Cassava disease and pest detection and early warning system

Abstract

The invention is suitable for the technical field of disease and pest detection and early warning, and provides a cassava disease and pest detection and early warning system. Comprising a satellite remote sensing module, an unmanned aerial vehicle spectrum acquisition module, a meteorological sensing network, a soil parameter monitoring module, a multi-source data fusion analysis module, an intelligent inspection path planning module, a handheld spectrum detection terminal, a terminal server and a cloud early warning platform. The satellite remote sensing module and the unmanned aerial vehicle spectrum acquisition module are used for positioning and spectrum monitoring, and the meteorological sensing network and the soil parameter monitoring module are used for acquiring environmental soil data; the multi-source data fusion analysis module screens a high-risk area, the intelligent routing inspection path planning module optimizes a detection path, and the handheld spectrum detection terminal performs routing inspection based on the optimized path; the terminal server runs a lightweight AI detection module to complete deep analysis and carry out feedback; the cloud early warning platform generates a customized prevention and treatment scheme; and the efficiency and accuracy of disease and pest detection are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of pest and disease detection and early warning, and particularly relates to a cassava pest and disease detection and early warning system. Background Art

[0002] As an important economic crop and food security guarantee crop globally, cassava is an important source of the agricultural economy in many regions and the livelihood of millions of farmers. However, in recent years, affected by pests and diseases such as cassava mosaic disease, brown streak disease, mites, and nematodes, the cassava industry has faced severe challenges, seriously threatening regional food security and farmers' income. Currently, the traditional pest and disease control system mainly relies on manual field inspections and chemical pesticide spraying. Manual inspections are judged by visually observing phenotypic characteristics such as leaf disease spots and insect holes. However, limited by factors such as differences in the experience of inspection personnel, strong concealment of early diseases, and complex field environments, it is difficult to respond promptly to pest and disease outbreaks; although chemical control can control pests in the short term, blind spraying leads to excessive use of pesticides, which not only increases production costs but also causes secondary problems such as soil pollution, enhanced pest and disease resistance, and excessive pesticide residues in agricultural products.

[0003] In recent years, detection methods based on remote sensing technology and artificial intelligence have been gradually applied to the agricultural field. However, the existing detection methods overly rely on high-resolution images and do not adapt to the characteristics of dense cassava canopies and complex leaf textures, resulting in poor model generalization ability. Although convolutional neural networks perform excellently in image classification, the mainstream models have a large number of parameters and rely on cloud servers for inference, leading to high latency in real-time field detection, high equipment deployment costs, and difficulty in adapting to scattered planting scenarios such as hilly areas, resulting in problems such as low detection efficiency, high cost, and poor timeliness. Summary of the Invention

[0004] The embodiments of the present application provide a cassava pest and disease detection and early warning system for solving the problems of low detection efficiency, high cost, and poor timeliness in cassava pest and disease detection.

[0005] The embodiments of the present application provide a cassava pest and disease detection and early warning system, including: a satellite remote sensing module, an unmanned aerial vehicle (UAV) spectral acquisition module, a meteorological sensor network, a soil parameter monitoring module, a multi-source data fusion analysis module, an intelligent inspection path planning module, a handheld spectral detection terminal, a terminal server, and a cloud early warning platform;

[0006] The satellite remote sensing module, the UAV spectral acquisition module, the meteorological sensor network, and the soil parameter monitoring module are respectively connected to the multi-source data fusion analysis module. The satellite remote sensing module is used to obtain wide-area remote sensing images of the cassava planting area. The UAV spectral acquisition module is used to collect canopy multi-spectral images. The meteorological sensor network is used to collect meteorological data. The soil parameter monitoring module is used to collect soil parameters. The multi-source data fusion analysis module is used to analyze the data transmitted by the satellite remote sensing module, the UAV spectral acquisition module, the meteorological sensor network, and the soil parameter monitoring module, and then output a probability map of potential pest and disease areas;

[0007] The intelligent inspection path planning module is respectively connected to the multi-source data fusion analysis module and the handheld spectral detection terminal. The intelligent inspection path planning module is used to analyze the received probability map of potential pest and disease areas and then output an inspection path, which is transmitted to the handheld spectral detection terminal, so that the handheld spectral detection terminal collects image data of cassava leaves and stems based on the inspection path;

[0008] The terminal server is connected to the handheld spectral detection terminal. The handheld spectral detection terminal is built-in with a data transmission module, and the data transmission module is used to transmit the image data of cassava leaves and stems to the terminal server; the terminal server is built-in with a lightweight AI detection module, and the lightweight AI detection module is used to analyze the pest and disease risks of cassava and feedback the analysis results to the handheld spectral detection terminal; the handheld spectral detection terminal outputs a visualization map including the pest and disease risk levels;

[0009] The cloud warning platform is respectively connected to the satellite remote sensing module, the UAV spectral acquisition module, the meteorological sensor network, the soil parameter monitoring module, the handheld spectral detection terminal, and the terminal server. The cloud warning platform is used to output a pest and disease control plan and feedback it to the handheld spectral detection terminal and the terminal server.

