Peanut disease intelligent monitoring method and system and electronic equipment

Abstract

The invention relates to the technical field of peanut disease intelligent monitoring scheme design, in particular to a peanut disease intelligent monitoring method and system and electronic equipment. RGB and near-infrared images are synchronously acquired through a dual-channel acquisition device, after illumination compensation, defogging and geometric correction, multi-scale features are extracted by using a transfer learning optimized Eff cientNet-B4 network, and spectral information is fused by using a dual-path CBAM attention mechanism to generate a disease sensitive feature vector. The cascade classifier realizes disease type identification and severity grading based on ResNet-34 and a random forest model, and predicts a disease development trend in combination with an LSTM time sequence model. And integrating a U-Net segmentation network to generate a visual report, and associating an expert knowledge base to output a prevention and treatment scheme. According to the method, the problems of low detection precision, poor model generalization ability and insufficient decision support of a traditional method are solved, and an efficient and accurate disease management tool is provided for peanut planting.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring scheme design of peanut diseases, and particularly relates to a peanut disease intelligent monitoring method and system and an electronic device. BACKGROUND

[0002] The existing peanut disease monitoring relies on manual visual inspection or single spectral imaging, which has significant defects. Single visible light or near-infrared imaging is difficult to fully capture multi-dimensional characteristics of diseases, such as early discoloration of rust and underground damage of root rot, which need to be analyzed by multiple spectra in cooperation. The existing technology is insufficient in correcting uneven light, fog and lens distortion, resulting in a decrease in feature extraction accuracy. Traditional machine learning relies on manual design of features, and has poor generalization ability for multi-scale disease spots (such as small spore piles and large areas of wilting), and has a high rate of missed detection. Time series prediction mostly uses linear models, which are difficult to model the nonlinear diffusion law of diseases, and lack a dynamic updating mechanism.

[0003] Although the existing deep learning scheme improves the classification accuracy, the model is bloated and the calculation cost is high, which makes it difficult to deploy to field edge devices. Most methods ignore the differentiated fusion of multi-modal data (RGB and near-infrared), resulting in low utilization of spectral information. In addition, the traditional system lacks a closed-loop support from identification to prevention, and farmers still rely on experience to apply pesticides, which is not scientific. Therefore, there is an urgent need for an intelligent monitoring technology that integrates multi-spectral imaging, adaptive feature learning and dynamic optimization to solve the core problems of low detection accuracy, weak model generalization and lack of decision support.

[0004] Therefore, the existing technology still needs further development. SUMMARY

[0005] The present application aims to overcome the above technical deficiencies, and provides a peanut disease intelligent monitoring method and system, and an electronic device to solve the problems existing in the prior art.

[0006] To achieve the above technical purpose, according to the first aspect of the present application, the present application provides a peanut disease intelligent monitoring method, comprising:

[0007] S1, acquiring a visible light image and a near-infrared image of a peanut plant by a multi-spectral imaging device;

[0008] S2, performing standardization preprocessing on the acquired images, including light compensation, fog removal processing and geometric correction;

[0009] S3, constructing a deep convolutional neural network containing disease characteristics, and extracting multi-scale image features by transfer learning;

[0010] S4, fusing visible light and near-infrared feature maps by using an attention mechanism to generate a disease sensitive feature vector;

[0011] S5, input the feature vector into a cascaded classifier to identify the disease type and severity level respectively;

[0012] S6, establish a disease development prediction model based on time series analysis;

[0013] S7, generate a visual report containing disease positioning information and control suggestions.

[0014] Specifically, the multi-modal data fusion in S1 specifically includes:

[0015] RGB images and NIR images are synchronously acquired by a dual-channel acquisition device to establish a pixel-level spatial registration relationship, and adaptive weighting algorithm is used to fuse the two spectral features.

[0016] Specifically, the transfer learning in S3 specifically includes:

[0017] A pre-trained EfficientNet-B4 is used as the backbone network, the shallow convolution kernel parameters are frozen, and the top fully connected layer is fine-tuned to realize disease feature adaptation.

[0018] Specifically, the attention mechanism in S4 uses a double-path CBAM module, which includes:

[0019] The spatial attention submodule generates a spatial weight map through maximum pooling and average pooling;

[0020] The channel attention submodule generates a channel weight vector through global average pooling.

[0021] Specifically, the cascaded classifier in S5 includes:

[0022] The first level ResNet-34 network realizes disease type classification, and outputs class probabilities including leaf spot, rust, and root rot;

[0023] The second level random forest model predicts disease severity based on the feature vector, and is divided into four levels: asymptomatic, mild, moderate, and severe.

[0024] Specifically, the time series analysis in S6 specifically includes:

[0025] An LSTM time series prediction model is constructed, which inputs continuous 7-day disease monitoring data and outputs 3-day disease spread trend prediction values.

[0026] Specifically, the visual report generation in S7 includes:

[0027] A disease area segmentation mask is generated based on the U-Net network, the positioning result is superimposed on the original image, and a control scheme containing the ratio of pesticides and the application cycle is generated combined with the expert knowledge base.

