectral range of 400-1000nm.

[0014] The radiation correction unit is based on standard reflector and dark current calibration to eliminate sensor response errors.

[0015] The geometric registration unit is used for multi-waveband image sub-pixel level alignment by using a SIFT feature matching algorithm, and the registration error is less than 0.3 pixels.

[0016] The noise suppression unit is used for removing Gaussian noise and salt and pepper noise in the image by using a method combining non-local mean filtering and wavelet transform.

[0017] The multispectral camera array is configured to synchronously acquire image data of at least 5 spectral wavebands, including but not limited to 450nm blue light, 550nm green light, 650nm red light, 760nm near-infrared, and 850nm short-wave infrared.

[0018] The visual intelligent analysis module comprises a waveband screening unit, a feature fusion unit, a decoupling reconstruction unit, and a sparse representation unit, wherein:

[0019] The waveband screening unit is configured to calculate spectral weights through a cross-waveband attention mechanism, and dynamically screen wavebands with high contribution to disease and pest identification.

[0020] The feature fusion unit is configured to process the spatial-spectral joint features by using a 3D convolution network to generate a fusion feature map containing multi-dimensional information.

[0021] The decoupling reconstruction unit is configured to separate disease and pest features from background noise based on an encoder-decoder architecture, and independently represent the disease and pest features by mutual information constraint.

[0022] The sparse representation unit is configured to compress feature dimensions by L1 regularization to generate a sparse feature vector focusing on sensitive information of the disease and pests.

[0023] The waveband screening unit calculates spectral weights through a cross-waveband attention mechanism, and dynamically screens wavebands with high contribution to disease and pest identification, and the specific operation is as follows:

[0024] A1: Process each waveband reflectance curve by a 1D convolution layer to generate a waveband feature vector f i ∈R d ;

[0025] A2: Calculate the waveband weight α i , and the specific formula is as follows:

[0026]

[0027] In the formula, f i is the i-th waveband feature vector, is the mean feature vector of all wavebands, W q , and W k are learnable parameter matrices.

[0028] A3: Retain the wavebands with weight α i >0.15 to form an optimized waveband combination.

[0029] The decoupling reconstruction unit separates disease and pest features from background noise based on an encoder-decoder architecture, and independently represents the disease and pest features by mutual information constraint, and the specific operation is as follows:

[0030] B1: Encoder feature separation:

[0031] ①A double-branch encoder network E=[E s ,E r ] is adopted, wherein E s extracts spatial features z s , and E r extracts spectral features z r ;

[0032] ② input multispectral image Through the convolutional layer and the fully connected layer to map to the latent space: z s = E s (X), z r = E r (X), where, D s , D r is the feature dimension;

[0033] B2: Mutual information constraint:

[0034] ①Decoupling by minimizing the mutual information between the latent feature z p and the background noise z b : where the mutual information I(·) is approximated by a neural network;

[0035] ②Introducing an adversarial training mechanism: constructing a discriminator D to distinguish the joint distribution p(z s ,z r ) and the marginal distribution product p(z s )p(z r ), the loss function is:

[0036] B3: Decoder reconstruction:

[0037] ①Design a condition generator G(z s ,z r ) based on the conditional adversarial network structure, reconstruct the feature atlas through z s and z r

[0038] ②Use a multi-scale reconstruction loss function, including:

[0039] a. Pixel-level L1 loss:

[0040] b. Perceptual loss: where VGG is a pre-trained visual feature extraction network;

[0041] B4: Decoupling evaluation and optimization:

[0042] ①Calculate the decoupling degree index When D < θ, trigger parameter update;

[0043] ②The joint optimization objective function is: L = L L1 + λ1L per + λ2L adv - λ3I(z​s ; z r ), wherein λ1=100, λ2=1, λ3=0.1 are weight coefficients.

[0044] The intelligent diagnosis module comprises a multi-model cascading unit, a confidence assessment unit, an explainability analysis unit, and a time series prediction unit, wherein:

[0045] The multi-model cascading unit is used to cascade Inception-v3 and LightGBM for end-to-end pest and disease identification.