[0010] Furthermore, the multi-source data fusion analysis module is used to analyze the data transmitted by the satellite remote sensing module, the UAV spectral acquisition module, the meteorological sensor network, and the soil parameter monitoring module, and then output a probability map of potential pest and disease areas, including:

[0011] Preprocess and extract features from the wide-area remote sensing image and the canopy multi-spectral image to obtain the normalized difference vegetation index and canopy texture features;

[0012] Use the normalized difference vegetation index, canopy texture features, meteorological data, and soil parameters as input data, and adopt a machine learning model to output a probability map of potential pest and disease areas.

[0013] Furthermore, the preprocessing and feature extraction of the wide-area remote sensing image and the canopy multispectral image are performed to obtain the normalized difference vegetation index (NDVI) and the canopy texture features, including:

[0014] Expression of the normalized difference vegetation index (NDVI):

[0015]

[0016] where: NIR is the reflectance of the near-infrared band, and Red is the reflectance of the red band;

[0017] Expression of the canopy texture features:

[0018]

[0019]

[0020]

[0021] where: Contrast, Entropy, and Correlation are the contrast, entropy, and correlation respectively, P(i, j) is the joint probability of pixel values i and j in the gray-level co-occurrence matrix, N is the number of gray levels, μ i and μ j are the mean values of pixel values i and j respectively, σ i and σ j are the standard deviations of pixel values i and j respectively.

[0022] Furthermore, the intelligent inspection path planning module is used to analyze the received probability map of potential pest and disease areas and output the inspection path, including:

[0023] Using the DBSCAN clustering algorithm to aggregate the discrete high-risk points in the probability map of potential pest and disease areas into areas to be inspected;

[0024] Adopting an improved genetic algorithm to optimize the inspection path, with the target parameters being minimizing the moving distance and maximizing the coverage rate of high-risk areas, and the constraint conditions including path continuity constraint, time window constraint, and terrain obstacle avoidance constraint;

[0025] Outputting the inspection navigation path and transmitting it to the handheld spectral detection terminal through the 5G network.

[0026] Furthermore, the use of the DBSCAN clustering algorithm to aggregate the discrete high-risk points in the probability map of potential pest and disease areas into areas to be inspected includes:

[0027] Traversing all discrete high-risk points, identifying core points and expanding the points within the neighborhood of the core points to form density-connected clusters;

[0028] Mark the unclassified points as noise, calculate the minimum bounding rectangle for each cluster, and generate the area to be inspected.

[0029] Furthermore, the improved genetic algorithm is used to optimize the inspection path. The objective parameters are to minimize the moving distance and maximize the coverage rate of high-risk areas. The constraint conditions include path continuity constraint, time window constraint, and terrain obstacle avoidance constraint, including:

[0030] The expression of the objective function F:

[0031]

[0032] Where: D is the total moving distance of the inspection path, D max is the distance of the farthest adjacent traversal among the neutral points in all areas to be inspected, C is the number of high-risk points covered by the path, C total is the total number of high-risk points in the area to be inspected, and w1 and w2 are the weights corresponding to the moving distance and the coverage rate of high-risk areas respectively;

[0033] The expression of the path continuity constraint:

[0034] Path(k + 1) must be connected to Path(k)

[0035] Where: Path(k + 1) and Path(k) are the coordinates of the (k + 1)-th and k-th nodes of the path respectively;

[0036] The expression of the time window constraint:

[0037]

[0038] Where: t k is the driving time of the k-th section of the path, T max is the maximum allowable inspection time, and n is the number of lines;

[0039] The expression of the terrain obstacle avoidance constraint:

[0040]

[0041] Where: Slope(Path(k)) is the terrain slope at the k-th path point.

[0042] Furthermore, the terminal server is built-in with a lightweight AI detection module. The lightweight AI detection module is used to analyze the pest and disease risks of cassava and feedback the analysis results to the handheld spectral detection terminal, including:

[0043] Normalize and continuum remove the multi-spectral image and reflection spectrum in the image data of the cassava leaves and stems respectively;

[0044] Use ResNet34 with pruning optimization to extract image features and output an image feature vector;

[0045] Use a 3-layer 1D-CNN to extract spectral features and output a spectral feature vector after gradually reducing the dimension;

[0046] Weightedly fuse the image feature vector and the spectral feature vector through an SE Block, and use a fully connected layer and Softmax to output the pest and disease category and confidence;

[0047] Calculate the risk density distribution of cassava pests and diseases according to the pest and disease category and confidence, normalize the risk density distribution and render a heat map, and output a dynamic pest and disease risk heat map.