[0028] Specifically, it further includes a model optimization step:

[0029] Deploy an online learning module to dynamically update the classifier parameters through newly collected field data, and use a knowledge distillation algorithm to compress the model.

[0030] According to the second aspect of the present application, a peanut disease intelligent monitoring system is provided, comprising:

[0031] An acquisition module is configured to collect visible light images and near-infrared images of peanut plants through a multispectral imaging device;

[0032] A control module is configured to perform standardized preprocessing on the collected images, including light compensation, defogging processing, and geometric correction; to construct a deep convolutional neural network containing disease characteristics, and extract multi-scale image features through transfer learning; to use an attention mechanism to fuse visible light and near-infrared feature maps to generate a disease-sensitive feature vector; to input the feature vector into a cascaded classifier to identify disease types and severity levels, respectively; to establish a disease development prediction model based on time series analysis; and to generate a visual report containing disease positioning information and control suggestions.

[0033] According to the third aspect of the present application, an electronic device is provided, comprising a memory and a processor, wherein the memory stores computer readable instructions, and the computer readable instructions are executed by the processor to implement the above-mentioned peanut disease intelligent monitoring method.

[0034] Advantages:

[0035] 1. Multi-modal data collaborative analysis: Through dual-channel imaging and pixel-level registration technology, visible light texture details and near-infrared spectral features are captured simultaneously, combined with an adaptive weighted fusion algorithm, which significantly improves the detection rate of early diseases, especially for the highly concealed root rot disease.

[0036] 2. Lightweight deep learning architecture: Using a transfer learning strategy to reuse the general feature extraction capabilities of pre-trained models, combined with an attention mechanism to dynamically focus on disease-sensitive areas, while reducing the demand for 80% of training data, maintaining classification accuracy, and adapting to edge computing device deployment.

[0037] 3. Cascaded classification and spatiotemporal prediction: The cascaded classifier identifies disease types and severity in stages, reducing model complexity; the LSTM time series model integrates continuous monitoring data to accurately predict disease spread trends, providing a scientific basis for early intervention.

[0038] 4. Dynamic optimization and knowledge transfer: The online learning module continuously optimizes model parameters based on newly collected data, combined with knowledge distillation technology to compress the model size, ensuring the long-term stability and adaptability of the system in complex field environments.

[0039] 5. Closed-loop decision support: visualize the report integrated disease positioning map and expert knowledge base recommended control scheme, guide precise pesticide application, reduce pesticide abuse, at the same time, realize real-time marking of disease area through AR interface, and improve the convenience of operation of farmers.

[0040] The above technical breakthroughs significantly improve the disease identification accuracy and computing efficiency, and provide whole-chain digital support from monitoring to decision-making for disease prevention and control, and have significant agricultural economic benefits and ecological value. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a flowchart of the peanut disease intelligent monitoring method provided in the embodiments of the present application;

[0042] Figure 2 is a system composition schematic diagram of the peanut disease intelligent monitoring system provided in the embodiments of the present application. DETAILED DESCRIPTION

[0043] In order for those skilled in the art to better understand the technical solutions of the present application, the technical solutions of the present application will be described clearly and completely below in combination with the drawings of the present application. Based on the embodiments in the present application, other similar embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application. In addition, the direction words mentioned in the following embodiments, such as "up", "down", "left", "right", etc. are only the directions of the drawings, therefore, the direction words used are used to illustrate but not to limit the present application.

[0044] The present application will be further described below in combination with the drawings and preferred embodiments.

[0045] Please refer to Figure 1 The present application provides a peanut disease intelligent monitoring method, comprising:

[0046] S1, acquiring visible light images and near-infrared images of peanut plants through a multispectral imaging device.

[0047] Specifically, the multi-modal data fusion in S1 specifically comprises:

[0048] The RGB image and the NIR image are synchronously acquired by the dual-channel acquisition device, the pixel-level spatial registration relationship is established, and the adaptive weighting algorithm is used to fuse the two spectral features.

[0049] It should be noted that, as for the dual-channel acquisition device, the present application designs the following schemes:

[0050] The visible light imaging unit adopts a Sony IMX415 CMOS sensor, is equipped with a focal length of 12 mm fixed focus lens, has a working waveband of 400-700 nm, a resolution of 3840x2160, a single pixel size of 1.45 μm, and a dynamic range of 120 dB at a frame rate of 30 fps;

[0051] The near-infrared imaging unit adopts an InGaAs sensor, has an effective photosensitive range of 700-1100 nm, is equipped with a band-pass filter (center wavelength 850 nm, bandwidth ±20 nm), a resolution of 1920x1080, and a quantum efficiency of 68% in the target waveband;

[0052] Dual-lens module: The optical axis distance is accurately controlled at 4.8±0.2 mm, is installed on a three-degree-of-freedom holder (horizontal rotation ±30°, pitch angle adjustment ±15°, lifting stroke 0.8-1.5 m), and a built-in gyroscope compensates for pixel deviation caused by mechanical vibration in real time.

