Insect situation forecasting remote management platform based on Internet of Things

Abstract

The invention provides an insect condition forecasting remote management platform based on the Internet of Things, and relates to the technical field of insect disease recognition, and the method comprises the steps: 1, carrying out the preprocessing of an obtained original image of a crop, obtaining a standardized image, inputting the standardized image into a pre-trained lightweight convolutional neural network, and carrying out the feature extraction, and obtaining a multi-scale feature map; 2, regarding the feature vector of each spatial position in the multi-scale feature map as a feature point to obtain a feature point set, projecting the feature point set to a virtual feature plane, constructing a regular geometric structure covering the distribution of all feature points, and dividing the regular geometric structure into a plurality of sub-regions; according to the invention, through multi-scale feature extraction, spatial structure analysis and weight modulation, corn disease and pest identification accuracy is improved, remote transmission and management of pest situation forecasting information are realized in combination with an Internet of Things platform, and the problems of insufficient utilization of disease and pest spatial information and low management efficiency are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of pest and disease identification technology, and in particular to a remote management platform for pest monitoring and forecasting based on the Internet of Things. Background Technology

[0002] In the application of intelligent agricultural monitoring, the use of lightweight convolutional neural networks for image recognition of maize diseases and pests has attracted considerable attention. These methods typically employ optimized network structures to achieve feasible recognition efficiency in field environments with limited computing resources, providing a potential technical path for the automation of maize pest forecasting. However, in practical applications, some existing recognition schemes still have room for improvement. A common issue is that some methods may have limitations in utilizing the spatial distribution and structural relationships of maize disease and pest areas during feature extraction and fusion. This is mainly reflected in the fact that the model may sometimes focus more on the recognition of apparent features such as color and texture, while paying relatively little attention to the implicit spatial contextual information such as the overall layout and relative position of maize lesions or damage points in the image.

[0003] For example, when analyzing maize leaf images, some leaf lesions may be similar in color and texture, but their distribution pattern on the leaf surface, extension trend along the leaf veins, or degree of local aggregation may be different. Existing methods may not be ideal in distinguishing maize diseases with similar appearance features but different spatial distribution patterns, and the identification results are easily affected by complex backgrounds or changes in imaging conditions. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide an IoT-based remote management platform for pest monitoring and forecasting, which improves the model's ability to distinguish between pests and diseases that appear similar but have different spatial structures, and reduces the misjudgment rate.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a remote management method for pest monitoring and forecasting based on the Internet of Things (IoT), the method comprising: Step 1: Preprocess the acquired original crop images to obtain standardized images, and input the standardized images into a pre-trained lightweight convolutional neural network for feature extraction to obtain multi-scale feature maps; Step 2: Treat the feature vector of each spatial location in the multi-scale feature map as a feature point to obtain a set of feature points. Project the set of feature points onto a virtual feature plane to construct a regular geometric structure that covers the distribution of all feature points, and divide the regular geometric structure into multiple sub-regions. Step 3: Map each feature point to the corresponding sub-region according to its projected coordinates, and aggregate the information of all feature points in each sub-region to generate a set of region aggregated features; Step 4: Based on the regional aggregation features, a spatial structure enhancement weight vector is calculated. The multi-scale feature map is modulated by the spatial structure enhancement weight vector to obtain the multi-scale features after spatial structure enhancement. Step 5: Perform feature fusion and classification on the multi-scale features after spatial structure enhancement to obtain the identification results of maize diseases and pests; Step 6: Based on the identification results of corn diseases and pests, generate pest monitoring information and transmit and manage it remotely through the Internet of Things platform.

[0006] Secondly, an IoT-based remote management platform for insect pest monitoring and forecasting includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0007] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0008] The above-described solution of the present invention has at least the following beneficial effects: By combining multi-scale feature extraction and regional feature aggregation, this design maintains a lightweight network architecture while preserving detailed information from shallow features (such as lesion edges and textures), local shape information from mid-level features (such as lesion outlines), and semantic information from deep features. Furthermore, it strengthens the local correlation of features through feature aggregation within sub-regions, compensating for the fragmented feature representation inherent in lightweight networks. Feature points are mapped to a virtual plane to construct regular geometric regions. Geometric analysis, including feature distribution centroids, principal directions, and quadrant density / distance, is used to uncover the spatial distribution patterns of pests and diseases, such as the symmetrical distribution of circular lesions and the linear layout of striped pests. Finally, spatial structure enhancement weights are generated to process the spatial features related to pests and diseases. The enhancement process, which suppresses background interference features, effectively improves the model's ability to distinguish between pests and diseases that appear similar but have different spatial structures, reducing the false positive rate. Size normalization and color channel standardization in the preprocessing stage eliminate color shifts caused by image size differences and lighting, providing stable input for feature extraction. Weight vectors generated through geometric relationship analysis focus solely on the core spatial structural features of pests and diseases, rather than disturbing apparent noise, effectively filtering irrelevant interference information from the feature map. This dual anti-interference design reduces fluctuations in recognition accuracy even under varying lighting conditions. A visual interface allows for remote viewing and operation, shortening the cycle from pest information collection to application and improving the timeliness and practicality of monitoring and forecasting. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the remote management method for insect pest monitoring and forecasting based on the Internet of Things provided in an embodiment of the present invention.

[0010] Figure 2 This is a flowchart illustrating step 2 provided in an embodiment of the present invention. Detailed Implementation

[0011] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0012] like Figure 1 As shown, embodiments of the present invention propose a remote management method for insect pest monitoring and forecasting based on the Internet of Things (IoT). The method includes the following steps: Step 1: Preprocess the acquired original crop images to obtain standardized images, and input the standardized images into a pre-trained lightweight convolutional neural network for feature extraction to obtain multi-scale feature maps; Step 2: Treat the feature vector of each spatial location in the multi-scale feature map as a feature point to obtain a set of feature points. Project the set of feature points onto a virtual feature plane to construct a regular geometric structure that covers the distribution of all feature points, and divide the regular geometric structure into multiple sub-regions. Step 3: Map each feature point to the corresponding sub-region according to its projected coordinates, and aggregate the information of all feature points in each sub-region to generate a set of region aggregated features; Step 4: Based on the regional aggregation features, a spatial structure enhancement weight vector is calculated. The multi-scale feature map is modulated by the spatial structure enhancement weight vector to obtain the multi-scale features after spatial structure enhancement. Step 5: Perform feature fusion and classification on the multi-scale features after spatial structure enhancement to obtain the identification results of maize diseases and pests; Step 6: Based on the identification results of corn diseases and pests, generate pest monitoring information and transmit and manage it remotely through the Internet of Things platform.

[0013] In this embodiment of the invention, a combined design of multi-scale feature extraction and regional feature aggregation is used to maintain a lightweight network architecture. This preserves detailed information of shallow features, such as lesion edges and textures, local shape information of mid-level features, such as lesion outlines, and semantic information of deep features. Furthermore, feature aggregation within sub-regions strengthens the local correlation of features, compensating for the fragmented feature representation of lightweight networks. Feature points are mapped to a virtual plane and regular geometric regions are constructed. Then, through geometric analysis of feature distribution centroids, principal directions, quadrant density / distance, and other parameters, the spatial distribution patterns of pests and diseases are mined, such as the symmetrical distribution of circular lesions and the linear layout of striped pests. Finally, spatial structure enhancement weights are generated to enhance the spatial distribution of pests and diseases. Feature enhancement and suppression of background interference effectively improve the model's ability to distinguish between pests and diseases that appear similar but have different spatial structures, reducing the false positive rate. Size normalization and color channel standardization in the preprocessing stage eliminate color shifts caused by image size differences and lighting, providing stable input for feature extraction. Weight vectors generated through geometric relationship analysis focus only on the core spatial structural features of pests and diseases, rather than disturbing apparent noise, effectively filtering irrelevant interference information from the feature map. This dual anti-interference design reduces fluctuations in recognition accuracy even under varying lighting conditions. A visual interface allows for remote viewing and operation, shortening the cycle from pest information collection to application and improving the timeliness and practicality of forecasting.