[0048] Furthermore, the step of weightedly fusing the image feature vector and the spectral feature vector through an SE Block, and using a fully connected layer and Softmax to output the pest and disease category and confidence includes:

[0049] Calculate the expression of the class probability distribution P:

[0050]

[0051] where: c ∈ {0, 1, 2} corresponds to three pest and disease categories of healthy, mosaic disease, and brown streak disease;

[0052] The confidence is the highest probability value confidence = max(P).

[0053] Furthermore, the step of calculating the risk density distribution of cassava pests and diseases according to the pest and disease category and confidence, normalizing the risk density distribution and rendering a heat map, and outputting a dynamic pest and disease risk heat map includes:

[0054] Calculate the formula for the risk value according to the pest and disease type and confidence:

[0055]

[0056]

[0057] where: R i is the risk value, P1 and P2 are the probabilities of mosaic disease and brown streak disease respectively, is the Gaussian kernel function, h = 50m is the bandwidth parameter to control the smoothness, and n is the number of detection points in the current visible area;

[0058] Normalize it to [0, 1] and render a heat map according to the gradient color band.

[0059] Furthermore, the cloud warning platform is respectively connected to the satellite remote sensing module, the UAV spectral acquisition module, the meteorological sensor network, the soil parameter monitoring module, the hand-held spectral detection terminal and the terminal server. The cloud warning platform is used to output pest control solutions and feedback them to the hand-held spectral detection terminal and the terminal server, including:

[0060] The cloud warning platform constructs a prediction model for the spread of pests and diseases based on the received multi-source data, and selects pesticides and corresponding pesticide application parameters according to the constructed prediction model;

[0061] The expression of the prediction model:

[0062]

[0063] Where: β is the transmission rate, γ is the recovery rate, and μ is the fatality rate.

[0064] From the above technical solutions, it can be seen that the embodiments of the present application have the following advantages:

[0065] The present invention realizes wide-area coverage and high-resolution monitoring through the satellite remote sensing module and the UAV spectral acquisition module. The meteorological sensor network and the soil parameter monitoring module provide environmental soil data, providing a comprehensive and accurate data basis for subsequent data analysis; the multi-source data fusion analysis module uses a machine learning model to screen high-risk areas, and the intelligent inspection path planning module optimizes the detection route. The hand-held spectral detection terminal then conducts inspections based on the optimized path and uploads the data to the terminal server, enabling the terminal server to run a lightweight AI detection module to complete in-depth analysis and feedback to the hand-held spectral detection terminal and the cloud warning platform, significantly improving the efficiency and accuracy of pest detection; the cloud warning platform generates customized control solutions, shortening the warning response time while reducing the amount of pesticide used, providing an efficient, accurate and low-cost solution for the prevention and control of cassava pests and diseases. Brief Description of the Drawings

[0066] Figure 1 It is a structural block diagram of an embodiment of a cassava pest and disease detection and warning system in the present invention;

[0067] Figure 2 It is a schematic flow chart of an embodiment of the intelligent inspection path planning module in the present invention;

[0068] Figure 3 It is a schematic flow chart of an embodiment of the lightweight AI detection module built in the terminal server in the present invention. Detailed Embodiments

[0069] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0070] In this embodiment, the cassava pest and disease detection and early warning system is used to improve the accuracy of pest and disease detection while shortening the early warning response time. The implementation method in this embodiment can be implemented in the system, on the server, or on the terminal, and no specific limitation is made.

[0071] Embodiment 1

[0072] Please refer to Figure 1 , a cassava pest and disease detection and early warning system in the present invention includes a satellite remote sensing module 1, an unmanned aerial vehicle (UAV) hyperspectral acquisition module 2, a meteorological sensor network 3, a soil parameter monitoring module 4, a multi-source data fusion and analysis module 5, an intelligent inspection path planning module 6, a handheld hyperspectral detection terminal 7, a terminal server 8, and a cloud early warning platform 9. Among them, the satellite remote sensing module 1, the UAV hyperspectral acquisition module 2, the meteorological sensor network 3, and the soil parameter monitoring module 4 are respectively connected to the multi-source data fusion and analysis module 5. The satellite remote sensing module 1 is used to obtain the wide-area remote sensing images of the cassava planting area. The UAV hyperspectral acquisition module 2 is used to acquire the canopy multi-spectral images. The meteorological sensor network 3 is used to collect meteorological data. The soil parameter monitoring module 4 is used to collect soil parameters. The multi-source data fusion and analysis module 5 is used to analyze the data transmitted by the satellite remote sensing module 1, the UAV hyperspectral acquisition module 2, the meteorological sensor network 3, and the soil parameter monitoring module 4 and then output a probability map of potential pest and disease areas; the intelligent inspection path planning module 6 is respectively connected to the multi-source data fusion and analysis module 5 and the handheld hyperspectral detection terminal 7. The intelligent inspection path planning module 6 is used to analyze the received probability map of potential pest and disease areas and then output an inspection path, which is transmitted to the handheld hyperspectral detection terminal 7, so that the handheld hyperspectral detection terminal 7 acquires the image data of cassava leaves and stems based on the inspection path; the terminal server 8 is connected to the handheld hyperspectral detection terminal 7. The handheld hyperspectral detection terminal 7 is internally provided with a data transmission module 71, and the data transmission module 71 is used to transmit the image data of cassava leaves and stems to the terminal server 8; the terminal server 8 is internally provided with a lightweight AI detection module 81, and the lightweight AI detection module 81 is used to analyze the pest and disease risks of cassava and feedback the analysis results to the handheld hyperspectral detection terminal 7; the handheld hyperspectral detection terminal 7 outputs a visualization map including the pest and disease risk levels; the cloud early warning platform 9 is respectively connected to the satellite remote sensing module 1, the UAV hyperspectral acquisition module 2, the meteorological sensor network 3, the soil parameter monitoring module 4, the handheld hyperspectral detection terminal 7, and the terminal server 8. The cloud early warning platform 9 is used to output pest and disease control solutions and feedback them to the handheld hyperspectral detection terminal 7 and the terminal server 8.