[0046] The confidence assessment unit is used to calculate the uncertainty of the prediction result based on Monte Carlo Dropout and output a 0-1 confidence score.

[0047] The explainability analysis unit is used to generate a heat map through Grad-CAM to visualize the key decision-making area.

[0048] The time series prediction unit is used to analyze historical data using a TCN time convolution network to predict the change rate of disease spread area in the next 5 days.

[0049] The TCN time convolution network is used in the time series prediction unit to analyze historical data and predict the change rate of disease spread area in the next 5 days, with the following specific operations:

[0050] C1: Collect pest and disease identification results, meteorological data, and soil moisture information in the past 14 days to construct a time series input vector.

[0051] C2: Input into a TCN network comprising multiple dilated causal convolution layers, wherein the dilation coefficients of each layer are exponentially increased to expand the receptive field.

[0052] C3: The network output layer generates a vector of length 5 representing the daily change rate of disease spread area in the next 5 days.

[0053] C4: The TCN network is trained by minimizing the following function:

[0054]

[0055] where y t is the true spread rate, is the predicted value, γ is the regularization coefficient, and W is the network parameter.

[0056] C5: The prediction result is used to assist in formulating a precise prevention and control operation plan.

[0057] The decision support module comprises an agronomic knowledge matching unit, a dosage optimization unit, a spatio-temporal decision generation unit, and a multi-objective optimization unit, wherein:

[0058] The agronomic knowledge matching unit is used for calling disease and pest control thresholds and standard schemes from the knowledge base according to crop varieties and growth periods;

[0059] The dose optimization unit is used for dynamically calculating pesticide application doses based on disease spot density and severity, and balancing control effect and environmental cost;

[0060] The spatio-temporal decision generation unit is used for mapping diagnosis results to GIS maps, and generating pesticide application time and area planning schemes with spatial heat maps;

[0061] The multi-objective optimization unit is used for generating trade-off decision schemes among control efficiency, cost input and ecological impact through a Pareto optimal algorithm.

[0062] The present application also proposes a crop disease and pest intelligent monitoring method based on multi-spectral imaging, comprising the following steps:

[0063] S1: crop images in the 400-1000nm waveband are collected through a multi-spectral camera array, high-quality multi-spectral data sets are generated through radiation correction, geometric registration and noise suppression processing;

[0064] S2: cross-waveband attention mechanism is used to dynamically screen high-contribution spectral wavebands, 3D convolution network is used to fuse spatial-spectral features, and disease and pest sensitive sparse feature representation is generated through decoupling reconstruction and L1 regularization;

[0065] S3: Inception-v3 and LightGBM cascade models are used to identify disease and pest types and severity, Monte Carlo Dropout is used to evaluate prediction confidence, and TCN time convolution network is used to predict disease spread trend in the next 5 days;

[0066] S4: pesticide application dose and operation time are dynamically calculated according to diagnosis results and agronomic knowledge base, disease control schemes with spatial heat maps are generated through GIS maps, and Pareto optimal algorithm is used to balance control efficiency, cost and ecological impact.

[0067] Compared with the prior art, the present application has the following beneficial effects:

[0068] The application realizes adaptive fusion and noise separation of multi-spectral features to generate sparse feature representation of focused pest sensitive information by dynamically screening high contribution spectral bands through cross-band attention mechanism and combining 3D convolution network and feature decoupling reconstruction algorithm constrained by mutual information, and based on multi-model fusion algorithm of Inception-v3 and LightGBM cascade, combined with Monte Carlo Dropout confidence evaluation, Grad-CAM interpretability analysis and TCN time convolution network time series prediction, an intelligent diagnosis system with reliability quantification, decision visualization and trend prediction ability is constructed to realize accurate identification of pest type and severity and dynamic prediction of 5-day spread trend in the future. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1 It is a system diagram of the crop pest intelligent monitoring system based on multi-spectral imaging of the application.

[0070] Figure 2 It is a flowchart of the crop pest intelligent monitoring method based on multi-spectral imaging of the application.

[0071] Figure 3 It is a hardware deployment schematic diagram in the crop pest intelligent monitoring system based on multi-spectral imaging and the use method thereof of the application.