[0014] In a preferred embodiment of the present invention, step 1 includes: Step 100: Perform size normalization processing on the original crop image by scaling it to a preset fixed width and height to obtain a size-normalized image; perform color channel normalization processing on the size-normalized image by calculating the mean and standard deviation of pixel values ​​for all pixels in the red, green, and blue color channels of the size-normalized image, and perform normalization calculation on the pixel values ​​of each color channel based on the mean and standard deviation of pixel values ​​to obtain a standardized image. Specifically, this includes: acquiring original crop images through an IoT terminal device deployed in the cornfield. This terminal device integrates a high-resolution industrial camera and an IoT communication module (supporting low-power wide-area network protocols such as LoRa and NB-IoT). (As suggested), shooting parameters, such as shooting time, angle, and resolution, can be remotely configured via a cloud-based IoT platform. The specific acquisition process is as follows: on a sunny day, between 9 and 11 AM or 3 and 5 PM, avoiding direct sunlight, the camera is remotely triggered via the cloud platform to automatically capture images of the entire corn plant and partial leaf areas at an angle 50 to 80 centimeters vertically to the corn plant leaves. During the shooting process, the terminal device transmits image preview data back in real time via the IoT module. Staff can confirm image quality in real time on the cloud platform, ensuring that the corn plant area occupies no less than 80% of the image and is free of significant obstructions such as weeds or soil. After acquisition, the image undergoes preliminary preprocessing via the edge computing module of the IoT terminal device, such as... Blurry images due to shooting errors are removed, and then transmitted in JPEG format via an IoT communication module to a local IoT gateway. The gateway then aggregates and synchronizes the data to a cloud server. After acquiring the original image, a fixed width W and a fixed height H are set based on the input specifications of a lightweight convolutional neural network, such as 224×224, to match the input requirements of MobileNet or ShuffleNet. A bilinear interpolation algorithm is used to perform scaling operations. First, the coordinates (x1, y1) of each pixel (x0, y0) in the original corn plant image are calculated using a coordinate mapping formula under the target size W×H, where x1=x0×(W / original width) and y1=y0×(H / original height). The pixel values ​​of the four neighboring pixels around the fixed coordinate (x1, y1) are calculated using the bilinear interpolation formula, which calculates weights based on the distance between the neighboring pixels and the target coordinate. The weights are inversely proportional to the distances. The weighted sum is then used to obtain the pixel value of the target coordinate (x1, y1). After completing the calculation by traversing all pixels, the original corn plant image is uniformly scaled to a fixed size of W×H to generate a size-normalized image. Subsequently, the size-normalized image undergoes color channel normalization processing. This involves traversing all pixels in the image, initializing the sum of pixel values ​​for the R, G, and B channels to 0, and sequentially reading the R, G, and B channel values ​​of each pixel, adding them to the sum of the corresponding channel values ​​to obtain the sum of pixel values ​​for each channel.Let the total number of pixels in the size-normalized image be N (N=W×H). Calculate the parameters of each channel according to the formula (channel mean μ=sum of pixel values ​​in that channel / N; channel standard deviation σ=sum of squares of the differences between all pixel values ​​in that channel and μ divided by the square root of N); perform normalization calculations on the R, G, and B channel pixel values ​​of each pixel separately, using the following formulas; =(x-μ) / σ, where x is the original channel value. If σ is 0 (all pixel values ​​are the same), then directly set... =0, after channel normalization of all pixels is completed, a standardized corn plant image is obtained, which is synchronized to the cloud server through an IoT gateway.

[0015] Step 101: Input the standardized image into a pre-trained lightweight convolutional neural network. The pre-trained lightweight convolutional neural network contains multiple cascaded convolutional blocks. The standardized image is sequentially passed through multiple cascaded convolutional blocks for hierarchical feature extraction, determining the output feature maps of three convolutional blocks at different depths. These three convolutional blocks at different depths correspond to the shallow, medium, and deep convolutional blocks of the pre-trained lightweight convolutional neural network, respectively. Specifically, this includes: selecting MobileNet as the basic architecture to construct a lightweight convolutional neural network for corn pest and disease identification. This network consists of multiple convolutional blocks connected in series. Since the input is an RGB format corn plant image, the number of input channels for the first convolutional block is set to 3 to match the three color channels of the image. The structure and operation process of each convolutional block are as follows: The number of convolutional kernels in the convolutional layers is set according to the feature map channel requirements, with a fixed 32 kernels for shallow convolutional blocks and a fixed 512 kernels for deep convolutional blocks. The kernel size is either 3×3 or 1×1, and only one size is used per convolutional block. The kernel weights are initialized using Xavier, with values ​​ranging from -0.5 to 0.5, specifically randomly sampled from the mean 0 and standard deviation. The input feature map follows a normal distribution. For example, when a 3×3 convolution kernel has 3 input channels and 32 output channels, the standard deviation is about 0.15, and the weights are distributed in the interval [-0.45, 0.45]. The input feature map is transformed and local features are extracted through convolution operation. The output pixel value of the 3×3 convolution kernel is the sum of the product of the pixel value and the weight in the corresponding 3×3 region. The operation process adopts a zero-padding strategy. The 3×3 convolution kernel is padded with 1 layer, and the 1×1 convolution kernel is padded with 0 layers to ensure that the size of the output feature map is completely consistent with that of the input feature map.

[0016] The batch normalization layer receives the feature maps output by the convolutional layer. It first calculates the mean and variance of each channel, and then obtains a normalized feature map through normalization. A scaling factor and an offset factor are introduced. The initial value of the scaling factor is set to 1.0, with a value range of [0.1, 2.0], and it stabilizes in the range of [0.8, 1.2] after training. The initial value of the offset factor is set to 0.0, with a value range of [-1.0, 1.0], and it stabilizes in the range of [-0.2, 0.2] after training. After adjustment by the linear adjustment formula, the pixel value = normalized pixel value × scaling factor + offset factor, so that the mean of each channel of the feature map is stabilized at 0 and the variance is stabilized at 1, thereby accelerating the convergence of network training.

[0017] The activation function layer uses the ReLU activation function to perform activation operations on the batch normalized pixel values. If the pixel value is greater than 0, the original value is retained, and the value range is approximately [-3, 3]. If it is less than or equal to 0, it is set to 0. Through this nonlinear transformation, the network's ability to represent the complex characteristics of corn diseases and pests is enhanced.

[0018] The network training phase adopts an IoT collaborative mode, which involves acquiring multiple batches of corn plant image data through IoT acquisition terminals distributed in different corn planting areas, following the acquisition method in step 100. This includes healthy corn plants and plants affected by various pests and diseases, such as leaf spot and rust. All images are aggregated to the cloud training server via the IoT platform, ensuring that the number of images for each category meets the training requirements. The LabelImg annotation tool is used to annotate the images in the cloud, drawing rectangular boxes to locate pest and disease areas, labeling them with categories, and generating XML format files stored in the cloud database. The annotated dataset is divided into training and validation sets in a 7:3 ratio (using stratified sampling to ensure consistent image proportions for each category in both sets). MD5 verification is used to ensure no duplicate images, avoiding uneven data distribution that could negatively impact training performance.

[0019] The initialization and setting process of training parameters is as follows: Convolutional layer weights are initialized using Xavier, with values ​​ranging from [-0.5, 0.5]. For example, with a 1×1 convolutional kernel having 32 input channels and 64 output channels, the standard deviation is approximately 0.15, and the weights are distributed in the range of [-0.45, 0.45]. The initial value of the batch normalization layer scaling factor is 1.0, with a value range of [0.1, 2.0], and the initial value of the offset factor is 0.0, with a value range of [-1.0, 1.0]. All layer bias parameters are uniformly set to 0.0. Training hyperparameter settings are then implemented, including a batch size of 32. It is compatible with GPUs with at least 4GB of video memory; the initial learning rate is 0.001, ranging from [0.0001, 0.01], and decreases to 0.1 times the original rate every 20 iterations; the maximum number of iterations is 100; in each iteration, 32 images and their corresponding labels are randomly read from the training set, scaled to 224×224 using bilinear interpolation, converted to float32 type tensors (pixel values ​​normalized to the range [0, 1]) and input into the network. The images are then processed sequentially through convolution operations (extracting local features), batch normalization, and ReLU in each convolutional block. Activation (introducing non-linearity) processing is used, and the feature map dimension is gradually adjusted with network depth (increasing the number of channels and decreasing the resolution). The final features are converted into one-dimensional vectors through global average pooling and input into a fully connected layer. Weights are initialized using Xavier, with values ​​ranging from -0.3 to 0.3. The predicted class probabilities are calculated using the Softmax function, ensuring the sum of all class probabilities is 1. The results are rounded to four decimal places, e.g., 0.0215 for healthy corn and 0.9785 for large leaf spot. The ground truth labels are then converted to one-hot encoding, with the corresponding class set to 1 and the rest set to 0. After calculating the loss value of a single image using the cross-entropy loss function, the average loss of all images in the batch is taken. The gradient of each layer's parameters is calculated using the backpropagation algorithm, with a value range of [-0.1, 0.1] and the absolute value of the gradient not exceeding 0.5. The parameters are updated using stochastic gradient descent, and the new parameter value = current parameter value - learning rate × gradient. For example, when the current parameter value is 0.2, the learning rate is 0.001, and the gradient is -0.0033, the new parameter value = 0.2 - 0.001 × (-0.0033). Momentum (momentum coefficient 0.9) and adaptive learning rate (0.9) are combined.999) Update parameters to ensure they remain within the preset range. After each iteration, calculate the prediction probability using validation set images and labels, and count the number of correct predictions. When counting correct predictions, for each validation set image, select the category with the highest prediction probability as the prediction category. If this category is completely consistent with the image's true label, it is counted as one correct prediction. After traversing all validation set images, sum the results to obtain the total number of correct predictions. When calculating the validation set classification accuracy, divide the total number of correct predictions by the total number of validation set images, and keep the result to two decimal places. For example, if there are 1400 images in the validation set and 1260 are correctly predicted, the accuracy = 1260 / 1400. When calculating the validation set loss value, convert the true labels into one-hot encodings. The cross-entropy loss function is used. The loss of each image is calculated across all validation set images, and the average value is taken, rounded to four decimal places. Training stops when the absolute value of the difference between consecutive validation set loss values ​​is less than 1e-4 for 5-10 consecutive rounds (e.g., loss of 0.1234 in round 20 and 0.1233 in round 21, with a difference of 0.0001), and the classification accuracy remains above 90.00%. All network parameters at this point (convolutional layer weights, batch normalization layer scaling and offset factors, fully connected layer weights, etc., all rounded to six decimal places) are saved, resulting in a trained lightweight convolutional neural network. This network is then distributed to various field IoT terminal devices via an IoT platform for local feature extraction.