[0073] Specifically, the satellite remote sensing module 1, the UAV spectral acquisition module 2, the meteorological sensor network 3, and the soil parameter monitoring module 4 respectively transmit the acquired data to the multi-source data fusion and analysis module 5, enabling the multi-source data fusion and analysis module 5 to analyze and output the probability map of potential pest and disease areas. Specifically, it includes the following:

[0074] 1. Preprocess and extract features from the wide-area remote sensing image and the canopy multi-spectral image to obtain the normalized difference vegetation index (NDVI) and the canopy texture features;

[0075] Convert the original digital values (DN) of the satellite and UAV images to surface reflectance. For satellite images, use an atmospheric correction model to eliminate the effects of atmospheric scattering, absorption, etc.; for UAV images, perform radiometric calibration through the whiteboard calibration method. Align the satellite image and the UAV image to the same geographic coordinate system, and perform spatial correction using ground control points (GCPs) or an automatic registration algorithm; ensure that the red-edge band and the short-wave infrared band are precisely aligned with the visible-near infrared band of the UAV.

[0076] Expression of the normalized difference vegetation index NDVI:

[0077]

[0078] Where: NIR is the reflectance of the near-infrared band, and Red is the reflectance of the red band;

[0079] Expression of the canopy texture features:

[0080]

[0081]

[0082]

[0083] Where: Contrast, Entropy, and Correlation are the contrast, entropy, and correlation respectively, P(i,j) is the joint probability of pixel values i and j in the gray-level co-occurrence matrix, N is the number of gray levels, μ i and μ j are the means of pixel values i and j respectively, and σ i and σ j are the standard deviations of pixel values i and j respectively.

[0084] 2. Use the normalized difference vegetation index, the canopy texture features, the meteorological data, and the soil parameters as input data, and adopt a machine learning model to output the probability map of potential pest and disease areas.

[0085] Meteorological data including temperature, humidity, and rainfall, soil parameters including humidity, pH, and conductivity, and canopy texture features were interpolated to the same spatial resolution as satellite images. Gray-level co-occurrence matrix (GLCM) analysis was performed on drone images to extract texture features such as contrast, entropy, and correlation, and then aggregated to 10-meter grid cells. Input feature vectors were constructed, including vegetation index, meteorological data, soil parameters, and canopy texture, and the features were Z-score standardized to eliminate dimensional differences. Historical annotation data was used as labels and feature vectors as input; a random forest model was used to output the probability of pests and diseases (0-1) for each grid cell, and a potential pest and disease area was determined when the probability threshold was set to ≥0.65; the prediction results were mapped to a grid layer, and high-probability areas were marked as red warnings. The model performance was evaluated through cross-validation, and the threshold or feature weight was dynamically adjusted in combination with field inspection data. The model output NDVI / NDRE layers intuitively displayed the health of vegetation; the probability map of potential pests and diseases was displayed through a GIS platform overlay.

[0086] Specifically, the cloud warning platform 9 is respectively connected to the satellite remote sensing module 1, the UAV spectrum acquisition module 2, the meteorological sensor network 3, the soil parameter monitoring module 4, the handheld spectrum detection terminal 7 and the terminal server 8. The cloud warning platform 9 is used to output the pest control plan and feed it back to the handheld spectrum detection terminal 7 and the terminal server 8, including the following:

[0087] The cloud early warning platform 9 constructs a prediction model for the spread of pests and diseases based on the received multi-source data, and selects pesticides and corresponding application parameters based on the constructed prediction model.