[0072] BRIEF DESCRIPTION OF DRAWINGS:

[0073] 100, multi-spectral data acquisition and preprocessing module; 101, spectral acquisition unit; 102, radiation correction unit; 103, geometric registration unit; 104, noise suppression unit; 200, visual intelligent analysis module; 201, band screening unit; 202, feature fusion unit; 203, decoupling reconstruction unit; 204, sparse representation unit; 300, intelligent diagnosis module; 301, multi-model cascade unit; 302, confidence evaluation unit; 303, interpretability analysis unit; 304, time series prediction unit; 400, decision support module; 401, agronomic knowledge matching unit; 402, dose optimization unit; 403, spatio-temporal decision generation unit; 404, multi-objective optimization unit. DETAILED DESCRIPTION

[0074] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0075] Embodiment 1:

[0076] As Figure 1 andFigure 3 As shown, the embodiment provides a crop disease and pest intelligent monitoring system based on multispectral imaging, which comprises a multispectral data acquisition and preprocessing module 100, a visual intelligent analysis module 200, an intelligent diagnosis module 300, and a decision support module 400. The multispectral data acquisition and preprocessing module 100 is used to acquire multispectral images of crops through a multispectral imaging device and preprocess the acquired original images. The visual intelligent analysis module 200 is used to dynamically screen waveband contribution through a cross-waveband adaptive feature fusion network, separate disease and pest sensitive features from background noise by using a feature decoupling reconstruction algorithm, and generate sparse feature representation. The intelligent diagnosis module 300 is based on a multi-model fusion algorithm with confidence evaluation and explainability, accurately identifies the type and severity of diseases and pests, and predicts the development trajectory of diseases in combination with time series analysis. The decision support module 400 is used to generate an executable scheme containing pesticide dosage, operation time, and recommended pesticide in combination with the diagnosis result and an agronomic knowledge base, and visualize the spatial distribution of diseases through a GIS map.

[0077] It should be noted that the multispectral data acquisition and preprocessing module 100 acquires and preprocesses multispectral images of crops, the visual intelligent analysis module 200 dynamically screens sensitive wavebands and decouples disease and pest features, the intelligent diagnosis module 300 fuses multiple models to achieve accurate identification and trend prediction of diseases and pests, and finally the decision support module 400 generates a visual pesticide application scheme in combination with agronomic knowledge.

[0078] In the embodiment, it should also be noted that the multispectral data acquisition and preprocessing module 100 comprises a spectral acquisition unit 101, a radiation correction unit 102, a geometric registration unit 103, and a noise suppression unit 104. The spectral acquisition unit 101 is used to synchronously acquire image data of crops in multiple wavebands through a multispectral camera array, covering a disease and pest sensitive spectral range of 400-1000 nm. The multispectral camera array is configured to synchronously acquire image data of at least 5 spectral wavebands, including but not limited to 450 nm blue light, 550 nm green light, 650 nm red light, 760 nm near-infrared, and 850 nm short-wave infrared. The radiation correction unit 102 is based on standard reflector and dark current calibration to eliminate sensor response errors. The geometric registration unit 103 is used to perform multi-waveband image sub-pixel alignment by using a SIFT feature matching algorithm, with a registration error <0.3 pixels. The noise suppression unit 104 is used to remove Gaussian noise and salt and pepper noise in the image by using a method combining non-local mean filtering and wavelet transform.

[0079] It should be noted that the spectral acquisition unit 101 synchronously acquires crop images in at least five characteristic wavebands (including 450 nm / 550 nm / 650 nm / 760 nm / 850 nm) in the range of 400-1000 nm, eliminates sensor errors by the radiation correction unit 102, realizes sub-pixel level multi-waveband alignment by the geometric registration unit 103, and finally removes noise by fusing non-local mean and wavelet transform by the noise suppression unit 104, to complete the whole process of preprocessing from raw data to a standardized spectral feature matrix.