[0053] It can be understood that after the visible light imaging unit and the near-infrared imaging unit are installed on the dual-lens module, the dual-lens module is installed on the field robot to collect peanut images.

[0054] It needs to be further explained that the establishment of the pixel-level spatial registration relationship includes homography matrix registration, and specifically includes:

[0055] The spatial correspondence relationship between the visible light (RGB) and near-infrared (NIR) images is established by homography transformation:

[0056]

[0057] In the formula:

[0058] (x, y): NIR image pixel coordinates;

[0059] (x', y'): registered RGB image coordinates;

[0060] H: 3x3 homography matrix (solved by RANSAC algorithm, iteration number ≥500 times);

[0061] The preferred basis is that the error of the solution of the H matrix is controlled within 0.3 pixels, and the edge alignment accuracy of the leaf blade is ensured to be above 95%.

[0062] It needs to be further explained that the fusion of the two spectral features by using the adaptive weighted algorithm includes designing a spectral fusion formula, and specifically includes:

[0063] The adaptive weighted fusion strategy is:

[0064] F fusion (i, j) = a·R norm(i,j) + β · N enh (i,j)

[0065] where:

[0066] R norm : normalized visible light image (dynamic range [0, 1]);

[0067] N enh : result of near-infrared image after CLAHE enhancement (contrast limiting threshold 2.0);

[0068] α = 0.6, β = 0.4: weight coefficients (determined by grid search optimization).

[0069] S2, standardization preprocessing is performed on the collected images, including light compensation, defogging processing and geometric correction.

[0070] It needs to be further explained that the application carries out light compensation based on the Retinex improved model, and the designed formula is as follows:

[0071]

[0072] where:

[0073] I(x,y): original input image (three-channel RGB value);

[0074] G σ : Gaussian convolution kernel (σ = 15, kernel size 25 × 25);

[0075] ∈ = 1e-6: a small constant to prevent division by zero;

[0076] Element-wise division operator;

[0077] Physical meaning: by estimating the illumination component G σ × I, the reflection component L is separated to eliminate the shadow effect.

[0078] It needs to be further explained that as for the defogging processing, the scheme designed by the application includes:

[0079] 1. Dark channel extraction:

[0080] For each pixel of the input image I(x,y), the dark channel value thereof is calculated:

[0081]

[0082] where:

[0083] Ω(x, y): 15x15 pixel window centered at (x, y), which is tested to effectively balance the defogging effect and edge preservation (SSIM > 0.85);

[0084] c: RGB color channel, physically means finding the minimum luminance value in the local region;

[0085] J dark (x, y): pixel value in the dark channel image;

[0086] min (x′,y′)∈Ω(x,y) : minimum value operation in the local window;

[0087] min c∈{R,G,B} : take the minimum value in the three channels of RGB;

[0088] I c (x', y'): intensity value of color channel c at position (x', y') in the original image.

[0089] 2. Atmospheric light value estimation:

[0090] Select the top 0.1% of the pixels in the dark channel image, and take the maximum value of the RGB of these points in the original image as the atmospheric light A:

[0091]

[0092] Where:

[0093] A: atmospheric light value, estimated ambient light intensity;

[0094] Top 0.1%: the top 0.1% of the brightest pixels in the image, sorted based on the maximum value of the RGB channel of each pixel;

[0095] I R (x, y), I G (x, y), I B (x, y): luminance value of red, green, and blue channels.

[0096] max operation: take the maximum value of the three channels at each pixel point, used to determine the brightness of the point, and select the brightest pixel point.

[0097] 3. Transmission rate calculation:

[0098] Estimate the fog concentration distribution by the following formula:

[0099]

[0100] Where:

[0101] ω = 0.95: retain natural fog parameter (empirical optimal value), on the peanut field dataset, the value makes the average gradient of the defogging image increase by 2.3 times;

[0102] t(x, y) e [0.05, 0.8]: constraint the reasonable range of transmittance.

[0103] 4. Image restoration:

[0104] Reconstruct the haze-free image according to the atmospheric scattering model:

[0105]

[0106] Wherein, t0 = 0.1: threshold to prevent denominator too small.

[0107] It needs to be further explained that, regarding the geometric correction, the scheme designed by the application is as follows:

[0108] 1. Lens distortion correction:

[0109] Eliminate radial and tangential distortion by Brown-Conrady model:

[0110] x corrected = x (1 + k1r 2 +k2r 4 )+2p1xy+p2(r 2 +2x 2 )

[0111] y corrected = y (1 + k1r 2 +k2r 4 )+p1(r 2 +2y 2 )+2p2xy

[0112] Wherein:

[0113] x corrected , y corrected : corrected pixel coordinates;

[0114] k1 = -0.12, k2 = 0.03: radial distortion coefficient (determined by 12x9 chessboard calibration);

[0115] p1 = 0.001, p2 = -0.0005: tangential distortion coefficient;

[0116] Normalized radius.