[0088] The cloud warning platform 9 receives multi-source data from terminals, drones, weather stations and soil sensors, writes them into the spatiotemporal database, and removes GPS drift points or sensor failure data. Based on the SEIRD model, the disease spread is predicted. The expression of the prediction model is as follows:

[0089]

[0090] Among them: β is the transmission rate, γ is the recovery rate, and μ is the mortality rate.

[0091] Using the finite difference method for discrete solution, the scope of infection in the next 7 days can be predicted.

[0092] Finally, the selected pesticide and the warning scheme generated by the corresponding application parameters are pushed to the handheld spectral detection terminal 7 and the terminal server 8, including push voice prompts that support minority recognition, pesticide application maps and pesticide purchase links, regional prevention and control priority heat maps and pesticide inventory warnings.

[0093] Embodiment 2

[0094] See alsoFigure 2 , in the embodiment where the intelligent inspection path planning module in the present invention is used to analyze the received probability map of potential pest and disease areas and output the inspection path, the steps are as follows:

[0095] S21. Use the DBSCAN clustering algorithm to aggregate the discrete high-risk points in the probability map of potential pest and disease areas into areas to be inspected;

[0096] 1. Traverse all discrete high-risk points, identify core points, and expand the points within the neighborhood of the core points to form density-connected clusters;

[0097] Set the GPS coordinate set of discrete high-risk points as P = {p1, p2,..., p n} where p i = (x i , y i ) (x i is longitude, y i is latitude). Set the neighborhood radius according to the density of the planting area, with a typical value of Eps = 50m; set the minimum number of neighborhood points of the core point MinPts to 5. Traverse all points p i ∈ P, calculate the typical value - the number of points N Eps (p i ) in the neighborhood of each point; if N Eps (p i ) ≥ MinPts, then mark p i as a core point. If p i is not classified into any cluster, create a new cluster C and add p i to C; traverse the points p i in the typical value - neighborhood of p k , if p k is not classified, add p k to cluster C j ; if p k is a core point, recursively expand its neighborhood points.

[0098] 2. Mark the unclassified points as noise, calculate the minimum bounding rectangle for each cluster, and generate the areas to be inspected.

[0099] Mark the points that are not assigned to any cluster as noise points; for each cluster C, calculate its minimum bounding rectangle (MBR) or convex hull (Convex Hull) as the boundary of the area to be inspected. The finally output set of areas to be inspected is R = {R1, R2,..., R m}, where R j (j ∈ 1, 2,..., m) contains the boundary coordinates and the number of points within the cluster.

[0100] S22. Optimize the inspection path using an improved genetic algorithm. The objective parameters are to minimize the moving distance and maximize the coverage rate of high-risk areas. The constraint conditions include path continuity constraint, time window constraint, and terrain obstacle avoidance constraint;

[0101] The expression of the objective function with the objective parameters of minimizing the moving distance and maximizing the coverage rate of high-risk areas is as follows:

[0102]

[0103] Where: D is the total moving distance of the inspection path, D max is the distance of the farthest adjacent traversal of the neutral points of all areas to be inspected, C is the number of high-risk points covered by the path, C total is the total number of high-risk points in the area to be inspected, and w1 and w2 are the weights corresponding to the moving distance and the coverage rate of high-risk areas respectively.

[0104] The expression of the path continuity constraint:

[0105] Path(k + 1) must be connected to Path(k)

[0106] Where: Path(k + 1) and Path(k) are the coordinates of the (k + 1)-th and k-th nodes of the path respectively;

[0107] The expression of the time window constraint:

[0108]

[0109] Where: t k is the driving time of the k-th section of the path, T max is the maximum allowable inspection time, and n is the number of lines;

[0110] The expression of the terrain obstacle avoidance constraint:

[0111]

[0112] Where: Slope(Path(k)) is the terrain slope at path point k.

[0113] Optimize the inspection path using an improved genetic algorithm as follows:

[0114] Adaptive crossover probability: Where g is the current iteration number, and G max is the maximum iteration number. Perform a high crossover rate in the early stage and reduce it in the later stage. Elite retention strategy: Retain the top 10% of individuals with the highest fitness in each generation and directly transfer them to the next generation to prevent the loss of excellent solutions. Dynamic mutation operator: Increase the mutation probability for individuals with low fitness to jump out of local optima.

[0115] S23. Output the inspection navigation path and transmit it to the handheld spectral detection terminal through the 5G network.

[0116] Convert the optimized path point sequence into tags in KML format; embed the expected time, the number of covered points, and slope warnings for each path segment in KML; push the KML file to the handheld terminal through 5G and display it in real time on the GIS map.