[0080] Further, it should be noted that the DJI Zenmuse P1 multi-spectral camera is used to synchronously acquire five characteristic wavebands (450 nm blue light, 550 nm green light, 650 nm red light, 760 nm near-infrared, and 850 nm short-wave infrared) in the range of 400-1000 nm, and the ground sampling distance reaches 2.5 cm, meeting the requirements of crop leaf-level disease identification. The radiation correction unit 102 deploys a SpectraVue standard reflector in the field, and acquires calibration data every 2 hours, and corrects the sensor response by the formula , wherein I(λ) is the original pixel value, S(λ) is the sensor gain, and Ref(λ) is the standard reflectivity; the noise suppression unit 104 first removes Gaussian noise by using non-local mean filtering (search window 11x11 and similar window 5x5), and then suppresses salt and pepper noise by wavelet transform (db4 wavelet, 3 layers of decomposition), so that the image signal-to-noise ratio is improved to more than 35 dB.

[0081] In the embodiment, it should also be noted that the visual intelligent analysis module 200 includes a waveband screening unit 201, a feature fusion unit 202, a decoupling reconstruction unit 203, and a sparse representation unit 204, wherein: the waveband screening unit 201 is used to calculate spectral weights by a cross-waveband attention mechanism, and dynamically screen wavebands with high contribution to disease and pest identification; the specific operation is as follows: A1: processing each waveband reflectance curve by a 1D convolution layer to generate a waveband feature vector f i ∈R d ; A2: calculating the waveband weight α i , and the specific formula is as follows:

[0082]

[0083] , wherein f i is the i-th waveband feature vector, is the mean feature vector of the waveband, and W q , W k is a learnable parameter matrix; A3: retaining the weight α i>0.15 bands constitute an optimized band combination. Feature fusion unit 202: used to process spatial-spectral joint features using a 3D convolutional network to generate a fused feature map containing multi-dimensional information; Decoupling and reconstruction unit 203: based on an encoder-decoder architecture, separates pest and disease features from background noise, and performs independent representation through mutual information constraints; specific operations are as follows: B1: Encoder feature separation: ① Use a dual-branch encoder network E = [E s E r In the formula, E s Extracting spatial features z s E r Extracting spectral features z r ② Input multispectral image Mapping to the latent space via convolutional layers and fully connected layers: z s =E s (X),z r =E r (X), where, D s D r B2: Feature dimension; Mutual information constraint: ① By minimizing the latent feature z p With background noise z b Decoupling through mutual information: In the formula, mutual information I(·) is approximated by a neural network; ② An adversarial training mechanism is introduced: a discriminator D is constructed to distinguish the joint distribution p(z) s ,z r The product of z and the marginal distribution p(z) s )p(z r The loss function is: B3: Decoder Reconstruction: ① Design Condition Generator G(z) s ,z r Based on the conditional adversarial network structure, through z s and z r Reconstructing Feature Maps ② Employ a multi-scale reconstruction loss function, including: a. Pixel-level L1 loss: b. Perceived loss: In the formula, VGG is a pre-trained visual feature extraction network; B4: Decoupling evaluation and optimization: ① Calculate the decoupling degree index Parameter updates are triggered when D < θ; ② The joint optimization objective function is: L = L L1 +λ1L per +λ2L adv -λ3I(z s ;z r), where λ1=100, λ2=1, λ3=0.1 are weight coefficients. The sparse representation unit 204 is configured to generate a sparse feature vector of focused pest-sensitive information by compressing the feature dimension through L1 regularization.

[0084] wherein it should be noted that the waveband screening unit 201 dynamically selects key spectral wavebands based on the cross-waveband attention mechanism, the spatial-spectral features are integrated by the feature fusion unit 202 using 3D convolution, the pest features and background noise are separated by the decoupling reconstruction unit 203 using an encoder-decoder architecture and mutual information constraint, and finally the sparse feature vector with high discriminability is generated by the sparse representation unit 204 through L1 regularization.

[0085] Further, it should be noted that the waveband feature vector f i ∈R d wherein d=64. The dual-branch encoder network of the decoupling reconstruction unit 203 adopts a ResNet18 architecture, and the spatial branch extracts the lesion edge texture features (such as the yellow rust-like texture of stripe rust), and the spectral branch captures the feature of the decrease in near-infrared reflectance caused by chlorophyll decay.