[0117] 2. Leaf deformation compensation:

[0118] Elastic deformation correction based on feature point matching:

[0119] Feature extraction: Detect SIFT feature points in near-infrared image (reserve 500 high response points);

[0120] Deformation modeling:

[0121] Construct thin-plate spline interpolation function:

[0122]

[0123] Where:

[0124] a0is a constant term, representing the translation amount of the reference coordinate system;

[0125] a1x, a2y are first-order terms, describing the linear change trend of the function in x and y directions;

[0126] φ(·) is a radial basis function, describing the local influence intensity of the control point (xi, yi) on the surrounding space, which decays with increasing distance.

[0127] ||(x,y)-(x i ,y i )|| is the Euclidean distance;

[0128] w i is the weight coefficient, and the determination method is to solve the equation group by least squares method, so that f(xi, yi) = zi (zi is the observation value of the control point).

[0129] N is preferably 50, and 50 pairs of matching points are randomly selected to solve the coefficient matrix.

[0130] 3. Dynamic compensation: When the local deformation caused by blade bending exceeds the threshold value (Δ>5 pixels), trigger local grid remapping:

[0131]

[0132] Where:

[0133] m ij represents the elements of the affine transformation matrix (estimated by optical flow of adjacent frames);

[0134] t x ,t y represents the translation compensation.

[0135] The parameters are preferably based on:

[0136] SIFT feature point number 500: balance calculation efficiency (processing time <50ms) and matching accuracy (reprojection error <1.2 pixels);

[0137] Deformation threshold Δ=5 pixels: tested, exceeding this value will cause lesion positioning error >3% (unacceptable).

[0138] S3, constructing a deep convolutional neural network containing disease characteristics, extracting multi-scale image features through transfer learning.

[0139] Specifically, the transfer learning in S3 specifically includes:

[0140] A pre-trained EfficientNet-B4 is used as the backbone network, the shallow convolution kernel parameters are frozen, and the disease feature adaptation is realized by fine-tuning the top fully connected layer.

[0141] Specifically, regarding step S3, the scheme designed by the application includes:

[0142] 1. Multi-scale feature extraction uses dilated convolution:

[0143]

[0144] In the formula:

[0145] F out (i,j) is the output feature map;

[0146] F in (i+r·m,j+r·n) is the input feature map;

[0147] r: dilated rate, preferably 2;

[0148] k: convolution kernel radius (fixed at 3);

[0149] W(m,n): weight value of the convolution kernel at position (m,n).

[0150] The preferred basis is that when r=2, the receptive field reaches 15x15 pixels, matching the common lesion size (5-10mm).

[0151] Loss function design for class imbalance problem:

[0152] FL(p t )=-α t (1-p t ) γ log(p t )

[0153] In the formula:

[0154] p t : model predicted true class probability (range [0, 1]);

[0155] α=0.25: class weight coefficient (disease sample proportion 20%);

[0156] γ = 2: difficult sample focus parameter (experiments show that it can improve small target detection AP value by 8.2%)

[0157] S4, visible light and near-infrared feature maps are fused by using an attention mechanism to generate a disease sensitive feature vector.

[0158] Specifically, the attention mechanism in S4 adopts a double-path CBAM module, including:

[0159] A spatial attention submodule generates a spatial weight map through maximum pooling and average pooling;

[0160] A channel attention submodule generates a channel weight vector through global average pooling.

[0161] Specifically, regarding step S4, the scheme designed by the application includes:

[0162] 1. Design a channel attention calculation formula:

[0163] M c (F)=σ(W1(W0(F avg ))+W1(W0(F max )))

[0164] In the formula:

[0165] F avg : global average pooling feature;

[0166] F max : global maximum pooling feature;

[0167] W0∈R C×C / 16 : dimension reduction fully connected layer weight;

[0168] W1∈R C / 16×C : dimension increasing fully connected layer weight;

[0169] σ: Sigmoid activation function.

[0170] 2. Design a spatial attention calculation formula:

[0171] M s (F)=σ(f 7×7 ([Avg(F);Max(F)]))

[0172] In the formula:

[0173] σ: Sigmoid activation function;

[0174] Avg(F): channel dimension average pooling result;

[0175] Max(F): channel dimension maximum pooling result;

[0176] f 7×7 : 7x7 convolution kernel (match typical lesion diameter 15 pixels).

[0177] S5, input the feature vector into the cascade classifier to identify the disease type and severity level respectively.

[0178] Specifically, the cascade classifier in S5 includes:

[0179] The first level ResNet-34 network realizes disease type classification, and outputs class probability including leaf spot, rust and root rot;

[0180] The second level random forest model predicts disease severity according to the feature vector, which is divided into four levels: asymptomatic, mild, moderate and severe.