[0117] In the above embodiment, through further analysis on the basis of the potential pest and disease areas screened out by the intelligent inspection path planning module, the optimal inspection path is obtained. The multi-source data fusion analysis module screening potential pest and disease areas provides an accurate and effective data basis for the intelligent inspection path planning module. The intelligent inspection path planning module conducts path planning with the goals of minimizing the moving distance and maximizing the coverage rate of high-risk areas, effectively shortening the inspection time and effectively improving the efficiency of pest and disease detection.

[0118] Embodiment III

[0119] Please refer to Figure 3 , the terminal server of the present invention is built-in with a lightweight AI detection module, and the lightweight AI detection module is used to analyze the pest and disease risks of cassava and feedback the analysis results to the handheld spectral detection terminal, including the following steps:

[0120] S31. Standardize and continuum remove the multi-spectral image and the reflection spectrum in the image data of cassava leaves and stems respectively;

[0121] 1. Assume that the multi-spectral image block in the image data of cassava leaves and stems is I ∈ R 224×224×C , where C is the number of spectral channels (such as RGB + red edge + near-infrared, C = 5). Standardization formula: Where: μ c is the mean of channel c, and σ c is the standard deviation of channel c. Finally, the output standardized image is I norm ∈ R 224×224×5 .

[0122] 2. The input spectrum for spectral data normalization is the reflectance curve S ∈ R 1080 , and the wavelength range is 340 - 850 nm. Continuum removal: i = 1, 2,..., 1080; where C(i) is the continuum curve fitted by the local maximum points of the spectrum; finally, the output normalized spectrum is S CR ∈ R 1080 .

[0123] S32. Use ResNet34 with pruning optimization to extract image features and output an image feature vector;

[0124] Each residual block contains the following operations. Taking the l-th layer as an example, the convolutional layer is: Z l = W l * X l―1 + b l ; where is the convolutional kernel, k = 3×3, C in 、C out are the number of input and output channels, and X l―1 is the input feature map. Batch normalization is where μ B 、 are the mean and variance of the current batch, and γ and β are learnable parameters. The activation function is: The channel importance score is the channel contribution based on Taylor expansion: where: Γ c is the loss function, and W c,x,y is the convolutional kernel weight of the c-th channel. When determining the pruning threshold, remove the channels where Γ c < 0.1·max(Γ). After pruning ResNet34, the final output image feature vector F img ∈ R 512 .

[0125] S33. Use a 3-layer 1D-CNN to extract spectral features, and output the spectral feature vector after gradually reducing the dimension; perform convolution operations. The input of the first layer is 1080 dimensions → the output is 512 dimensions (convolutional kernel 5, stride 2); the input of the second layer is 512 dimensions → the output is 256 dimensions (convolutional kernel 5, stride 2); the input of the third layer is 256 dimensions → the output is 128 dimensions (convolutional kernel 5, stride 2). After global average pooling (GAP), output the spectral feature vector F spec ∈ R 128 .

[0126] S34. Weightedly fuse the image feature vector and the spectral feature vector through the SE Block, and use the fully connected layer and Softmax to output the pest and disease categories and confidence levels;

[0127] Concatenate the image and spectral features into a joint feature as: The calculation of the channel attention weight includes compression, excitation, and scaling. Among them, compression is to obtain channel statistics through global average pooling: Excitation is to learn the weight through a fully connected layer: s = σ(W2·δ(W1·z)), where W1 ∈ R r×640 , r is the reduction ratio, δ is the ReLU activation, W2 ∈ R 640×r , and σ is the Sigmoid function. Scaling is to adjust the feature according to the weight: F fused = s·F concat .

[0128] S35. Calculate the risk density distribution of cassava pests and diseases according to the pest and disease categories and confidence levels, normalize the risk density distribution to render a heat map, and output a dynamic pest and disease risk heat map.

[0129] Fully connected layer classification: Input the fused feature F fused ∈R 640 , passing through the fully connected layer: logits = W fc ·F fused +b fc , where W fc ∈R 3×640 , b fc ∈R 3 . Calculate the category probability distribution:

[0130]

[0131] Among them: c ∈ {0, 1, 2} corresponds to three pest and disease categories of healthy, mosaic disease, and brown streak disease;

[0132] The confidence level is the highest probability value confidence = max(P).

[0133] The calculation formula for the risk value according to the pest and disease type and confidence level:

[0134]

[0135]

[0136] Among them: R i is the risk value, P1 and P2 are the probabilities of mosaic disease and brown streak disease respectively, is the Gaussian kernel function, h = 50m is the bandwidth parameter, controlling the smoothness, and n is the number of detection points in the current visible area;

[0137] Normalize it to [0, 1] and render the heat map according to the gradient color band.

[0138] Finally, call the built-in map SDK in the terminal (such as the Gaode Map API), and overlay the heat map layer on the satellite base map. After a new detection point is added, the local KDE is recalculated and the heat map is incrementally updated. If the user moves beyond the current range, the surrounding pre-cached data is automatically loaded (based on the 4G / 5G network). The user can zoom in and out of the map and click on the hotspot to view the details (disease type, confidence level, recommended recheck time).