[0020] The standardized corn image obtained in step 100, with a size of 224×224 and pixel values ​​ranging from [-3, 3], is input into the pre-trained network. After the image passes through all convolutional blocks in sequence, feature maps (all four-dimensional tensors, float32 type, with values ​​ranging from [0, 3]) are extracted from convolutional blocks at three different depths. Specifically, shallow convolutional blocks are located in the first 1 / 3 of the network depth. For example, when the network has a total of 18 convolutional blocks, the first 6 are shallow convolutional blocks. The output feature map resolution is 112×112 (half the size of the input image), with 32 or 64 channels and pixels distributed in the range of [0, 2.5]. This feature map can retain detailed information such as corn leaf texture and the edges of small lesions. Middle convolutional blocks are located in the middle 1 / 3 of the network depth. For example, when the network has a total of 18 convolutional blocks, the first 6 are shallow convolutional blocks. The output feature map resolution is 112×112 (half the size of the input image), with 32 or 64 channels and pixels distributed in the range of [0, 2.5]. This feature map can retain detailed information such as corn leaf texture and the edges of small lesions. When the network has a total of 18 convolutional blocks, the 7th to 12th are mid-level convolutional blocks, with an output feature map resolution of 56×56 (1 / 4 of the input image), 128 or 256 channels, and pixels distributed in the range of [0, 2.0], containing local features such as lesion outlines and leaf vein structures; deep convolutional blocks are located in the last 1 / 3 of the network depth. For example, when the network has a total of 18 convolutional blocks, the 13th to 18th are deep convolutional blocks, with an output feature map resolution of 28×28 or 14×14 (1 / 8 or 1 / 16 of the input image, respectively), 512 channels, and pixels distributed in the range of [0, 1.5], containing global semantic information such as lesion areas and healthy areas; the extracted three types of feature maps are transmitted back to the cloud server in real time through the IoT module, and are also cached and backed up on the local terminal device.

[0021] Step 102 involves combining the output feature maps of the shallow, mid, and deep convolutional blocks into a multi-scale feature map. Specifically, this includes integrating the output feature maps of the shallow, mid, and deep convolutional blocks extracted in Step 101 using tensor stitching (before stitching, the IoT collaboration module on the cloud server ensures that the feature map data uploaded by each field terminal has a unified format). This involves first performing resolution adaptation on the mid and deep feature maps (based on the edge computing capabilities of the cloud server, bilinear interpolation is used to adjust the resolution of the mid and deep feature maps to match that of the shallow feature maps). Figure 1The three types of feature maps are then stitched together along the channel dimension (only expanding the channel dimension, maintaining spatial dimension consistency). Together, they form the multi-scale feature map required for corn disease and pest identification. The specific feature types and functions of each feature map are as follows: The shallow convolutional block output feature map retains detailed features of the corn plant image, such as parallel textures on the corn leaf surface, edges of small lesions caused by pests, and outlines of small borer holes left by corn borers. This helps identify early-stage small pests and diseases. This type of feature is preferentially transmitted to the local terminal via the IoT module, supporting rapid preliminary judgment. The middle convolutional block output feature map contains local shape features of the corn plant image, such as the overall outline of lesions on corn leaves, local structure of leaf veins, and local distribution of pest and disease areas. This helps determine the local morphology and distribution patterns of pests and diseases. This type of feature is similar to... The data is stored on both the local terminal and the cloud server for local inference and secondary verification in the cloud. The feature map output by the deep convolutional block contains global semantic features of the corn plant image, such as global information on corn lesion areas, healthy leaf areas, and pest-infected areas. It can distinguish between pest-infected areas and background areas, such as leaves and stems, providing a global basis for pest classification. This type of feature is mainly stored on the cloud server and is combined with multi-region feature data for collaborative analysis. The integrated multi-scale feature map is synchronized to the cloud pest monitoring database through IoT communication protocols (such as MQTT), while a lightweight copy is kept on the local IoT terminal device. The cloud server indexes and annotates the feature map (associating it with auxiliary data returned by the IoT terminal, such as collection time, collection device location, and field environmental parameters), forming a complete pest data package.

[0022] In this embodiment, size normalization unifies the size of all input images, avoiding size differences caused by different shooting angles and distances of the original images. Color channel normalization eliminates the influence of lighting changes on image color, ensuring that the pixel values ​​of each channel are within a uniform data distribution range, reducing feature extraction fluctuations caused by input differences. By constructing a lightweight network containing multiple convolutional blocks, the hierarchical feature extraction capability is retained while using small-sized convolutional kernels and optimizing the network structure. The pre-training process utilizes a large-scale dataset to optimize network parameters, enabling the network to possess preliminary crop image feature recognition capabilities, avoiding the slow convergence and poor generalization problems caused by training from scratch. The integrated multi-scale feature maps cover different levels of features required for crop disease and pest identification. This multi-dimensional feature data avoids recognition bias caused by single features, improving recognition accuracy.

[0023] In a preferred embodiment of the present invention, step 2 includes: Step 200: Define the feature vector at each spatial location in the multi-scale feature map as a feature point. The set of all feature points constitutes the feature point set. Specifically, this includes: obtaining the multi-scale feature maps (shallow layer 112×112×32 / 64, middle layer 56×56×128 / 256, deep layer 28×28×512 or 14×14×512) extracted layer by layer from the corn plant image by the pre-trained network in step 101. These feature maps contain detailed features of corn diseases and pests, such as the edges of small lesions, local shape features such as lesion outlines, and global semantic features such as the distribution of lesion areas. For each spatial location of each feature map, taking the shallow feature map as an example, the spatial location coordinates are (i, j), where the high-dimensional feature vector corresponding to i∈[0, 111] and j∈[0, 111] is defined as a feature point. This vector integrates the feature information of the maize plant at the corresponding spatial location. For example, the 32-dimensional or 64-dimensional feature vector corresponding to a certain spatial location (i, j) in the shallow feature map contains the features of maize leaf texture and traces of minor pests and diseases at that location. By traversing all spatial locations of all multi-scale feature maps and collecting all feature vectors, a set of feature points covering different levels of maize pests and diseases is formed.

[0024] Step 201 involves linearly transforming each feature point in the feature point set through a fully connected layer in a pre-trained lightweight convolutional neural network. This projects the high-dimensional feature vector of the feature point onto a two-dimensional coordinate system, obtaining the two-dimensional projected coordinates of each feature point on a virtual feature plane. Specifically, this includes using the fully connected layer in the lightweight convolutional neural network pre-trained in step 101 (used for predicting corn pest and disease categories) to perform a linear transformation on each feature point in the feature point set, thus mapping the high-dimensional features to a two-dimensional space. Let the high-dimensional feature vector of a certain feature point be X, with dimension C, where C is the number of channels in the corresponding feature map, such as 32. 64, 128, etc., correspond to different levels of maize disease and pest feature dimensions. The weight matrix of the fully connected layer is W, with a dimension of 2×C. The weight values ​​follow Xavier initialization and range [-0.3, 0.3]. The bias vector is b, with a dimension of 2 and a fixed value of 0.0. Then the linear transformation formula is Y=W×X+b, where Y is a two-dimensional vector (y1, y2), y1 is the horizontal coordinate value, and y2 is the vertical coordinate value. Through this linear transformation, the high-dimensional vector X containing maize disease and pest features is projected onto a two-dimensional coordinate system to obtain the two-dimensional projected coordinates (y1, y2) of each feature point on the virtual feature plane.

[0025] Step 202: Based on the two-dimensional projected coordinates of all feature points in the feature point set, calculate the minimum and maximum values ​​of the two-dimensional projected coordinates in the horizontal and vertical directions, determine the minimum bounding rectangle that can cover the projected coordinates of all feature points, and use the minimum bounding rectangle as a regular geometric structure. Specifically, this includes: traversing the two-dimensional projected coordinates (y1, y2) of all feature points in the feature point set, and calculating the minimum value in the horizontal direction for each feature point. (Minimum value among all y1 values) and maximum value (Maximum value among all y1 values), minimum value in the ordinate direction. (Minimum value among all y² values) and maximum value (The largest value among all y² values), with For the left boundary, For the right boundary, For the lower boundary, Construct a rectangle with the upper boundary that completely encloses the projected coordinates of all corn disease and pest feature points. This rectangle is the smallest bounding rectangle that can cover the projected coordinates of all feature points, and it is used as a regular geometric structure.

[0026] Step 203: Divide the minimum bounding rectangle into equally spaced rectangular grids along its length and width, each grid representing a sub-region. Specifically, determine the number of grid divisions N, where N is the number of equal divisions along both the length and width directions of the minimum bounding rectangle. The number of divisions in both directions must be consistent, i.e., each area is divided into N equal parts. The value of N needs to be dynamically set based on the density of corn disease and pest feature points in the feature point set. A higher feature point density means a higher number of points per unit area containing detailed features such as corn leaf texture and lesion edges. When there are many features, N should be a larger value, such as 8 or 16, to ensure that each sub-region contains an appropriate number of pest and disease feature points, avoiding analysis bias caused by excessive density of local lesion features or healthy area features. When the feature point density is low, that is, when the number of pest and disease related feature points per unit area is small, N should be a smaller value, such as 4 or 8, to prevent some sub-regions from lacking corn pest and disease feature points, avoiding waste of healthy and pest and disease feature information. Calculate the geometric parameters of the minimum bounding rectangle, which encloses the two-dimensional projected coordinates of all corn pest and disease feature points. The length L of the minimum bounding rectangle is the difference between the maximum and minimum values ​​of the x-coordinate, i.e., L = - The width W of the minimum bounding rectangle is the difference between the maximum and minimum values ​​of the ordinate, i.e., W = - Based on the number of divisions N, calculate the length ΔL and width ΔW of each sub-region, where ΔL=L / N and ΔW=W / N.