[0181] It should be noted that regarding step S5, the scheme designed by the present application includes:

[0182] 1. Disease type identification module:

[0183] Network architecture: based on ResNet-34 modification, remove the original classification layer, add a 512-dimensional fully connected layer with Dropout (0.3 probability)

[0184] Classification threshold setting:

[0185] Determine the threshold: draw the PR curve on the validation set, select the precision turning point (0.72) corresponding to the recall rate ≥ 90% as the alarm threshold

[0186] Fuzzy processing: when the highest class probability is in the interval of 0.65-0.72, mark it as "suspected disease" and start the review mechanism

[0187] 1. Severity assessment module:

[0188] Feature engineering: from the 1024-dimensional features extracted from the deep network, the top 50 key features are selected through the Gini importance evaluation of the random forest.

[0189] Classification standard:

[0190] Asymptomatic (0 level): lesion coverage <5%;

[0191] Mild (1 level): 5%-15% and no wilting;

[0192] Moderate (2 level): 15%-30% or local wilting;

[0193] Severe (3 level): >30% or overall plant yellowing.

[0194] S6, establish a disease development prediction model based on time series analysis.

[0195] Specifically, the time series analysis in S6 specifically includes:

[0196] An LSTM time series prediction model is constructed, and continuous 7-day disease monitoring data is input, and 3-day future disease spread trend prediction values are output.

[0197] It should be noted that, regarding step S6, the scheme designed by the present application includes:

[0198] 1. Input data processing:

[0199] Time window: 7-day disease index sequence (statistical analysis shows that peanut leaf disease completes the incubation period within 7 days);

[0200] Feature construction: 12-dimensional features are extracted daily.

[0201] Specifically, the 12-dimensional features include: lesion coverage (%), lesion area daily growth rate (%), R channel mean, G channel mean, B channel mean, color gradient (the degree of color change), texture contrast, texture entropy, spatial distribution entropy (the uniformity of lesion distribution), environmental temperature (℃), environmental relative humidity (%), light intensity (lux).

[0202] Further, the calculation method of the color gradient is:

[0203]

[0204] Wherein:

[0205] S: the saturation channel of the image in the HSV color space, the value range is [0, 1], reflecting the purity of the color;

[0206] The spatial derivative of saturation in the horizontal and vertical directions is calculated by the Scharr operator;

[0207] G: color gradient amplitude, the larger the value, the more intense the color change (such as lesion edge).

[0208] Further, the calculation method of the color gradient is:

[0209]

[0210] Wherein:

[0211] N: the number of gray levels of the image (usually 256);

[0212] P(i,j): the element in the gray level co-occurrence matrix (GLCM), representing the probability of the co-occurrence of gray levels i and j at a specified direction and distance (such as 1 pixel horizontally).

[0213] Contrast: the larger the value, the greater the local pixel difference (e.g. raised texture of rust pustules).

[0214] Further, the calculation method of the texture entropy is:

[0215]

[0216] Wherein:

[0217] P(i,j): joint probability distribution in GLCM (needs to be normalized);

[0218] Entropy: the larger the value, the more complex and disordered the texture (e.g. random spots of necrotic tissue); the smaller the value, the more uniform the texture (e.g. healthy leaves).

[0219] Further, the calculation method of the spatial distribution entropy is:

[0220]

[0221] Wherein:

[0222] M: the number of blocks uniformly divided from the leaf image (e.g. 32x32 pixel blocks);

[0223] p k : the proportion of diseased pixels in the kth block

[0224] H: the higher the entropy value, the more random the distribution of disease spots (airborne diseases); the lower the entropy value, the more concentrated the distribution (soil-borne diseases).

[0225]

[0226] It should be noted that in the preferred embodiment of the present application, the ambient temperature (℃), the ambient relative humidity (%) and the light intensity (lux) are obtained by temperature sensors, humidity sensors and light sensors mounted on the field robot.

[0227] 2. LSTM network configuration:

[0228] Memory unit: 64 LSTM units in the first layer, retaining the time step output; 32 units in the second layer, outputting the final hidden state;

[0229] Regularization measures: interlayer insertion of Dropout layer (0.2 probability), application of L2 regularization (λ = 0.001) to the weight;

[0230] Predicted output: probability distribution of severity level in the next 3 days (corresponding to the optimal intervention window period for pesticide application).

[0231] It should be further noted that the method further comprises:​

[0232] LSTM cell state update:

[0233] f t = σ(W f · [h t-1 , x t ] + b f )

[0234] i t = σ(W i · [h t-1 , x t ] + b i )

[0235]

[0236] o t = σ(W o · [h t-1 , x t ] + b o )

[0237] h t = o t ⊙ tanh(C t )

[0238] wherein:

[0239] x t : input feature at time t (12-dimensional feature);

[0240] h t : hidden state (dimension 64);

[0241] C t : cell state (memory disease development rule);

[0242] f, i, o: forget gate, input gate, output gate.

[0243] S7, generating a visual report containing disease positioning information and control suggestions.

[0244] Specifically, the visual report generation in S7 includes:

[0245] Generating a disease area segmentation mask based on a U-Net network, superimposing the positioning result on the original image, and combining an expert knowledge base to generate a control scheme containing pesticide proportioning and application cycle.