[0139] In the above embodiments, by normalizing and continuum removing the multi-spectral images and reflection spectra of cassava leaves and stems respectively, the data quality can be improved; the ResNet34 optimized by pruning is used to extract image feature vectors, combined with the spectral feature vectors extracted by 3-layer 1D-CNN, and the dual-channel features are weighted and fused through the SE Block to improve the accuracy of disease classification; based on the risk density distribution calculated from the disease category and confidence, a dynamic pest and disease risk heat map is generated through normalized rendering, realizing the precise positioning and visual warning of pests and diseases, providing real-time and intuitive decision-making support for farmers, and significantly improving the efficiency and prevention and control effect of cassava pest and disease detection.

[0140] Those of ordinary skill in the art can realize that the units of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components of each example have been generally described according to their functions in the above description. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0141] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored, etc. In addition, the functional units in each embodiment of the present invention can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units.

[0142] If the above integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as USB flash drives, read-only memories (ROMs), random access memories (RAMs), mobile hard disks, magnetic disks, or optical discs.

[0143] It will be understood that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention, and they should all be covered within the scope of the claims and the description of the present invention.

Claims

1. A cassava pest and disease detection and early warning system, characterized in that, Including: Satellite remote sensing module, UAV spectral acquisition module, meteorological sensor network, soil parameter monitoring module, multi-source data fusion and analysis module, intelligent inspection path planning module, handheld spectral detection terminal, terminal server and cloud warning platform; The satellite remote sensing module, UAV spectral acquisition module, meteorological sensor network and soil parameter monitoring module are respectively connected to the multi-source data fusion and analysis module. The satellite remote sensing module is used to obtain the wide-area remote sensing images of the cassava planting area. The UAV spectral acquisition module is used to collect canopy multi-spectral images. The meteorological sensor network is used to collect meteorological data. The soil parameter monitoring module is used to collect soil parameters. The multi-source data fusion and analysis module is used to analyze the data transmitted by the satellite remote sensing module, UAV spectral acquisition module, meteorological sensor network and soil parameter monitoring module and then output a probability map of potential pest and disease areas; The intelligent inspection path planning module is respectively connected to the multi-source data fusion and analysis module and the handheld spectral detection terminal. The intelligent inspection path planning module is used to analyze the received probability map of potential pest and disease areas and then output an inspection path, and transmit it to the handheld spectral detection terminal, so that the handheld spectral detection terminal collects image data of cassava leaves and stems based on the inspection path; The terminal server is connected to the handheld spectral detection terminal. The handheld spectral detection terminal is built-in with a data transmission module. The data transmission module is used to transmit the image data of cassava leaves and stems to the terminal server. The terminal server is built-in with a lightweight AI detection module. The lightweight AI detection module is used to analyze the pest and disease risks of cassava and feedback the analysis results to the handheld spectral detection terminal. The handheld spectral detection terminal outputs a visualization map including the pest and disease risk level; The cloud warning platform is respectively connected to the satellite remote sensing module, UAV spectral acquisition module, meteorological sensor network, soil parameter monitoring module, handheld spectral detection terminal and terminal server. The cloud warning platform is used to output pest and disease control plans and feedback them to the handheld spectral detection terminal and the terminal server.

2. The cassava pest and disease detection and early warning system according to claim 1, characterized in that, The multi-source data fusion and analysis module is used to analyze the data transmitted by the satellite remote sensing module, UAV spectral acquisition module, meteorological sensor network and soil parameter monitoring module and then output a probability map of potential pest and disease areas, including: Preprocessing and feature extraction are performed on the wide-area remote sensing images and canopy multi-spectral images to obtain the normalized difference vegetation index (NDVI) and canopy texture features; Taking the normalized difference vegetation index (NDVI), canopy texture features, meteorological data and soil parameters as input data, a machine learning model is used to output a probability map of potential pest and disease areas.

3. The cassava pest and disease detection and early warning system according to claim 2, characterized in that, The preprocessing and feature extraction of the wide-area remote sensing images and canopy multi-spectral images to obtain the normalized difference vegetation index (NDVI) and canopy texture features include: Expression of the normalized difference vegetation index NDVI: Where: NIR is the reflectance of the near-infrared band, and Red is the reflectance of the red band; Expression of canopy texture features: where Contrast, Entropy, and Correlation are the contrast, entropy, and correlation respectively, P(i,j) is the joint probability of pixel values i and j in the gray-level co-occurrence matrix, N is the number of gray levels, μ i and μ j are the means of pixel values i and j respectively, and σ i and σ j are the standard deviations of pixel values i and j respectively.