[0086] In the present embodiment, it should also be noted that the intelligent diagnosis module 300 includes a multi-model cascading unit 301, a confidence evaluation unit 302, an explainability analysis unit 303, and a time series prediction unit 304, wherein: the multi-model cascading unit 301 is configured to cascade Inception-v3 and LightGBM for end-to-end pest identification; the confidence evaluation unit 302 is configured to calculate the uncertainty of the prediction result based on Monte Carlo Dropout and output a 0-1 confidence score; the explainability analysis unit 303 is configured to generate a heat map through Grad-CAM to visualize the key decision area; and the time series prediction unit 304 is configured to analyze historical data using a TCN time convolution network to predict the change rate of the disease spread area in the next 5 days. The specific operations are as follows: C1: collect the pest identification results, meteorological data, and soil moisture information in the past 14 days to construct a time series input vector; C2: input the time series input vector into a TCN network including multiple dilated causal convolution layers, wherein the dilation coefficients of each layer increase exponentially to expand the receptive field; C3: the network output layer generates a vector of length 5, representing the daily change rate of the disease spread area in the next 5 days; C4: the TCN network is trained by minimizing the following function:

[0087]

[0088] wherein y t is the true spread rate, is the predicted value, γ is the regularization coefficient, and W is the network parameter; C5: the prediction result is used to assist in formulating a precise prevention and control operation plan.

[0089] It should be noted that the multi-model cascade unit 301 fuses Inception-v3 and LightGBM to realize accurate identification of diseases and pests, the confidence evaluation unit 302 quantifies the prediction reliability, the explainability analysis unit 303 generates a heat map to reveal the decision basis, and the TCN network of the time series prediction unit 304 predicts the disease spread trend in the next 5 days.

[0090] Further, it should be noted that the cascade architecture of Inception-v3 (input size 299x299, output 2048-dimensional features) and LightGBM (tree depth = 7, leaf number = 63) is adopted. The time series prediction unit 304 adopts a 4-layer TCN network (expansion coefficient [1, 2, 4, 8], convolution kernel k = 3).

[0091] In the present embodiment, it should also be noted that the decision support module 400 includes an agronomic knowledge matching unit 401, a dose optimization unit 402, a spatio-temporal decision generation unit 403, and a multi-objective optimization unit 404, wherein: the agronomic knowledge matching unit 401 is used to retrieve disease and pest control thresholds and standard schemes from a knowledge base according to crop varieties and growth periods; the dose optimization unit 402 is used to dynamically calculate pesticide application doses based on disease spot density and severity, and balance control effect and environmental cost; the spatio-temporal decision generation unit 403 is used to map diagnosis results to GIS maps to generate pesticide application time and regional planning schemes with spatial heat maps; and the multi-objective optimization unit 404 is used to generate trade-off decision schemes among control efficiency, cost input, and ecological impact through a Pareto optimal algorithm.

[0092] It should be noted that the agronomic knowledge matching unit 401 calls a crop control knowledge base, the dose optimization unit 402 calculates a precise pesticide application scheme, the spatio-temporal decision generation unit 403 realizes disease spatial visualization on a GIS platform, and finally the multi-objective optimization unit 404 balances control efficiency, economic cost, and ecological impact through a Pareto optimal algorithm.

[0093] Further, it should be noted that the pesticide dose calculation formula is: In the formula, D is the pesticide application amount (ml / m 2 ), S d is the disease spot density (pieces / cm 2 ), S v is the severity score (1-5 levels), and a, b are crop growth stage parameters.

[0094] Example 2:

[0095] As Figure 2 and Figure 3As shown, in this embodiment, the crop disease and pest intelligent monitoring method based on multispectral imaging specifically includes the following steps:

[0096] S1: Multispectral data acquisition and standardized preprocessing

[0097] 1. Spectral acquisition:

[0098] DJI Zenmuse P1 multispectral camera (configured with five wavebands of 450 nm blue light, 550 nm green light, 650 nm red light, 760 nm near-infrared, and 850 nm short-wave infrared) is used;

[0099] Flight height is 20 m, ground resolution is 2.5 cm, and frame rate is 30 fps;