[0246] It should be noted that, regarding step S7, the scheme designed by the present application includes:

[0247] 1. Lesion positioning technology:

[0248] Segmentation network: Based on U-Net architecture, encoder adopts VGG16 (pre-trained on plant dataset), decoder contains 5 levels of up-sampling.

[0249] Post-processing procedure:

[0250] Morphological opening operation (kernel size 3x3) on output probability map to eliminate small noise;

[0251] Set double threshold processing: probability > 0.7 is determined as the core area of disease spot, 0.4-0.7 is the transition zone;

[0252] Calculate the centroid coordinates of the disease spot, and superimpose a semi-transparent red mask (transparency 60%) on the original image.

[0253] 2. Reasoning rules for control program:

[0254] ① Knowledge base construction:

[0255] Collect the decision-making experience of 12 agricultural experts (service length ≥10 years) through Delphi method

[0256] Integrate 5 industry standard documents such as "Peanut Disease and Pest Control Manual"

[0257] Analyze historical plant protection records (covering 10,000 operation data in Shandong, Henan and other places from 2018 to 2023).

[0258] ② Design rule extraction standard:

[0259] Input variables: disease type (leaf spot / rust / root rot), severity level (0-3), environmental temperature and humidity (15-35℃ / 40-100% RH);

[0260] Output variables: pesticide type, ratio (g / acre), application cycle (days), and operation mode (unmanned aerial vehicle / manual).

[0261] ③ Construct fuzzy reasoning system:

[0262] Reasoning mechanism design:

[0263] Rule activation algorithm design:

[0264]

[0265] Among them:

[0266] α i : Activation strength of the ith rule;

[0267] The membership degree of input variable xj belonging to fuzzy set Aj, quantifying semantic description (such as "high humidity");

[0268] wi ∈[0.6,1.0]: Expert confidence weight (E=0.1*years of experience), 10 years expert weight=1.0, 6 years=0.6, to avoid interference from inexperienced experts;

[0269] De-masking method:

[0270] The proportion of the medicament adopts the gravity method:

[0271]

[0272] Among them:

[0273] α k : The activation strength of the kth decision rule;

[0274] d k : The recommended dose of the kth activation rule;

[0275] Dose: The pesticide application amount of the final decision (unit: g / acre).

[0276] The pesticide application cycle adopts the maximum membership degree principle to select discrete values.

[0277] Specifically, the application also designs a conflict resolution mechanism, including:

[0278] Priority ranking:

[0279] 1. National expert rule (weight + 0.15);

[0280] 2. Localized prevention and control records (match geographical coordinates + 0.1);

[0281] Time effectiveness factor: w new = w base × (1+0.05Δt), Δt is the rule update time difference (years). The typical decision logic designed by the application is shown in Tables 1-4:

[0282] Table 1 Rust disease decision logic

[0283]

[0284] Table 2 Leaf spot disease decision logic

[0285]

[0286]

[0287] Table 3 Root rot disease decision logic

[0288]

[0289] Table 4 Compound disease decision logic

[0290]

[0291]

[0292] It should be noted that the method further comprises:

[0293] When the new prevention cases accumulate to 100, the following steps are taken:

[0294] 1. Calculate the rule matching degree:

[0295] 2. If Match < 70%, start rule weight attenuation: w new = w old × 0.9;

[0296] 3. Add new derived rules: mine frequent item sets (support > 5%, confidence > 80%) through Apriori algorithm.

[0297] Specifically, it further comprises a model optimization step:

[0298] Deploy an online learning module to dynamically update the classifier parameters through newly collected field data, and use a knowledge distillation algorithm to compress the model.

[0299] It should be further noted that, regarding the design of the online learning module, the scheme designed by the present application comprises:

[0300] 1. Data stream processing architecture

[0301] ① Edge data collection:

[0302] Deploy a ring buffer (capacity 5000 images) to store the unannotated data collected in the field in the last 72 hours;

[0303] Trigger annotation conditions: when the detection confidence is < 0.6 or the model prediction is inconsistent with the feedback from farmers, upload to the cloud annotation platform.

[0304] ② Cloud annotation mechanism:

[0305] Use a semi-supervised active learning strategy:

[0306] Use CoreSet algorithm to select the Top 5% samples with the most information (based on feature space density estimation);

[0307] Annotators annotate the selected samples for disease type / severity (annotation time is reduced by 60%).

[0308] 2. Incremental learning algorithm

[0309] ① Optimizer design:

[0310]

[0311] where:

[0312] θt: model's parameters (weights) at time step t;

[0313] θt+1: new parameters after update at time step t+1;

[0314] Loss function gradient computed on new data batch Bnew;

[0315] Gradient weighting term computed on old data batch B old

[0316] L(·): loss function;

[0317] η = 5e -4 : learning rate (adjusted by linear warm-up);

[0318] B new : new batch data (256 latest pest and disease images collected per batch);

[0319] B old : historical data random sampling (10% of past planting season's disease records to prevent catastrophic forgetting);

[0320] λ = 0.3: memory retention coefficient (validation set accuracy drop < 2%).