[0027] Along the length of the minimum bounding rectangle, the x-coordinate starts from... arrive The distribution range of corn pest and disease characteristic points in the horizontal direction is covered, and the area is divided into dividing lines at intervals of ΔL. The horizontal coordinates of the dividing lines are as follows: +ΔL、 +2ΔL、…、 +(N-1)ΔL; along the width direction (vertical coordinate from...) arrive (Covering the vertical distribution range of corn pest and disease characteristic points), dividing the area into segments according to ΔW intervals, with the vertical coordinates of the segments being... +ΔW、 +2ΔW、…、 +(N-1)ΔW, through the intersection of the above horizontal and vertical dividing lines, divides the smallest bounding rectangle containing the projection of all corn disease and pest feature points into N×N rectangular grids of equal size, each rectangular grid being a sub-region.

[0028] This embodiment integrates the spatial location feature vectors of multi-scale feature maps to uniformly transform the detailed features of corn diseases and pests, such as leaf texture, small lesions, and local shape features, such as lesion outlines, into feature points, along with global semantic features, such as the distribution of disease and pest areas. This avoids missing key information due to single-scale features. High-dimensional feature dimensionality reduction is achieved using the fully connected layers of a pre-trained lightweight convolutional neural network, reducing computational costs. Simultaneously, the high-dimensional features are mapped to a two-dimensional plane, causing points with similar disease and pest features to cluster on the plane, facilitating the intuitive capture of spatial correlation patterns among disease and pest features. This improves the efficiency of feature analysis. By calculating the extreme values ​​of the projected coordinates, the minimum bounding rectangle is determined, ensuring that the rectangle tightly encloses all pest and disease feature points. This avoids wasting space resources and fully covers the distribution range of pest and disease features, guaranteeing the integrity of feature analysis after segmentation. The rectangle is divided into uniform sub-regions at equal intervals, so that local pest and disease features are concentrated in each sub-region. This allows for targeted analysis of feature correlations in different sub-regions, such as local lesion density and the boundary features between healthy and pest and disease areas. This reduces interference from feature mixing in the global analysis and improves the accuracy of local representation of pest and disease features.

[0029] In a preferred embodiment of the present invention, step 3 includes: Step 300: Based on the two-dimensional projected coordinates of each feature point and the boundary coordinates of multiple rectangular grid sub-regions, determine the corresponding rectangular grid to which each feature point belongs. Group and aggregate the high-dimensional feature vectors of all feature points belonging to the same rectangular grid. Specifically, this includes: obtaining the two-dimensional projected coordinates (y1, y2) of each corn disease and pest feature point obtained in step 201 and the boundary coordinates of the rectangular grid sub-regions determined in step 203. The lateral boundary of each sub-region is […]. , ]( The left x-coordinate of the sub-region (where x is the right x-coordinate of the sub-region), and the vertical boundary is [ , ]( The lower ordinate of the sub-region. (The upper ordinate of the sub-region); for each feature point (containing feature information such as corn leaf texture and lesion outline), determine whether its projected coordinates satisfy... ≤y1≤ and ≤y2≤ If the condition is met, the feature point belongs to the sub-region. Traverse all feature points and store the high-dimensional feature vectors (derived from the multi-scale maize disease and pest features in step 200) corresponding to the feature points belonging to the same rectangular grid, forming the feature vector group of the grid, thereby realizing the aggregation of disease and pest features in the local area.

[0030] Step 301: Perform mean pooling and max pooling on the high-dimensional feature vectors within each rectangular grid group to obtain the mean pooling feature vector and max pooling feature vector corresponding to each rectangular grid. Specifically, this includes: performing pooling on the feature vector group of each rectangular grid, which contains corn pest and disease related features within that region. When performing mean pooling, suppose a certain grid's feature vector group contains M high-dimensional vectors, each vector containing pest and disease features at that location, and each vector has a dimension of C, such as 32, 64, etc., where the i-th element of the k-th vector is... ( =1, 2, ..., M; i = 1, 2, ..., C); the i-th element of the average pooling feature vector is calculated as the average of the corresponding elements of all vectors. This vector reflects the overall distribution trend of maize pest and disease characteristics in the region, such as the average lesion intensity, and finally obtains the average pooling feature vector with dimension C; when using max pooling, the i-th element of the max pooling feature vector is calculated as the maximum value of the corresponding elements of all vectors, i.e. This vector highlights the most significant maize pest and disease characteristics within the region, such as the characteristics of the largest lesions, and ultimately yields a max-pooling feature vector of dimension C.

[0031] Step 302 involves concatenating the average pooling feature vector and the max pooling feature vector of each rectangular grid along the channel dimension to form the region aggregation feature vector for each rectangular grid. Specifically, this includes: for each rectangular grid, concatenating the average pooling feature vector and the max pooling feature vector obtained in step 301 along the channel dimension. Let the average pooling vector be... It reflects the overall trend of pest and disease characteristics in the region, and the max pooling vector is This reflects the most significant pest and disease characteristics within the region, and the resulting region aggregated feature vector after splicing is [ The dimension is 2C.

[0032] Step 303: Organize and arrange the region aggregation feature vectors of all rectangular grids according to the row and column order of the grids to form a set of region aggregation features. Specifically, this includes: arranging the region aggregation feature vectors of all grids sequentially from left to right and from top to bottom according to the row and column order of the rectangular grids in step 203, to form a complete set of region aggregation features; for example, the feature vectors of N×N grids are arranged in the order of grids (1,1), (1,2), ..., (1,N), (2,1), ..., (N,N) to form a complete feature set containing N×N region features, which fully preserves the spatial distribution information of corn pest and disease features on the virtual plane, such as the location association of lesion areas and the boundary distribution between healthy areas and pest and disease areas.

[0033] In this embodiment, grouping is achieved through feature point attribution judgment, which concentrates the spatially distributed maize disease and pest features, strengthens the correlation of local features, and avoids information fragmentation caused by feature dispersion. The combination of mean pooling and max pooling can capture the overall distribution trend of disease and pest features in the region, such as the average density of lesions, and highlight key salient features, such as the features of the largest lesion area, thereby improving the comprehensiveness of feature representation. Channel splicing integrates the two types of pooling features, combining the overall rules and prominent features of local features, and enhancing the ability to express the complex features of maize diseases and pests.

[0034] In a preferred embodiment of the present invention, step 4 includes: Step 400: Based on the regional aggregation features, a feature correction vector is generated through geometric relationship analysis. Specifically, step 400a involves calculating the arithmetic mean of the x-coordinates and y-coordinates of all two-dimensional projected coordinates in the feature point set. The points corresponding to these arithmetic mean values ​​are used as the centroids of the feature distribution. This includes obtaining the two-dimensional projected coordinates of all corn disease and pest feature points in the feature point set obtained in step 201. Each coordinate corresponds to a spatial location containing disease and pest features. The coordinates of each feature point contain both x-coordinate and y-coordinate values. The total number of all feature points is counted, i.e., the number of points containing disease and pest features. The total number of spatial locations of pest-related features is calculated by summing the horizontal coordinates of all feature points (accumulating the horizontal coordinates of all pest-related features) and dividing by the total number of feature points to obtain the arithmetic mean of the horizontal coordinates. Similarly, the total number of vertical coordinates of all feature points is calculated by summing the vertical coordinates of all pest-related features and dividing by the total number of feature points to obtain the arithmetic mean of the vertical coordinates. The point corresponding to the arithmetic mean of the horizontal and vertical coordinates is defined as the centroid of the distribution of maize pest-related features. This point reflects the concentrated distribution location of maize pest-related features on the virtual plane and can preliminarily locate the core distribution area of ​​pest-related features.