[0100] Crop canopy reflectivity data in the range of 400-1000 nm is synchronously acquired;

[0101] 2. Data preprocessing:

[0102] Radiation correction: based on SpectraVue standard reflectance board, according to the formula Calibration (I(λ) is the original pixel value, S(λ) is the sensor gain, and Ref(λ) is the standard reflectivity);

[0103] Geometric registration: improved SIFT algorithm (feature point matching error <0.3 pixels) is used to realize multispectral sub-pixel alignment;

[0104] Noise suppression: first, non-local mean filtering (11×11 search window) is performed, and then db4 wavelet 3-layer decomposition denoising is implemented;

[0105] S2: Cross-modal feature intelligent analysis

[0106] 1. Waveband dynamic screening:

[0107] Each waveband feature vector f is extracted through 1D convolution i ∈R d ;

[0108] According to the specific formula Q=W q f i , Weight (d=64) is calculated;

[0109] Wavebands with weight α i >0.15 are retained to form an optimized combination;

[0110] 2. Feature decoupling reconstruction:

[0111] Encoding stage:

[0112] Spatial branch: ResNet18 extracts texture features (output 256 dimensions);

[0113] Spectral branch: 3-layer 1D CNN extracts spectral features (output 128 dimensions);

[0114] Decoupling constraint: minimize mutual information I(z p ; z b ) by discriminator;

[0115] Reconstruction output: jointly optimized with L1 loss (weight 0.7) + perceptual loss (VGG16 features, weight 0.3);

[0116] S3: Multi-model collaborative diagnosis and prediction

[0117] 1. Pest and disease identification:

[0118] Inception-v3 extracts 2048-dimensional image features, which are reduced to 128 dimensions by PCA and then input into LightGBM (maximum depth 7);

[0119] Output pest and disease type and severity (1-5 rating);

[0120] 2. Time series prediction:

[0121] Construct a 14-day historical data time series (including disease index, temperature, humidity, etc. 8 features);

[0122] TCN network configuration:

[0123] 4 residual blocks (expansion coefficients [1, 2, 4, 8]);

[0124] Each block contains: dilated causal convolution (kernel size = 3), LayerNorm layer, ReLU activation;

[0125] Output layer: linear fully connected (output dimension 5);

[0126] Output 5-day future disease area change rate prediction value

[0127] S4: Precision pesticide application decision making

[0128] 1. Knowledge matching:

[0129] According to the crop variety (such as winter wheat) and growth stage (such as jointing stage), call the control threshold library;

[0130] Example: Rust control threshold = 5 lesions / cm 2 ;

[0131] 2. Dose optimization:

[0132] The amount of drug administration (a = 1.2, b = 0.8 is the parameter of winter wheat) is calculated according to the formula

[0133] Constraint: environmental carrying capacity < 15 ml / m 2 ;

[0134] 3. GIS visualization:

[0135] Generate heat map on ArcGIS Pro platform:

[0136] Red warning area (lesion density > threshold 120%);

[0137] Yellow monitoring area (threshold 80%-120%);

[0138] Green safe area (<80%).

[0139] In the description of the present specification, the description of the terms "one embodiment", "example", "specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0140] The preferred embodiments of the application disclosed above are only used to help explain the application. The preferred embodiments do not describe all the details and limit the application to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of the present specification. The present specification selects and describes these embodiments in order to better explain the principles and practical applications of the application, so that those skilled in the art can well understand and utilize the application. The application is limited only by the claims and their full scope and equivalents.​

