Pest time sequence prediction method based on layered convolution
The pest timing data is processed through the stratified convolution model UniTCN and specific loss function strategies, and the nonlinear and non-stable state problems of pest prediction in the prior art are solved, and efficient and accurate pest warning is achieved.
Patent Information
- Application Number
- CN202510297452.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-13
- Publication Date
- 2025-07-04
AI Technical Summary
Existing pest prediction technologies are difficult to deal with high-dimensional nonlinear relationships, have limited long-term periodic law modeling capabilities, and lack generalization capabilities in non-steady-state scenarios, resulting in early warning lag or misjudgment.
The hierarchical convolution model UniTCN is adopted, which includes the RevIN layer, the timing decomposition module and the hierarchical convolution attention module. It combines the Huber loss function and OneCycleLR learning rate strategy to process multivariable timing data, suppress meteorological mutation interference, and explicitly learn dynamic associations of multi-source variables.
Lightweight and low-latency pest timing prediction is realized, which improves the robustness and prediction stability of the model in non-steady-state scenarios, reduces memory footprint and inference delay, and improves prediction accuracy and generalization capabilities.
Smart Images
Figure CN120258205A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of pest prediction, and particularly relates to a pest time series prediction method based on hierarchical convolution. Background Art
[0002] With the growth of the global population and the development of agricultural intensification, ensuring the yield and quality of crops has become the core issue of food security. However, pest infestation poses a serious threat to agricultural production. According to statistics, the global annual crop losses caused by pests reach as high as 15%-20%, accompanied by the ecological imbalance and the risk of drug resistance caused by the abuse of pesticides. Taking the codling moth and the tomato leaf miner as examples, they cause billions of dollars in economic losses every year: the codling moth causes losses of more than 700 million US dollars to fruit farmers in the United States, while the tomato leaf miner once caused an epidemic in Guangdong, resulting in a halving of tomato production and direct losses of more than 2 billion yuan. Traditional pest monitoring relies on manual inspections or simple statistical models, which have defects such as poor real-time performance, insufficient long-term dependence modeling, and weak adaptability to non-steady data, and it is difficult to cope with complex scenarios such as meteorological mutations and seasonal pest outbreaks, resulting in late warnings or misjudgments.
[0003] Existing pest prediction technologies mainly rely on two types of methods: traditional statistical models and machine learning models. However, both of these two types of methods require manual feature engineering, have high requirements for domain knowledge, and the quality of feature selection and extraction directly affects the model effect. In addition, traditional statistical models are limited by linear assumptions, and machine learning models are limited by model structures and parameters. Both are difficult to handle high-dimensional non-linear relationships and have limited ability to model long-term periodic patterns.
[0004] In recent years, deep learning technologies (such as LSTM, Transformer) have provided new ideas for time series prediction, but still face core challenges in the pest warning scenario. The traditional time convolutional network (TCN) is limited by locality and is difficult to model the cross-seasonal periodic patterns in pest data. And Transformer relies on a high-complexity attention mechanism and is difficult to be deployed on resource-constrained field edge devices. In addition, pest data is frequently shifted in distribution due to external factors such as meteorological mutations and pesticide applications, which further poses high requirements for the generalization ability of the model in non-steady scenarios. Therefore, it is particularly important to establish a lightweight pest time series prediction method with noise suppression and strong generalization ability. Summary of the Invention
[0005] In order to make up for the deficiencies of the existing technology, the present invention aims to provide a pest time series prediction method based on hierarchical convolution to solve the problems encountered in pest prediction in the existing technology. The technical problems solved by the present invention can be realized through the following technical solutions:
[0006] The described pest time series prediction method based on hierarchical convolution includes the following specific steps:
[0007] S1. Collect multivariate time series data, including the occurrence quantity of pests and environmental parameters such as temperature, humidity, light, soil temperature, and soil humidity.
[0008] S2. Preprocess the data, including filling missing values, cleaning outliers, and normalizing.
[0009] S3. Divide the preprocessed dataset into three subsets: a training set, a validation set, and a test set.
[0010] S4. Construct a hierarchical convolutional model UniTCN, which includes a RevIN layer, a time series decomposition module, and a hierarchical convolutional attention module.
[0011] S5. Use the Huber loss function and the OneCycleLR learning rate strategy for model evaluation and update. If the preset number of iterations is reached, end the training.
[0012] S6. Deploy the hierarchical convolutional model UniTCN to achieve real-time prediction and early warning of the occurrence quantity of pests.