[0035] Step 400b: Establish a temporary coordinate system with the centroid of the feature distribution as the origin, and calculate the covariance matrix of all two-dimensional projected coordinates in the temporary coordinate system. Specifically, this includes: establishing a temporary coordinate system with the centroid of the maize disease and pest feature distribution obtained in step 400a as the origin. The horizontal axis of the temporary coordinate system corresponds to the horizontal direction of the original two-dimensional projected coordinates, and the vertical axis corresponds to the vertical direction of the original two-dimensional projected coordinates, ensuring consistency with the original spatial distribution direction of the feature points; for each maize disease and pest feature point, subtract the arithmetic mean of the centroid's horizontal coordinate from its original horizontal coordinate value to obtain the horizontal coordinate of the feature point in the temporary coordinate system, reflecting the horizontal offset of the feature point relative to the centroid. The magnitude of the offset is related to the horizontal distance between the disease and pest feature and the core area; subtract the arithmetic mean of the centroid's vertical coordinate from its original vertical coordinate value to obtain the vertical coordinate of the feature point in the temporary coordinate system, reflecting the vertical offset of the feature point relative to the centroid; based on the covariance of all maize disease and pest feature points... The covariance matrix is ​​calculated using time coordinates. This matrix is ​​a 2x2 matrix, and the elements are calculated as follows: The element in the first row and first column is the sum of the squares of the temporary horizontal coordinates of all feature points (the sum of the squares of the horizontal offsets of all pest and disease feature points relative to the centroid), divided by (total number of feature points minus 1), reflecting the dispersion of pest and disease feature points in the horizontal direction; The element in the first row and second column is the sum of the products of the temporary horizontal and vertical coordinates of all feature points (the sum of the products of the horizontal and vertical offsets of pest and disease feature points), divided by (total number of feature points minus 1), reflecting the correlation between the horizontal and vertical distribution of pest and disease feature points; The element in the second row and first column is equal to the element in the first row and second column; The element in the second row and second column is the sum of the squares of the temporary vertical coordinates of all feature points (the sum of the squares of the vertical offsets of all pest and disease feature points relative to the centroid), divided by (total number of feature points minus 1).

[0036] Step 400c involves performing eigenvalue decomposition on the covariance matrix, extracting the eigenvector corresponding to the largest eigenvalue as the first principal direction, and extracting the eigenvector corresponding to the second largest eigenvalue as the second principal direction; using the line containing the first principal direction as the first reference axis and the line containing the second principal direction as the second reference axis. Specifically, this includes performing eigenvalue decomposition on the covariance matrix obtained in step 400b to obtain two eigenvalues ​​and two corresponding eigenvectors, where the first eigenvalue is greater than or equal to the second eigenvalue (the magnitude of the eigenvalue reflects the dispersion of the pest and disease characteristics distribution in the corresponding direction); defining the eigenvector corresponding to the first eigenvalue as the first... The primary direction is the direction in which the distribution of maize disease and pest feature points is most dispersed (i.e., the range of disease and pest features is widest in this direction, possibly corresponding to the main direction of lesion spread). The eigenvector corresponding to the second eigenvalue is defined as the second primary direction, which is the direction in which the distribution of maize disease and pest feature points is less dispersed (corresponding to the secondary direction of disease and pest feature extension). The straight line passing through the centroid of the maize disease and pest feature distribution and extending along the first primary direction is defined as the first reference axis, and the straight line passing through the centroid of the feature distribution and extending along the second primary direction is defined as the second reference axis. The two reference axes are perpendicular to each other and together constitute a reference system describing the primary direction of the distribution of maize disease and pest feature points.

[0037] Step 400d: Using the centroid of the feature distribution as the vertex and the first and second reference axes as boundaries, divide the feature projection plane into four quadrants. Count the number of feature points contained in each quadrant and calculate the feature point density of each quadrant. Calculate the average Euclidean distance from all feature points in each quadrant to the centroid of the feature distribution. Specifically, using the centroid of the maize disease and pest feature distribution from step 400a as the vertex and the first and second reference axes from step 400c as boundaries, divide the feature projection plane into four quadrants (each quadrant corresponds to a sub-region of disease and pest features distributed in different principal directions). The first quadrant is the region between the positive direction of the first reference axis and the positive direction of the second reference axis, corresponding to the positive superposition region of the main and secondary extension directions of disease and pest features. The second quadrant is the region between the negative direction of the first reference axis and the positive direction of the second reference axis. The third quadrant is the region between the negative direction of the first reference axis and the negative direction of the second reference axis. The fourth quadrant is the region between the positive direction of the first reference axis and the negative direction of the second reference axis. The region between directions; traverse all maize disease and pest feature points, determine their quadrant based on their temporary coordinates (offset relative to the centroid) (determine the specific distribution sub-region of each disease and pest feature point in the main direction reference system), count the number of feature points in each quadrant, i.e., the number of disease and pest features contained in each sub-region, and the sum of the number of feature points in the four quadrants equals the total number of feature points; calculate the feature point density of each quadrant by dividing the number of feature points in that quadrant by the area of ​​that quadrant. Since the two reference axes are perpendicular to each other, the areas of the four quadrants are equal, each being one-quarter of the total area of ​​the feature projection plane (the density reflects the density of disease and pest features in each sub-region, and quadrants with high density may correspond to areas with concentrated lesions); for maize disease and pest feature points in each quadrant, calculate the Euclidean distance from each feature point to the centroid, add up the Euclidean distances of all feature points in that quadrant, and then divide by the number of feature points in that quadrant to obtain the average Euclidean distance of each quadrant, reflecting the overall concentration of disease and pest features in that sub-region relative to the core region.

[0038] Step 400e: Concatenate the feature point density and average Euclidean distance of the four quadrant regions in quadrant order to form an eight-dimensional spatial distribution feature vector. Perform linear transformation and dimensionality compression on the eight-dimensional spatial distribution feature vector to generate a feature correction vector. Specifically, this includes: arranging the feature point density of the four quadrants in the order of the first to fourth quadrants, and then arranging the average Euclidean distance of the four quadrants in the same order. Concatenate the two sets of data sequentially to form an eight-dimensional spatial distribution feature vector containing eight values ​​(this vector fully characterizes the global spatial distribution pattern of corn pest and disease features); call the linear transformation layer of the pre-trained lightweight convolutional neural network, with the preset correction dimension set to 8. Therefore, the weight matrix has a dimension of 8×8 and a weight value range of [-0.2, 0.2]. The bias vector has a dimension of 8 and a bias value range of [-0.1, 0.1]. A linear transformation is performed on this eight-dimensional vector. The transformation process is as follows: the weight matrix of the linear transformation layer is multiplied by the eight-dimensional vector, and the bias vector of the same layer is added to obtain the vector before correction (dimension of 8 and element value range of [-1.8, 1.8]). Subsequently, the vector before correction is compressed to 2 dimensions (removing redundant information) using the PCA dimensionality reduction method to generate the feature correction vector (vector element value range of [-1.5, 1.5]). This vector condenses the spatial distribution pattern of corn pest and disease characteristics.

[0039] Step 401: The feature correction vector and the regional aggregation feature are fused to obtain the geometrically enhanced regional aggregation feature. Specifically, the feature correction vector obtained in step 400e, which contains the global spatial distribution pattern of maize pests and diseases, is a 2-dimensional vector. The regional aggregation feature obtained in step 303, which contains local pests and diseases, such as the edge of lesions and leaf texture details, is 2×C, where C is the number of channels of the corresponding feature map. For example, 32 channels in the shallow layer correspond to 64 dimensions, and 512 channels in the deep layer correspond to 1024 dimensions. After splicing, a geometrically enhanced regional aggregation feature containing (2×C+2) values ​​is formed, with the vector element values ​​ranging from [-2.0, 2.0].

[0040] Step 402: Input the geometrically enhanced region aggregation features into the weight generation network of the pre-trained lightweight neural network to obtain the initial weight vector. Specifically, this includes: inputting the geometrically enhanced region aggregation features obtained in step 401, fused with local features and global spatial distribution features of maize diseases and pests, with a dimension of 2×C+2, where C is the number of channels corresponding to the multi-scale feature map, into the weight generation network of the pre-trained lightweight neural network. This network is a 3-layer fully connected structure, and the weights of all layers are optimized through pre-training on the maize disease and pest identification task. The parameters of each layer are set as follows: the first fully connected layer weight matrix dimension... The first fully connected layer has a dimension of 128×(2×C+2), with 128 rows corresponding to the output feature dimension and 2×C+2 columns matching the dimension of the geometrically enhanced region aggregation feature of the input. The weight values ​​range from [-0.15, 0.15]. The bias vector dimension is 128, and the bias value is uniformly set to 0.0. The second fully connected layer has a weight matrix dimension of 64×128, with 64 rows corresponding to the output feature dimension and 128 columns matching the output dimension of the first fully connected layer. The weight values ​​range from [-0.15, 0.15]. The bias vector dimension is 64, and the bias value is uniformly set to 0.0. The third fully connected layer... The weight matrix has a dimension of C×64, with C rows corresponding to the number of channels in the multi-scale feature map and 64 columns matching the output dimension of the second fully connected layer. The weight values ​​range from [-0.1, 0.1]. The bias vector has a dimension of C, and the bias values ​​are uniformly set to 0.0. After geometric enhancement, the region aggregation features are first input into the first fully connected layer, where a linear transformation is performed. The 128×(2×C+2) weight matrix of this layer is multiplied by the input features to obtain 128-dimensional intermediate features. Then, the 128-dimensional bias vector of this layer is added, with all bias values ​​set to 0.0, completing the linear transformation. Finally, Re is applied to the linear transformation result. LU activation is performed to obtain 128-dimensional features from the first layer output. These 128-dimensional features are then fed into the second fully connected layer, and the above operation is repeated to obtain 64-dimensional features from the second layer output. The 64-dimensional features from the second layer output are then fed into the third fully connected layer, where a linear transformation is performed. This involves multiplying the 64-dimensional features by a C×64 weight matrix and adding a C-dimensional bias vector with a bias value of 0.0. No activation is required, and a vector of dimension C is directly output. This vector is the initial weight vector, with element values ​​ranging from -0.8 to 0.8. The dimension is exactly the same as the number of channels C in the multi-scale feature map.