4. The cassava pest and disease detection and early warning system according to claim 1, wherein, The intelligent inspection path planning module is used to analyze the received probability map of potential pest and disease areas and output an inspection path, including: Using the DBSCAN clustering algorithm to aggregate discrete high-risk points in the probability map of potential pest and disease areas into areas to be inspected; Adopting an improved genetic algorithm to optimize the inspection path, with the objective parameters being minimizing the moving distance and maximizing the coverage rate of high-risk areas, and the constraint conditions including path continuity constraint, time window constraint, and terrain obstacle avoidance constraint; Outputting an inspection navigation path and transmitting it to the handheld spectral detection terminal through the 5G network.

5. The cassava pest and disease detection and early warning system according to claim 4, wherein The step of using the DBSCAN clustering algorithm to aggregate discrete high-risk points in the probability map of potential pest and disease areas into areas to be inspected includes: Traversing all discrete high-risk points, identifying core points and expanding the points within the neighborhood of the core points to form density-connected clusters; Marking the unclassified points as noise, calculating the minimum bounding rectangle for each cluster, and generating areas to be inspected.

6. The cassava pest and disease detection and early warning system according to claim 4, characterized in that The step of adopting an improved genetic algorithm to optimize the inspection path, with the objective parameters being minimizing the moving distance and maximizing the coverage rate of high-risk areas, and the constraint conditions including path continuity constraint, time window constraint, and terrain obstacle avoidance constraint, includes: The expression of the objective function F: Where: D is the total moving distance of the inspection path, D max is the distance of the farthest adjacent traversal of the neutral points of all areas to be inspected, C is the number of high-risk points covered by the path, C total is the total number of high-risk points in the area to be inspected, and w1 and w2 are the weights corresponding to the moving distance and the high-risk area coverage rate respectively; The expression of the path continuity constraint: Path(k + 1) must be connected to Path(k). Where: Path(k + 1) and Path(k) are the coordinates of the (k + 1)-th and k-th nodes of the path respectively; The expression of the time window constraint: where: t k is the travel time of the k-th path, T max is the maximum allowable inspection time, and n is the number of lines; The expression of the terrain obstacle avoidance constraint: Slope(Path(k)) ≤ 15% Where: Slope(Path(k)) is the terrain slope at path point k.

7. The cassava pest and disease detection and early warning system according to claim 1, wherein The terminal server is built-in with a lightweight AI detection module, and the lightweight AI detection module is used to analyze the pest and disease risks of cassava and feedback the analysis results to the handheld spectral detection terminal, including: Normalizing and continuum removing the multi-spectral images and reflection spectra in the image data of the cassava leaves and stems respectively; Adopting a pruned ResNet34 to extract image features and output an image feature vector; Using a 3-layer 1D-CNN to extract spectral features and output a spectral feature vector after gradually reducing the dimension; Weightedly fusing the image feature vector and the spectral feature vector through an SE Block, and using a fully connected layer and Softmax to output the pest and disease categories and confidence levels; Calculating the risk density distribution of cassava pest and disease according to the pest and disease categories and confidence levels, normalizing the risk density distribution and rendering a heat map, and outputting a dynamic pest and disease risk heat map.

8. The cassava pest and disease detection and early warning system according to claim 7, characterized in that, The step of weightedly fusing the image feature vector and the spectral feature vector through an SE Block, and using a fully connected layer and Softmax to output the pest and disease categories and confidence levels, includes: The expression for calculating the class probability distribution P: Where: c ∈ {0, 1, 2} corresponds to three pest and disease categories of healthy, mosaic disease, and brown streak disease; The confidence level is the highest probability value confidence = max(P).

9. The cassava pest and disease detection and early warning system according to claim 8, characterized in that, The step of calculating the risk density distribution of cassava pest and disease according to the pest and disease categories and confidence levels, normalizing the risk density distribution and rendering a heat map, and outputting a dynamic pest and disease risk heat map, includes: The calculation formula for calculating the risk value according to the pest and disease type and confidence level: Where: R i is the risk value, P1 and P2 are the probabilities of mosaic disease and brown stripe disease respectively, is the Gaussian kernel function, h = 50m is the bandwidth parameter that controls the smoothness, and n is the number of detection points in the current visible area; Normalize to [0, 1] and render the heat map according to the gradient color bar.

10. The cassava pest and disease detection and early warning system according to claim 1, characterized in that, The cloud warning platform is respectively connected to the satellite remote sensing module, the UAV spectral acquisition module, the meteorological sensor network, the soil parameter monitoring module, the handheld spectral detection terminal and the terminal server. The cloud warning platform is used to output the pest control plan and feedback it to the handheld spectral detection terminal and the terminal server, including: The cloud warning platform constructs a prediction model for the spread of pests and diseases based on the received multi-source data, and selects pesticides and corresponding pesticide application parameters according to the constructed prediction model; The expression of the prediction model: Where: β is the transmission rate, γ is the recovery rate, and μ is the case fatality rate.

Images