[0013] Further, in step S1, data on the occurrence quantity of Tuta absoluta in Hotan, Changji, Yili and other places in Xinjiang are collected by deploying environmental monitoring devices (temperature and humidity sensors, soil probes, light meters), and environmental parameters are recorded synchronously with a time resolution of 1 hour, covering 1 year; samples are generated using a sliding window (window length = 24 hours, step size = 1 hour).
[0014] Further, in step S2, the missing data are filled with the sliding window mean of adjacent time series data, and outliers are identified using a statistical method based on the 3σ principle and replaced with locally weighted smoothed values; finally, the obtained multivariate time series data are respectively normalized by Z-score to eliminate the dimensional differences between different features and make the data comparable. The formula is:
[0015]
[0016] where x is a single sample, μ is the mean of all samples, and σ is the standard deviation of all samples.
[0017] Further, in step S3, the processed dataset is divided into three subsets: a training set, a validation set, and a test set, in a ratio of 8:1:1.
[0018] Furthermore, in step S4, the RevIN layer of the hierarchical convolutional model UniTCN normalizes and denormalizes each individual sample to address the distribution shift problem; the time series decomposition module uses the moving average method to decompose the time series into seasonal and trend terms, suppressing the interference of non-stationary noises such as meteorological mutations on the prediction results. The hierarchical convolutional attention module includes a point-level attention sub-module, a block-level attention sub-module, and a variable-level attention sub-module. The point-level attention sub-module extracts local time series features based on small kernel convolution and parameter sharing mechanisms, avoiding the dependence on position encoding; the block-level attention sub-module models cross-block time series patterns through multi-scale depth convolution to capture long-term dependencies; the variable-level attention sub-module explicitly learns the dynamic associations between multi-source variables (such as the non-linear coupling between temperature, humidity, and pest occurrence), and fuses cross-variable information through grouped convolution.
[0019] Furthermore, the specific construction of the hierarchical convolutional model UniTCN is as follows:
[0020] ① Let the historical time step be T and the input variable dimension be M, then the input matrix is expressed as:
[0021] X = [x1, x2, …, x T ∈ R T*M ,
[0022] where, x t ∈ R M represents the multi-variable observation value at the t-th time step;
[0023] ② Predict the pest occurrence in the next S time steps, and the output matrix is:
[0024] Y = [y T+1 , y T+2 , …, y T+S ∈ R T*1 ,
[0025] where, y T+s represents the predicted value of the pest occurrence at the s-th time step in the future;
[0026] ③ The input sequence is first normalized and denormalized for each individual sample by the RevIN layer to address the distribution shift problem, and its formula is:
[0027]
[0028] x = x norm * σ norm + μ norm ,
[0029] where, x norm is the normalized individual sample, μ norm is the mean of the individual sample, σnorm is the standard deviation of a single sample;
[0030] ④ The time series decomposition module uses the moving average method to decompose the input sequence into a seasonal term and a trend term:
[0031] T t = MA(X t ; kernel size = K),
[0032] S t = X t - T t ,
[0033] where K is the moving average window size, X t ∈R T*M is the input time series data, the decomposed seasonal term S t is used to model periodic changes, and the trend term T t is used to capture long-term trends;
[0034] ⑤ The hierarchical convolutional model UniTCN extracts multi-scale time series features through the hierarchical convolutional module.
[0035] Furthermore, the specific content of the hierarchical convolutional model UniTCN extracting multi-scale time series features through the hierarchical convolutional module is as follows:
[0036] First, in the point-level attention sub-module, local general time series features are extracted based on the parameter sharing mechanism using a small convolutional kernel, that is:
[0037]
[0038] where, represents the output of the block-level attention sub-module in the previous layer, P represents the stacking layer number of the point-level attention sub-module and the block-level attention sub-module, PtWB(·) represents the point-level attention sub-module, Conv(·) represents the convolutional operation, and GELU(·) represents the activation function based on the Gaussian distribution;
[0039] Then, in the block-level attention sub-module, depthwise separable convolution is combined with multi-scale convolution to model the cross-block long-term dependence relationship and fuse the time series patterns under different time spans, that is:
[0040]
[0041]
[0042] where, represents the output of the previous - layer dot - level attention sub - module, PhWB(·) represents the block - level attention sub - module, Conv1(·) and Conv2(·) represent two multi - scale convolution operations based on depth - separable convolution, and Drop(·) represents a dropout operation;
[0043] Finally, the variable - level attention sub - module uses grouped convolution to learn the multi - variable dynamic association, explicitly modeling the non - linear coupling relationship between environmental parameters and pest occurrence. The formula is as follows:
[0044] O l = VWB(O l-1 ), l = 1, …, Q,
[0045] VWB(O l-1 ) = Drop(GELU(Conv(O l-1 ))),
[0046] where O l-1 represents the output of the previous - layer variable - level attention sub - module, Q represents the stacking layer number of the variable - level attention sub - module, and VWB(·) represents the variable - level attention sub - module.