[0041] Step 403 involves inputting the initial weight vector into a normalization function for processing to obtain a numerically normalized spatial structure enhancement weight vector. Specifically, this includes inputting the initial weight vector from step 402 (used to adjust the intensity of pest and disease features, with dimension C) into the Sigmoid normalization function. The result is the normalized value. This function maps the value of each element in the initial weight vector to the interval [0.2, 0.8] to obtain a numerically normalized spatial structure enhancement weight vector with dimension C. The normalization process ensures that the weight values ​​are appropriate, avoiding excessive enhancement of pest and disease features (masking detailed information) due to excessively large weights, or insufficient feature enhancement (failing to highlight key pest and disease information) due to excessively small weights.

[0042] Step 404: Expand the channel dimension of the spatial structure enhancement weight vector to match the number of channels in the multi-scale feature map, obtaining the expanded spatial structure enhancement weight vector; perform channel-by-channel product operation between the expanded spatial structure enhancement weight vector and the multi-scale feature map to obtain the enhanced multi-scale features. Specifically, this includes: expanding the channel dimension of the spatial structure enhancement weight vector obtained in step 403, with dimension C and element values ​​ranging from [0.2 to 0.8], so that the number of channels in the expanded weight vector completely matches the number of channels in the multi-scale feature map extracted in step 101. If it is a shallow feature map (32 or 64 channels) and the weight vector dimension C matches the number of channels in the feature map, then the weight vector is directly used; if it is a deep feature map (512 channels) and the weight vector dimension C is less than 512, then the weight vector elements are cyclically repeated until the number of channels reaches 512, ensuring that the expanded weight vector... Each element remains within the range of [0.2, 0.8]. After expansion, the expanded weight vector is multiplied by the multi-scale feature map through each channel. For each channel of the feature map (each channel corresponds to a type of corn pest and disease feature information, such as lesion texture and edge contour), all pixel values ​​of that channel are taken (reflecting the intensity of the corresponding feature at different spatial locations, in the range of [0, 3]). Each pixel value is multiplied by the weight value of the corresponding channel of the weight vector, in the range of [0.2, 0.8], to obtain the enhanced pixel value of that channel, in the range of [0, 2.4]. Channels with high weight values ​​have their corresponding pest and disease feature intensity amplified; channels with low weight values ​​have irrelevant background information suppressed. After traversing all channels and completing the operation, the multi-scale feature with enhanced spatial structure is obtained. This feature integrates the original pest and disease features with the spatial structure weight information, which can highlight the key regional features of corn pests and diseases, such as lesion concentration areas and typical disease texture areas.

[0043] This embodiment captures the global spatial distribution patterns of maize disease and pest characteristics by calculating the centroid, covariance matrix, and principal direction. Quadrant statistics and feature correction vector generation transform spatial distribution differences into quantifiable features, avoiding indiscriminate processing of uniformly distributed features. By integrating regional aggregation features and spatial distribution features, it preserves local disease and pest details, such as lesion edges and textures, while supplementing global positional correlations, such as the distribution of lesion clusters, thus solving the problem of one-sided information from a single feature dimension and improving the completeness of feature representation. The weight generation network combined with normalization processing ensures that the generated weight vectors are highly adaptable to maize disease and pest characteristics. The normalized weights can precisely control the enhancement degree of features in different channels, avoiding feature distortion. Channel-by-channel product operations realize the spatial structure enhancement of multi-scale features, highlighting key disease and pest areas, such as the feature response of high-density lesion areas, and suppressing background interference, such as healthy leaf areas.

[0044] In a preferred embodiment of the present invention, step 5 includes: Step 500 involves concatenating the shallow, mid-level, and deep feature maps from the enhanced spatial structure multi-scale features by channel dimension to form preliminary fused features. Specifically, this includes: first, obtaining the enhanced spatial structure multi-scale features obtained in step 404. This feature contains three types of feature maps: shallow feature maps with 32 or 64 channels and a spatial resolution of 112×112, corresponding to detailed features of maize diseases and pests, such as lesion edge texture and micro-spot outlines; mid-level feature maps with 128 or 256 channels. The spatial resolution is 56×56, corresponding to local features of pests and diseases, such as the overall outline of lesions and the extent of leaf damage. The deep feature map, with a fixed number of channels (512) and a spatial resolution of 28×28 or 14×14, corresponds to global features of pests and diseases, such as the distribution pattern of lesions on leaves and the overall boundary between healthy and diseased areas. Because the three types of feature maps have different spatial resolutions, upsampling is required on the middle and deep feature maps to unify the resolution. Upsampling of the middle feature map uses bilinear interpolation, with a resolution of 56×56. Based on the original resolution of 6, new pixels are filled by calculating the linear weighted values ​​of adjacent pixels, thus enlarging the spatial resolution to 112×112. This ensures that each pixel in the mid-level feature map can still accurately reflect the original local pest and disease characteristics after upsampling. The deep feature map upsampling also uses bilinear interpolation. If the original resolution is 28×28, it is enlarged by 4 times to 112×112; if the original resolution is 14×14, it is enlarged by 8 times to 112×112, ensuring the global pest and disease characteristics of the deep feature map after upsampling. Without losing features, after unifying the resolution, the three types of feature maps are stitched together along the channel dimension. That is, the upsampled shallow feature map, middle feature map, and deep feature map are superimposed in sequence. After stitching, the number of channels of the preliminary fused feature is the sum of the number of channels of the three types of feature maps. For example, 32 channels of shallow layer + 128 channels of middle layer + 512 channels of deep layer = 672 channels. The spatial resolution is maintained at 112×112. Finally, a preliminary fused feature is formed, which simultaneously contains the detailed, local, and global scale features of corn diseases and pests.

[0045] Step 501: Input the preliminary fused features into the convolutional layer of the pre-trained lightweight neural network for channel dimensionality reduction and feature integration to obtain optimized fused features with a uniform number of channels. Perform global average pooling on the optimized fused features to obtain a one-dimensional global feature vector. Specifically, the preliminary fused features obtained in step 500 are input into the convolutional layer of the pre-trained lightweight neural network, taking 112×112×672 as an example. This convolutional layer is used to achieve channel dimensionality reduction and cross-scale feature integration. The specific parameters and calculation process are as follows: the convolutional layer uses a 3×3 kernel, the number of kernels is set to a uniform number of channels, such as 128, the stride is set to 1, and the padding (edge ​​padding) is set to 1, padding by 1 pixel to ensure that the spatial resolution of the feature map after convolution remains 112×112. The calculation formula is: output resolution = (input resolution - kernel size + 2 × padding) / stride + 1. Each convolutional kernel corresponds to one bias value. The bias value is controlled within the range of [-0.1, 0.1] to avoid excessive bias causing feature shift. The convolution calculation process is performed on each 3×3 convolution kernel, whose depth is consistent with the number of channels of the initial fused feature, 672. During the calculation, the convolution kernel is multiplied element-wise with the local region (3×3×672) of the initial fused feature, and then all the product results are summed to obtain the convolution output value of the local region. Then, the output value is added to the bias value of the corresponding convolution kernel to complete the calculation of a single local region. All local regions of the initial fused feature are traversed with a stride of 1 to generate a 112×112 single-channel feature map. This process is repeated until all 128 convolution kernels have been calculated, and finally, an optimized fused feature of 112×112×128 is generated, realizing the dimensionality reduction from 672 to 128. At the same time, through the weight learning of different convolution kernels, the correlation between shallow details, mid-level local features, and deep global features is strengthened, highlighting the cross-scale feature combination that is key to the identification of pests and diseases.

[0046] A global average pooling operation is performed on the optimized fusion features to condense global pest and disease feature information. The pooling calculation process is as follows: for each of the 128 channels of the optimized fusion features, first, all pixel values ​​of the feature map of that channel (112×112=12544 pixels) are counted, and all pixel values ​​are added together to obtain the pixel sum of that channel. Then, the pixel sum is divided by the total number of pixels (12544) to obtain the average pixel value of that channel. This value represents the global average intensity of the pest and disease feature corresponding to that channel. The average pixel values ​​of the 128 channels are arranged in channel order and combined to form a 128-dimensional one-dimensional global feature vector. This vector discards the spatial location information of the feature map and only retains the pest and disease feature intensity information at the global level, which reduces the complexity of subsequent calculations and avoids local redundant information from interfering with classification.

[0047] Step 502: Input the one-dimensional global feature vector into the fully connected classification layer of the pre-trained lightweight neural network. Perform a linear transformation through the fully connected classification layer to obtain the original classification score vector for each pest / disease category. Specifically, input the 128-dimensional one-dimensional global feature vector obtained in step 501 into the fully connected classification layer of the pre-trained lightweight neural network. This layer is pre-trained to adapt to the corn pest / disease identification task and is used to map the global features to scores for each category. The specific parameters and calculation process are as follows: The weight matrix of the fully connected classification layer has a dimension of K×128, where K is the number of corn pest / disease categories, such as corn leaf blight, corn rust, corn sheath blight, and corn stalk rot. Let K=5. Each element of the weight matrix represents the correlation strength between the corresponding global feature dimension and the corresponding category. The vector dimension is set to K, and each element corresponds to a score bias for a type of pest or disease. The bias value is controlled within the range of [-0.1, 0.1]. According to the matrix multiplication rule, the K×128 weight matrix of the fully connected classification layer is multiplied with the 128-dimensional global feature vector. Specifically, the 128 elements of the u-th row of the weight matrix are multiplied element by element with the 128 elements of the global feature vector. Then, all the product results are summed to obtain the intermediate score value of the u-th type of pest or disease. Subsequently, this intermediate score value is added to the u-th bias value of the bias vector to obtain the preliminary classification score of the u-th type of pest or disease. This process is repeated until all K categories have been calculated. The results are combined to form a K-dimensional original classification score vector. Each value in the vector corresponds to the preliminary classification score of a type of corn pest or disease. The higher the score, the higher the matching degree between the global features and the features of that category.