[0321] ② Loss function:

[0322] L = L CE + 0.5L KL (T = 2)

[0323] where:

[0324] L CE : new data cross-entropy loss;

[0325] L KL : KL divergence loss, constraining parameter update direction to align with original model's knowledge distribution.

[0326] 3. Model update strategy

[0327] ① Dynamic update threshold:

[0328] When validation set F1-score drops consecutively for 3 periods, trigger full parameter fine-tuning (learning rate reset to 1e-3);

[0329] Normal update: only fine-tune the last 3 fully connected layers (save 75% computation resources).​

[0330] Version rollback mechanism:

[0331] Keep the last 5 model snapshots (timestamp + performance indicators).

[0332] When the misjudgment rate of the new version in A / B testing is more than 1.5 times that of the old version, automatically roll back to the stable version.

[0333] Further, regarding the design of the knowledge distillation compression scheme, the scheme designed by the present application includes:

[0334] 1. The classroom-student model architecture is shown in Table 5:

[0335] Table 5 Classroom-student model architecture

[0336] Model type Backbone network Parameter amount FLOPs Applicable scenario Teacher model EfficientNet-B4 19M 4.4B Cloud high-precision inference Student model MobileNetV3-S 2.3M 0.6B Edge real-time processing

[0337] 2. Design the distillation loss function:

[0338] L distill = α·L CE (y,σ(z s ))+β·L KL (σ(z t / T),σ(z s / T))+γ·L AT

[0339] Wherein:

[0340] α: weight coefficient of cross-entropy loss (in the preferred embodiment of the present application, preferably α = 0.7, to ensure the basic classification ability first);

[0341] L CE : standard cross-entropy loss, used to measure the difference between the student model prediction and the real label;

[0342] y: real label;

[0343] σ(z s ): probability distribution after Sigmoid activation of the original output z s of the student model;

[0344] β: KL divergence weight coefficient (in the preferred embodiment of the present application, preferably β = 0.3, to control the knowledge transfer strength);

[0345] L KL : Kullback-Leibler divergence, which measures the similarity between the teacher and student output distributions;

[0346] σ(z t / T): probability distribution after Sigmoid activation of the output z t"Soft label" probability distribution after temperature scaling (T>1) and Sigmoid;

[0347] σ(z s / T): probability distribution of student model output zs after same processing;

[0348] T: temperature parameter (in preferred embodiments of the invention, preferably T=3, softens probability distribution to convey dark knowledge);

[0349] α·L CE (y,σ(z s )) : cross-entropy of student model prediction vs. true label (a = 0.7);

[0350] β·L KL (σ(z t / T),σ(z s / T)): KL divergence of softened logits.

[0351] γ·L AT : attention transfer loss (g = 0.1), matches teacher-student intermediate layer feature maps:

[0352]

[0353] where:

[0354] Teacher model c-th channel attention map (after GAP);

[0355] Student model corresponding layer feature map.

[0356] 3. Distillation training strategy

[0357] ① Two-stage training method:

[0358] Offline pre-distillation:

[0359] Train teacher model on historical dataset (100,000) to convergence (accuracy 92.7%);

[0360] Fix teacher parameters, train student model for 50 rounds with learning rate 5e-4.

[0361] Online distillation:

[0362] Continuously fine-tune student model with new data after deployment (learning rate reduced to le-5);

[0363] Update teacher model parameters synchronously every 500 new data received.

[0364] ② Quantization compression:

[0365] Dynamic 8-bit integer quantization (Dynamic Quantization) is adopted:

[0366] Weights: symmetric quantization (scale = 127 / max(|w|));

[0367] Activation: asymmetric quantization (zero_point = 128).

[0368] Referring to Figure 2 The present application provides another embodiment, which provides a peanut disease intelligent monitoring system, comprising:

[0369] The acquisition module 100 is configured to collect visible light images and near-infrared images of peanut plants by a multispectral imaging device;

[0370] The control module 200 is configured to perform standardization preprocessing on the collected images, including light compensation, defogging processing, and geometric correction; to construct a deep convolutional neural network containing disease characteristics, extract multi-scale image features through transfer learning; to fuse visible light and near-infrared feature maps by using an attention mechanism to generate a disease-sensitive feature vector; to input the feature vector into a cascaded classifier to identify disease types and severity levels, respectively; to establish a disease development prediction model based on time series analysis; and to generate a visual report containing disease positioning information and control suggestions.