[0047] Furthermore, the Huber loss function in step S5 is a loss function that combines MSE and MAE. When dealing with regression problems, the Huber loss function can both smooth the training process and reduce the influence of outliers. Its formula is:
[0048]
[0049] where y is the true pest occurrence, is the predicted pest occurrence, and δ is the parameter that controls the contributions of MSE and MAE.
[0050] Furthermore, OneCycleLR in step S5 is different from the traditional learning rate adjustment strategy. During training, the learning rate first linearly increases from a small value to a large value, and then linearly decreases to the set final value, forming a single - cycle learning rate change curve. The OneCycleLR strategy allows the use of a larger learning rate in the early stage of training to prompt the model to quickly jump out of the local optimal solution, accelerate the convergence speed, and reduce the training time required; at the same time, using a smaller learning rate in the later stage can help the model make fine - tuning in the region close to the optimal solution, which is helpful to improve the generalization ability of the model.
[0051] Compared with the prior art, the present invention has the following advantages:
[0052] (1) The present invention adopts a lightweight architecture, combines depthwise separable convolution and dynamic pruning to achieve low-latency inference. Compared with traditional Transformers, the memory occupancy is saved by about 50%, and the inference efficiency is improved by about 43%.
[0053] (2) The present invention is based on a sequence decomposition method of small kernel convolution, which separately processes the trend and seasonal components, can effectively improve the robustness of the model in non-steady-state scenarios, and enhance the prediction stability in extreme weather scenarios.
[0054] (3) The hierarchical convolutional attention module designed by the present invention can capture local temporal features and cross-local model temporal patterns while extracting local temporal features, capture long-term dependencies (such as the impact of weekly meteorological changes on the occurrence of pests), and can also explicitly learn the dynamic associations between multi-source variables. Brief Description of the Drawings
[0055] Figure 1 It is the flowchart of the method of the present invention. Detailed Embodiment
[0056] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0057] In this embodiment, taking the tomato leaf miner as an example, a pest temporal prediction method based on hierarchical convolution is provided.
[0058] As Figure 1 shown, a pest temporal prediction method based on hierarchical convolution, the specific steps include: collecting multivariate temporal data; filling missing values, cleaning outliers and normalizing the data for preprocessing; dividing the preprocessed data set; constructing a hierarchical convolution model UniTCN, including a RevIN layer, a temporal decomposition module and a hierarchical convolutional attention module; using the Huber loss function and the OneCycleLR learning rate strategy for model evaluation and update; deploying the UniTCN model to achieve real-time prediction and early warning. The details are as follows:
[0059] (1) Collect multivariate temporal data to construct a data set.
[0060] Environmental parameters of tomato planting areas in Hotan, Changji, and Yili in Xinjiang were collected by deploying temperature and humidity sensors, soil temperature and humidity probes, and light meters, with a time resolution of 1 hour, covering a one-year period from January 2024 to December 2024; the daily number of adult tomato leafminers was recorded synchronously. The original data contained 8,760 samples, and after processing with a sliding window (window length = 24 hours, step size = 1 hour), 8,736 time series samples were generated. Each sample contained 24-hour historical data and a 24-hour prediction target. The input dimension of each sample was 6 (including six variables: pest occurrence, temperature, humidity, light intensity, soil temperature, and soil humidity), and the output dimension was 1 (i.e., single-variable prediction of pest occurrence).
[0061] (2) Perform preprocessing on the data, including filling missing values, cleaning outliers, and standardizing.
[0062] For missing values caused by sensor failures, use the mean value of the sliding window (window size = 6 hours) for filling. Identify outliers based on the 3σ principle and replace them with locally weighted smoothed values. For example, when the pest occurrence at a certain moment exceeds the range of the historical mean ± 3 times the standard deviation, use the weighted smoothed value of the adjacent 12-hour data for correction. Perform Z-score standardization on multi-variable data respectively to eliminate the dimensional differences between different features and make the data comparable. The formula is:
[0063]
[0064] where x is a single sample, μ is the mean of all samples, and σ is the standard deviation of all samples.
[0065] (3) Divide the preprocessed dataset into three subsets: training set, validation set, and test set, in a ratio of 8:1:1, to avoid data leakage.
[0066] (4) Construct a hierarchical convolutional model, the UniTCN model.