[0048] Step 503: Input the original classification score vector into the normalization function to calculate the final predicted probability of each pest / disease category. Select the category with the highest final predicted probability as the corn pest / disease identification result. Specifically, this includes: inputting the original classification score vector obtained in step 502 into the Softmax normalization function to convert the preliminary classification score into an intuitive category prediction probability. The specific calculation and result determination process is as follows: Let the original classification score vector be... ,in For the first The initial score for pests and diseases is first calculated by determining the natural index value for each score. (where e is the natural constant), then calculate the sum of all natural index values ​​= + +…+ Finally, use the first The natural index value of the class ( Divide by the sum to get the first... The final predicted probability of each disease and pest is calculated. The sum of the predicted probabilities of all categories is 1, and the value of each final predicted probability is in the range of [0, 1]. The larger the probability value, the higher the probability that the category is the real disease and pest category. The final predicted probabilities of all categories are iterated through, and the category with the largest probability value is obtained. This category is then identified as the disease and pest of corn.

[0049] This embodiment avoids information fragmentation caused by resolution differences between features at different scales by upsampling to unify spatial resolution and stitching together multi-scale features. It integrates shallow details, mid-level local features, and deep global features. Convolutional layers reduce feature redundancy by reducing dimensionality and strengthen the correlation between pest and disease features across scales. Global average pooling transforms two-dimensional features into one-dimensional vectors, preserving global feature information while reducing the computational cost of subsequent fully connected layers and improving feature processing efficiency. The fully connected classification layer maps global features to the category space through linear transformation, and the generated classification score accurately reflects the matching degree between features and each category. The Softmax function transforms the classification score into a probability, making the results more intuitive. The category with the highest probability is selected as the recognition result, ensuring the clarity of the recognition decision and improving the accuracy and reliability of corn pest and disease identification.

[0050] In a preferred embodiment of the present invention, step 6 includes: Step 600: Based on the identification results of corn diseases and pests, extract the disease and pest category identifiers and corresponding predicted probability values. Specifically, this includes: extracting two types of core information from the final identification results of corn diseases and pests generated in step 503. The first is the disease and pest category identifier, which is the predefined label of the disease and pest with the highest predicted probability value, including corn leaf blight, corn rust, corn sheath blight, healthy corn, etc. The label must be consistent with the category code pre-stored in the pest database to ensure subsequent data matching. The second is the corresponding predicted probability value, which is the final predicted probability corresponding to the category identifier, such as 0.92, 0.85, etc. After extraction, the predicted probability value needs to be validated to determine whether the probability value is within a reasonable range of 0 to 1. If it exceeds the range, such as a negative number or a value greater than 1, a recalculation instruction is triggered on the field IoT terminal. That is, the terminal calls the one-dimensional global feature vector stored in step 502, re-inputs the fully connected classification layer and the Softmax function, and generates a new predicted probability value until a valid probability value is extracted.

[0051] Step 601: Combining the image acquisition timestamp and the geographical location information of the acquisition device, generate structured pest monitoring data containing pest categories, predicted probability values, timestamps, and geographical locations. Specifically, the image acquisition timestamp comes from the terminal's built-in clock module, which synchronizes with the cloud time via the NTP protocol of the IoT platform with an error of ≤1 second. The geographical location information of the acquisition device comes from the GPS / BeiDou positioning module integrated into the terminal, with a positioning accuracy of ≤10 meters. It is also associated with field environmental parameters collected in real time by the terminal, such as temperature and humidity, and the data comes from the temperature and humidity sensors on the terminal. During structured integration, the data is organized according to a predefined field format. The fields include pest category identifier, predicted probability value, acquisition timestamp, geographical location of the acquisition device (latitude and longitude), acquisition device number, field temperature, and field humidity. Each field is encapsulated in JSON format to ensure a consistent data structure.

[0052] Step 602 involves transmitting the structured insect pest monitoring data to the cloud-based IoT platform via an IoT communication interface. This includes: For communication protocol and interface selection, low-power wide-area network protocols such as LoRa and NB-IoT are chosen based on the field network environment; alternatively, MQTT protocols are suitable for scenarios with larger data volumes. The communication interface is the corresponding protocol module built into the terminal, such as a LoRa RF module or an NB-IoT module, with data forwarding achieved through an IoT gateway. In the data encapsulation and verification stage, the structured data in JSON format is encapsulated before transmission. If the LoRa / NB-IoT protocol is used, the JSON data is converted into a binary stream, and the checksum is calculated using the CRC32 checksum algorithm. The checksum is the result of performing cyclic redundancy calculations on the binary stream data using a CRC32 polynomial, where x represents a variable in the polynomial formula. The polynomial formula is: + + + + + + + + + The checksum +x+1 is appended to the end of the binary stream to form a complete data packet. If the MQTT protocol is used, a message subject needs to be set in the format of crop type / pest / data / device number, where the device number is the unique identifier of the data collection device. The JSON data is then encapsulated as the message body. For reliable transmission, the terminal monitors the transmission status in real time. If a transmission failure occurs, such as no response from the gateway or failure to verify, the data will be retransmitted after 5 seconds, with a maximum of 3 retries. If a retry fails, the data is marked as pending transmission and will be transmitted first after the network is restored to ensure that no data is lost.

[0053] Step 603: The IoT platform receives and parses the structured insect pest monitoring data, and stores the parsed structured insect pest monitoring data in the insect pest database. Specifically, during data reception and verification, the platform receives data packets through corresponding protocol ports, such as the LoRa gateway port and MQTTBroker port. For LoRa / NB-IoT data packets, the CRC32 checksum at the end is extracted, and the CRC32 value of the data packet body is recalculated. If the two match, the verification is successful. For MQTT messages, the platform verifies whether the sending device has registered on the platform and whether the message subject format is compliant. If the verification is successful, the message body is extracted. In the data parsing stage, according to predefined structured data fields, pest category identifiers, predicted probability values, etc., are extracted from the JSON message body or binary stream. The system verifies the validity of fields, such as the predicted probability value (which must be between 0 and 1), the latitude and longitude (which must be within a reasonable geographical range for corn-producing areas), and the temperature (which must be within the crop growth environment range of -10℃ to 45℃). If a field is invalid, it is marked as data to be verified and the corresponding terminal is notified to re-upload. Valid data is then stored. During data storage, valid data is stored in the pest database, using the collection timestamp and collection device number as a composite primary key in the main pest monitoring table. Simultaneously, the collection device number is associated with the database device information table, which stores the device's fixed location and hardware model. The collection timestamp is associated with the environmental parameter supplementary table, which stores the temperature and humidity at the corresponding time, forming a main table-supplementary table relationship structure. This ensures that the data can be traced back to complete scene information through the device number and timestamp.

[0054] Step 604: Based on historical records and current data in the pest database, generate a pest early warning report or pest management decision-making suggestions, and display and remotely manage them through the visual interface of the IoT platform. Specifically, this includes: When generating a pest early warning report, first determine the statistical dimensions, namely the time dimension such as a single day or a single week, and the spatial dimension based on the geographical location of the collection device, dividing the area into plots and townships; calculate the pest incidence rate, the formula is: Pest incidence rate = (Number of positive collections of a certain pest within the statistical period ÷ Total number of collections in the area within the statistical period) × 100%, where the number of positive collections refers to the number of collections with a predicted probability value ≥ 0.7 (a positive threshold predefined based on historical control effects) and the category is identified as that pest; compare the current incidence rate with the average incidence rate of the same period in the past 3 years in the database. If the current incidence rate is higher than 1.5 times the average incidence rate (early warning threshold), then generate a pest early warning report, the content of which includes the pest category, the affected area (accurate to the plot), the current incidence rate, historical data for the same period, and the expected spread direction (combined with wind direction and plot layout).

[0055] In the pest and disease management decision-making recommendation generation stage, the platform calls upon a database of historical control effect tables. This table stores the control measures and effects of a certain pest or disease under similar environmental parameters. Combining current environmental parameters, such as prioritizing ventilation and chemical control in high humidity environments, and considering the incidence rate, the platform generates a recommendation to use triadimefon pesticides, spraying once every 7 days for two consecutive times, if the incidence rate of corn rust is 18%, which is higher than the historical average of 9%, and the current humidity is 75%. At the same time, the platform recommends turning on field ventilation equipment to reduce humidity to below 60%. The recommendation also references historical data, such as the control effect of this measure reaching 85% in similar scenarios in the same period of 2023. In terms of remote visualization management, the IoT platform's visualization interface (web dashboard, mobile APP) displays data in chart form, such as a regional pest and disease distribution heat map (the darker the color, the higher the incidence rate) and an incidence rate trend line graph (compared to historical data). It also displays early warning reports and decision-making recommendations. Managers can view real-time data, export PDF reports, or issue remote commands through the interface, such as sending a control command to field irrigation / ventilation equipment to adjust the humidity to 60%, thus realizing remote monitoring and management of pests.