Claims

1. A crop disease and pest intelligent monitoring system based on multispectral imaging, characterized in that, The system comprises a multispectral data acquisition and preprocessing module, a visual intelligent analysis module, an intelligent diagnosis module, and a decision support module, wherein: The multispectral data acquisition and preprocessing module is configured to acquire multispectral images of crops by a multispectral imaging device and preprocess the acquired original images; The visual intelligent analysis module is configured to dynamically screen waveband contribution degrees by a cross-waveband adaptive feature fusion network, separate disease and pest sensitive features from background noise by a feature decoupling reconstruction algorithm, and generate sparse feature representations; The intelligent diagnosis module is configured to accurately identify disease and pest types and severity based on a multi-model fusion algorithm with confidence evaluation and explainability, and predict disease development trajectories by combining time series analysis; The decision support module is configured to generate executable schemes including pesticide dosage, operation time, and recommended pesticides by combining diagnosis results and an agronomic knowledge base, and visualize disease spatial distribution by a GIS map. The visual intelligent analysis module comprises a waveband screening unit, a feature fusion unit, a decoupling reconstruction unit, and a sparse representation unit, wherein: The waveband screening unit is configured to calculate spectral weights by a cross-waveband attention mechanism and dynamically screen wavebands with high contribution degrees to disease and pest identification; The feature fusion unit is configured to process spatial-spectral joint features by a 3D convolution network and generate fusion feature maps containing multi-dimensional information; The decoupling reconstruction unit is configured to separate disease and pest features from background noise based on an encoder-decoder architecture and independently represent them by mutual information constraints; The sparse representation unit is configured to compress feature dimensions by L1 regularization and generate sparse feature vectors focusing on disease and pest sensitive information; The decoupling reconstruction unit separates disease and pest features from background noise based on an encoder-decoder architecture and independently represents them by mutual information constraints, and the specific operations are as follows: B1: Encoder feature separation ① a dual-branch encoder network is adopted wherein, extracting spatial features , extracting spectral features , a spatial feature extraction branch, a spectral feature extraction branch; ② input multispectral image , mapped to the latent space by convolutional and fully connected layers: , where, , , , are the spatial and spectral feature dimensions, respectively, H is the image height, W is the image width, and B is the number of bands, is a real tensor of dimension H x W x B. B2: Mutual information constraint decoupled by minimizing mutual information with background noise: where mutual information is approximated by a neural network, is a latent feature vector containing pest and disease information, is a background noise feature vector;​​ ii. Introduce a mechanism of adversarial training: construct a discriminator Discriminate joint distribution Product with marginal distribution The loss function is: In the formula, is a discriminator, used to distinguish the joint distribution And the area of the marginal distribution , is the expectation of the distribution ; B3: Decoder reconstruction , based on a conditional generative adversarial network structure, by and reconstructing feature maps : , wherein G is a conditional generator;​ ②Adopt a multi-scale reconstruction loss function, including: a. Pixel-wise LI loss: where X is the input original multispectral image tensor, is the expectation of the input original multispectral image X; b. perceptual loss: wherein, is a pre-trained visual feature extraction network; B4: Decoupling evaluation and optimization ①calculating the decoupling index triggering a parameter update when triggering a parameter update when The joint optimization objective function is: wherein , , is a weight coefficient, is mutual information between spatial features and spectral features , is a perception loss, is an adversarial training loss, is a preset decoupling threshold.

2. The multispectral imaging based crop disease and pest intelligent monitoring system according to claim 1, characterized in that, The multispectral data acquisition and preprocessing module comprises a spectral acquisition unit, a radiation correction unit, a geometric registration unit, and a noise suppression unit, wherein: The spectral acquisition unit is configured to synchronously acquire image data of crops in multiple wavebands by a multispectral camera array, covering a disease and pest sensitive spectral range of 400-1000nm; The radiation correction unit is configured to eliminate sensor response errors based on standard reflector and dark current calibration; The geometric registration unit is configured to perform multi-waveband image sub-pixel alignment by a SIFT feature matching algorithm, with a registration error of less than 0.3 pixels; The noise suppression unit is configured to remove Gaussian noise and salt and pepper noise in the image by a combination of non-local mean filtering and wavelet transform.

3. The multispectral imaging based crop disease and pest intelligent monitoring system according to claim 2, characterized in that, The multispectral camera array is configured to synchronously acquire image data of at least 5 spectral wavebands, including but not limited to 450nm blue light, 550nm green light, 650nm red light, 760nm near-infrared, and 850nm short-wave infrared.