[0371] In a preferred embodiment, the present application further provides an electronic device, comprising:

[0372] a memory and a processor, wherein the memory has computer readable instructions stored thereon, and the computer readable instructions, when executed by the processor, implement the peanut disease intelligent monitoring method. The computer device can be a server, a terminal, or any other electronic device with necessary computing and / or processing capabilities. In one embodiment, the computer device can include a processor, a memory, a network interface, a communication interface, and the like connected by a system bus. The processor of the computer device can be used to provide necessary computing, processing and / or control capabilities. The memory of the computer device can include a non-volatile storage medium and an internal memory. The non-volatile storage medium or the non-volatile storage medium can store an operating system, a computer program, and the like. The internal memory can provide an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface and the communication interface of the computer device can be used to connect and communicate with external devices through a network. The computer program is executed by the processor to execute the steps of the method of the present application.

[0373] The application can be implemented as a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, causes the steps of the method of the embodiments of the application to be performed. In one embodiment, the computer program is distributed over a network of coupled computer devices or processors such that the computer program is stored, accessed and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, can be performed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations can be performed by one or more computer devices or processors and one or more other method steps / operations can be performed by one or more other computer devices or processors. One or more computer devices or processors can perform a single method step / operation, or perform two or more method steps / operations.

[0374] It will be appreciated by the person of ordinary skill in the art that the method steps of the application can be instructed by a computer program to relevant hardware such as a computer device or processor, which computer program can be stored in a non-transitory computer-readable storage medium, which computer program, when executed, causes the steps of the application to be performed. Any reference herein to a memory, storage, database or other medium can include non-volatile and / or volatile memory, as the case can be. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state disk, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0375] The various technical features described above can be combined in any manner. Although not all possible combinations of the technical features are described, any combination of the technical features should be considered to be within the scope of the present description, as long as such a combination does not result in a contradiction.

[0376] The above description of specific embodiments of the application does not limit the scope of protection of the application. Any various other corresponding changes and modifications made to the technical concept of the application should be included within the scope of protection of the claims of the application.

Claims

1. A method for intelligent monitoring of peanut diseases, characterized in that: The method comprises: S1. Collect visible light images and near-infrared images of peanut plants using multispectral imaging equipment; S2, perform standardized preprocessing on the collected images, including illumination compensation, defogging and geometric correction; S3. Build a deep convolutional neural network containing disease features and extract multi-scale image features through transfer learning; S4, using the attention mechanism to fuse visible light and near-infrared feature maps to generate disease-sensitive feature vectors; S5, inputting the feature vector into the cascade classifier to identify the disease type and severity level respectively; S6. Establish a disease development prediction model based on time series analysis; S7. Generate a visual report containing disease location information and prevention and control suggestions.

2. The method according to claim 1, characterized in that The multimodal data fusion in S1 specifically includes: The RGB image and NIR image are acquired synchronously through a dual-channel acquisition device, a pixel-level spatial registration relationship is established, and an adaptive weighted algorithm is used to fuse the two spectral features.

3. The method according to claim 1, characterized in that The transfer learning in S3 specifically includes: The pre-trained EfficientNet-B4 is used as the backbone network, the shallow convolution kernel parameters are frozen, and disease feature adaptation is achieved by fine-tuning the top fully connected layer.

4. The method according to claim 1, wherein The attention mechanism in S4 adopts a dual-path CBAM module, including: The spatial attention submodule generates a spatial weight map through maximum pooling and average pooling; The channel attention submodule generates a channel weight vector through global average pooling.

5. The method according to claim 4, characterized in that The cascade classifier in S5 includes: The first-level ResNet-34 network classifies disease types and outputs the probability of leaf spot, rust, and root rot. The second-level random forest model predicts the severity of the disease based on the eigenvector and divides it into four levels: asymptomatic, mild, moderate, and severe.

6. The method according to claim 1, wherein The time series analysis in S6 specifically includes: Build an LSTM time series prediction model, input 7 consecutive days of disease monitoring data, and output the predicted value of disease spread trend in the next 3 days.

7. The method according to claim 1, characterized in that The visualization report generation in S7 includes: The diseased area segmentation mask is generated based on the U-Net network, the positioning results are superimposed on the original image, and the prevention and control plan including the pesticide ratio and application cycle is generated in combination with the expert knowledge base.

8. The method according to claim 1, characterized in that Also includes the model optimization step: An online learning module is deployed to dynamically update the classifier parameters through newly collected field data, and the knowledge distillation algorithm is used to compress the model.

9. An intelligent monitoring system for peanut diseases, characterized in that: include: an acquisition module, used for collecting visible light images and near-infrared images of peanut plants using a multispectral imaging device; The control module is used to perform standardized preprocessing on the collected images, including illumination compensation, dehazing, and geometric correction; to construct a deep convolutional neural network containing disease features and extract multi-scale image features through transfer learning; to fuse visible light and near-infrared feature maps using an attention mechanism to generate disease-sensitive feature vectors; to input feature vectors into a cascade classifier to identify disease types and severity levels respectively; to establish a disease development prediction model based on time series analysis; and to generate a visual report containing disease location information and prevention and control recommendations.

10. An electronic device, characterized in that: include: Memory; and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the intelligent monitoring method for peanut diseases according to any one of claims 1 to 8 is implemented.

Images