[0056] This embodiment extracts core information and associates it with timestamps, geographical locations, and environmental parameters to form structured data. It then uses database association and storage to achieve full-chain traceability of equipment, time, pest infestation, and environment. Relying on IoT communication protocols and a verification and retry mechanism, it ensures rapid transmission of field data to the cloud, with automatic retries in case of transmission failure, preventing the spread of pests and diseases from going undetected due to data lag and allowing time for early intervention. By quantitatively calculating the incidence rate and comparing it with historical data to determine early warning thresholds, it avoids empirical errors. Simultaneously, it combines historical control effects with real-time environmental parameters to generate decision-making suggestions, ensuring highly targeted measures, reducing pesticide waste and environmental impact from indiscriminate control, and improving control efficiency. A visual interface displays pest infestations in multiple regions in real time, allowing managers to monitor the situation without being on-site and remotely issue control commands, thus improving management efficiency.

[0057] Embodiments of the present invention also provide an IoT-based remote management platform for insect pest monitoring, comprising: a processor and a memory storing a computer program, wherein the computer program, when run by the processor, executes the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0058] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0059] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A remote management method for insect pest monitoring and forecasting based on the Internet of Things, characterized in that, The method includes: Step 1: Preprocess the acquired original crop images to obtain standardized images, and input the standardized images into a pre-trained lightweight convolutional neural network for feature extraction to obtain multi-scale feature maps; Step 2: Treat the feature vector of each spatial location in the multi-scale feature map as a feature point to obtain a set of feature points. Project the set of feature points onto a virtual feature plane to construct a regular geometric structure that covers the distribution of all feature points, and divide the regular geometric structure into multiple sub-regions. Step 3: Map each feature point to the corresponding sub-region according to its projected coordinates, and aggregate the information of all feature points in each sub-region to generate a set of region aggregated features; Step 4: Based on the regional aggregation features, a spatial structure enhancement weight vector is calculated. The multi-scale feature map is modulated by the spatial structure enhancement weight vector to obtain the multi-scale features after spatial structure enhancement. Step 5: Perform feature fusion and classification on the multi-scale features after spatial structure enhancement to obtain the identification results of maize diseases and pests; Step 6: Based on the identification results of corn diseases and pests, generate pest monitoring information and transmit and manage it remotely through the Internet of Things platform.

2. The method for remote management of insect pest monitoring and forecasting based on the Internet of Things according to claim 1, characterized in that, Step 1 includes: The original crop image is scaled to a preset fixed width and height to obtain a size-normalized image. The size-normalized image is then subjected to color channel standardization. The mean and standard deviation of pixel values ​​for all pixels in the red, green, and blue color channels of the size-normalized image are calculated. Based on the mean and standard deviation of pixel values, the pixel values ​​of each color channel are normalized to obtain a standardized image. A standardized image is input into a pre-trained lightweight convolutional neural network. The pre-trained lightweight convolutional neural network contains multiple concatenated convolutional blocks. The standardized image is sequentially passed through multiple concatenated convolutional blocks for hierarchical feature extraction. The output feature maps of three convolutional blocks at different depths are determined. The three convolutional blocks at different depths correspond to the shallow, medium, and deep convolutional blocks of the pre-trained lightweight convolutional neural network, respectively. The output feature maps of shallow convolutional blocks, mid-level convolutional blocks, and deep convolutional blocks are used together as a multi-scale feature map.

3. The method for remote management of insect pest monitoring and forecasting based on the Internet of Things according to claim 2, characterized in that, Step 2 includes: Each feature vector at a spatial location in a multi-scale feature map is defined as a feature point, and the set of all feature points constitutes the feature point set. Each feature point in the feature point set is linearly transformed through a fully connected layer in a pre-trained lightweight convolutional neural network, and the high-dimensional feature vector of the feature point is projected onto a two-dimensional coordinate system to obtain the two-dimensional projected coordinates of each feature point on the virtual feature plane. Based on the two-dimensional projected coordinates of all feature points in the feature point set, calculate the minimum and maximum values ​​of the two-dimensional projected coordinates in the horizontal and vertical directions, determine the minimum bounding rectangle that can cover the projected coordinates of all feature points, and use the minimum bounding rectangle as a regular geometric structure. Divide the smallest bounding rectangle into equal-sized rectangular grids along its length and width, with each rectangular grid being a sub-region.

4. The method for remote management of insect pest monitoring and forecasting based on the Internet of Things according to claim 3, characterized in that, Step 3 includes: Based on the two-dimensional projected coordinates of each feature point and the boundary coordinates of multiple rectangular grid sub-regions, the corresponding rectangular grid to which each feature point belongs is determined, and the high-dimensional feature vectors of all feature points belonging to the same rectangular grid are grouped and aggregated. The high-dimensional feature vectors within each rectangular grid group are subjected to mean pooling and max pooling to obtain the mean pooling feature vector and max pooling feature vector corresponding to each rectangular grid. The average pooling feature vector and the max pooling feature vector of each rectangular grid are concatenated along the channel dimension to form the region aggregated feature vector of each rectangular grid. All rectangular grid region aggregation feature vectors are organized and arranged according to the row and column order of the grid, and combined into a set of region aggregation features.

5. The method for remote management of insect pest monitoring and forecasting based on the Internet of Things according to claim 4, characterized in that, Step 4 includes: Based on the region aggregation characteristics, feature correction vectors are generated through geometric relationship analysis; The feature correction vector is fused with the region aggregation feature to obtain the geometrically enhanced region aggregation feature; The geometrically enhanced region aggregation features are input into the weight generation network in a pre-trained lightweight neural network to obtain the initial weight vector; The initial weight vector is processed by a normalization function to obtain a numerically normalized spatial structure enhancement weight vector. The channel dimension of the spatial structure enhancement weight vector is expanded to match the number of channels in the multi-scale feature map to obtain the expanded spatial structure enhancement weight vector. The expanded spatial structure enhancement weight vector is then multiplied channel-by-channel with the multi-scale feature map to obtain the spatial structure enhancement multi-scale feature.

6. The method for remote management of insect pest monitoring and forecasting based on the Internet of Things according to claim 5, characterized in that, Based on region aggregation features, feature correction vectors are generated through geometric relationship analysis, including: Calculate the arithmetic mean of the x-coordinates and the arithmetic mean of the y-coordinates of all two-dimensional projected coordinates in the feature point set, and take the points corresponding to the arithmetic mean of the x-coordinates and the arithmetic mean of the y-coordinates as the centroids of the feature distribution; Establish a temporary coordinate system with the centroid of the characteristic distribution as the origin, and calculate the covariance matrix of all two-dimensional projected coordinates in the temporary coordinate system. The covariance matrix is ​​decomposed into eigenvalues. The eigenvector corresponding to the largest eigenvalue is extracted as the first principal direction, and the eigenvector corresponding to the second largest eigenvalue is extracted as the second principal direction. The line containing the first principal direction is taken as the first reference axis, and the line containing the second principal direction is taken as the second reference axis. Using the centroid of the feature distribution as the vertex and the first and second reference axes as the boundaries, the feature projection plane is divided into four quadrants. The number of feature points contained in each quadrant is counted, and the feature point density of each quadrant is calculated. The average Euclidean distance from all feature points in each quadrant to the centroid of the feature distribution is calculated. The feature point density and average Euclidean distance of the four quadrant regions are concatenated in quadrant order to form an eight-dimensional spatial distribution feature vector. The eight-dimensional spatial distribution feature vector is then subjected to linear transformation and dimensional compression to generate a feature correction vector.

7. The method for remote management of insect pest monitoring and forecasting based on the Internet of Things according to claim 6, characterized in that, Step 5 includes: After enhancing the spatial structure, the shallow, medium, and deep feature maps in the multi-scale features are spliced ​​together along the channel dimension to form preliminary fused features. The initial fused features are input into the convolutional layers of a pre-trained lightweight neural network for channel dimensionality reduction and feature integration to obtain optimized fused features with a uniform number of channels; global average pooling is then performed on the optimized fused features to obtain a one-dimensional global feature vector. The one-dimensional global feature vector is input into the fully connected classification layer in the pre-trained lightweight neural network. The fully connected classification layer performs a linear transformation to obtain the original classification score vector for each pest category. The original classification score vector is input into the normalization function to calculate the final predicted probability of each disease and pest category. The category with the highest final predicted probability value is selected as the corn disease and pest identification result.

8. The method for remote management of insect pest monitoring and forecasting based on the Internet of Things according to claim 7, characterized in that, Step 6 includes: Based on the identification results of corn diseases and pests, disease and pest category identifiers and corresponding predicted probability values ​​are extracted; By combining image acquisition timestamps and the geographic location information of the acquisition device, structured pest monitoring data containing pest categories, predicted probability values, timestamps, and geographic locations is generated. Structured insect monitoring data is transmitted to the cloud-based IoT platform via an IoT communication interface; The IoT platform receives and parses structured insect pest monitoring data, and stores the parsed structured insect pest monitoring data in the insect pest database; Based on historical records and current data in the insect infestation database, insect infestation early warning reports or pest management decision-making suggestions are generated and displayed and remotely managed through the visual interface of the Internet of Things platform.

9. A remote management platform for insect pest monitoring and forecasting based on the Internet of Things, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images