4. The multispectral imaging based crop disease and pest intelligent monitoring system according to claim 1, wherein, The waveband screening unit calculates spectral weights through a cross-waveband attention mechanism, dynamically screens wavebands with high contribution to disease and pest identification, and specifically operates as follows: A1: processing each band reflectance curve through a 1D convolution layer to generate a band feature vector , R is the value range of each element in the band feature vector , R is the value range of each element in the band feature vector is the dimension of the band feature vector; A2: Calculate band weight The specific formula is as follows: ; wherein, is the i-th waveband feature vector, is the mean feature vector of all wavebands, , is a learnable parameter matrix, Q and K are the query vector and key vector, respectively, is a scaling factor, is the normalized attention weight of the i-th waveband; A3: Retain weights of the wavebands, constituting an optimized waveband combination.

5. The multispectral imaging based crop disease and pest intelligent monitoring system according to claim 1, wherein, The intelligent diagnosis module includes a multi-model cascade unit, a confidence evaluation unit, an explainability analysis unit, and a time series prediction unit, wherein: The multi-model cascade unit is used to cascade Inception-v3 and LightGBM for end-to-end disease and pest identification; The confidence evaluation unit is used to calculate the uncertainty of the prediction result based on Monte Carlo Dropout, and output a 0-1 confidence score; The explainability analysis unit is used to generate a heat map through Grad-CAM to visualize the key decision-making area; The time series prediction unit is used to analyze historical data using a TCN time convolution network to predict the disease spread area change rate in the next 5 days.

6. The multispectral imaging based crop disease and pest intelligent monitoring system according to claim 5, characterized in that, In the time series prediction unit, a TCN time convolution network is used to analyze historical data and predict the disease spread area change rate in the next 5 days, and specifically operates as follows: C1: Collect disease and pest identification results, meteorological data, and soil moisture information in the past 14 days to construct a time series input vector; C2: Input into a TCN network containing multiple dilated causal convolution layers, where the dilation coefficient of each layer increases exponentially to expand the receptive field; C3: The network output layer generates a vector of length 5, representing the daily change rate of disease spread area in the next 5 days; C4: The TCN network is trained by minimizing the following function: ； In the formula, is the true diffusivity on the tth day, is the predicted value on the tth day, is the regularization coefficient, W is the network parameter, and T is the total time step of the training data, is the L2 norm of the parameters as a regularization term; C5: The prediction result is used to assist in formulating a precise prevention and control operation plan.

7. The multispectral imaging based crop disease and pest intelligent monitoring system according to claim 1, wherein, The decision support module includes an agronomic knowledge matching unit, a dosage optimization unit, a spatio-temporal decision generation unit, and a multi-objective optimization unit, wherein: The agronomic knowledge matching unit is used to retrieve disease and pest control thresholds and standard schemes from the knowledge base according to crop varieties and growth cycles; The dosage optimization unit is used to dynamically calculate the pesticide dosage based on disease spot density and severity, balancing prevention and control effect and environmental cost; The spatio-temporal decision generation unit is used to map the diagnosis results to a GIS map to generate a spatial heat map-based pesticide application time and area planning scheme; The multi-objective optimization unit is used to generate a trade-off decision scheme between prevention efficiency, cost input, and ecological impact through a Pareto optimal algorithm.

8. A crop disease and pest intelligent monitoring method based on multispectral imaging, characterized in that: The multi-spectral imaging-based crop disease and pest intelligent monitoring system of any one of claims 1-7 is used, including the following steps: S1: Collect crop images in the 400-1000nm waveband through a multi-spectral camera array, and generate high-quality multi-spectral data sets after radiation correction, geometric registration, and noise suppression processing; S2: Dynamically screen high-contribution spectral wavebands using a cross-waveband attention mechanism, fuse spatial-spectral features through a 3D convolution network, and generate sparse feature representations sensitive to diseases and pests through decoupling reconstruction and L1 regularization; S3: Identify disease and pest types and severity based on a cascade model of Inception-v3 and LightGBM, evaluate prediction confidence through Monte Carlo Dropout, and predict disease spread trends in the next 5 days through a TCN time convolution network; S4: According to the diagnosis results and the agricultural knowledge base, dynamically calculate the pesticide dosage and operation time, generate the prevention and control scheme with spatial heat map through GIS map, and balance the prevention and control efficiency, cost and ecological impact by using Pareto optimal algorithm.

Images