[0067] The hierarchical convolutional model UniTCN includes the RevIN layer, the time series decomposition module, and the hierarchical convolutional attention module. The RevIN layer performs normalization and inverse normalization on a single sample to solve the problem of distribution shift. The time series decomposition module uses the moving average method to decompose the time series into seasonal and trend terms, suppressing the interference of non-stationary noises such as meteorological mutations on the prediction results. The hierarchical convolutional attention module includes the point-level attention sub-module, the block-level attention sub-module, and the variable-level attention sub-module. The point-level attention sub-module extracts local time series features based on small kernel convolution and parameter sharing mechanisms, avoiding the dependence on position encoding. The block-level attention sub-module models cross-block time series patterns through multi-scale depth convolution to capture long-term dependencies. The variable-level attention sub-module explicitly learns the dynamic associations between multi-source variables (such as the non-linear coupling between temperature, humidity, and pest occurrence), and fuses cross-variable information through grouped convolution.
[0068] Let the historical time step be \(T\), and the dimension of the input variable be \(M\) (including the occurrence amount of pests and environmental parameters). Then the input matrix is expressed as:
[0069] \(X = [x_1, x_2, \ldots, x T \in\mathbb{R} T*M ,
[0070] where \(x t \in\mathbb{R} M represents the multi-variable observation value at the \(t\)-th time step.
[0071] Predict the occurrence amount of pests in the next \(S\) time steps. The output matrix is:
[0072] \(Y = [y T+1 , y T+2 , \ldots, y T+S \in\mathbb{R} T*1 ,
[0073] where \(y T+s represents the predicted value of the occurrence amount of pests at the \(s\)-th time step in the future.
[0074] The input sequence is first normalized and denormalized for a single sample via the RevIN layer to solve the distribution shift problem. The formula is:
[0075]
[0076] x = x norm *\sigma norm +\mu norm ,
[0077] where \(x norm is the single sample after normalization, \(\mu norm is the mean of the single sample, and \(\sigma norm is the standard deviation of the single sample.
[0078] The time series decomposition module uses the moving average method to decompose the input sequence into a seasonal term and a trend term:
[0079] T t = MA(X t ; kernel size = K),
[0080] S t = X t - T t ,
[0081] where \(K\) is the moving average window size, \(X t \in\mathbb{R} T*M is the input time series data, and the decomposed seasonal term \(S tFor modeling periodic variations, trend term T t For capturing long-term trends.
[0082] The hierarchical convolutional model, the UniTCN model, extracts multi-scale time series features through hierarchical convolutional modules. First, in the point-level attention sub-module, local general time series features are extracted based on a small convolutional kernel through a parameter sharing mechanism, avoiding dependence on positional encoding.
[0083]
[0084] Among them, represents the output of the block-level attention sub-module in the previous layer, P represents the stacking layers of the point-level attention sub-module and the block-level attention sub-module, PtWB(·) represents the point-level attention sub-module, Conv(·) represents the convolutional operation, and GELU(·) represents the activation function based on the Gaussian distribution.
[0085] Next, in the block-level attention sub-module, depthwise separable convolution is used in combination with multi-scale convolution to model cross-block long-term dependencies and fuse time series patterns at different time spans.
[0086]
[0087] Among them, represents the output of the point-level attention sub-module in the previous layer, PtWB(·) represents the block-level attention sub-module, Conv1(·), Conv2(·) represent two multi-scale convolution operations based on depthwise separable convolution, and Drop(·) represents one dropout operation.
[0088] Finally, the variable-level attention sub-module uses grouped convolution to learn the dynamic associations of multiple variables and explicitly models the non-linear coupling relationship between environmental parameters and the occurrence amount of pests (such as the interaction effect between temperature and pest damage).
[0089] O l = VWB(O l-1 ), l = 1, …, Q,
[0090] VWB(O l-1 ) = Drop(GELU(Conv(O l-1 ))),
[0091] Among them, O l-1 represents the output of the variable-level attention sub-module in the previous layer, Q represents the stacking layers of the variable-level attention sub-module, and VWB(·) represents the variable-level attention sub-module.
[0092] (5) The Huber loss function and the OneCycleLR learning rate strategy are used for model evaluation and update. If the preset number of iterations is reached, the training ends.
[0093] The Huber loss is adopted, combining the advantages of MAE and MSE, while smoothing the training process and reducing the influence of outliers. Its formula is as follows:
[0094]
[0095] where y is the actual occurrence amount of pests, is the predicted occurrence amount of pests, and δ is the parameter that controls the contributions of MSE and MAE.
[0096] In addition, combined with the OneCycleLR learning rate adjustment strategy, the learning rate is first increased from a small value to a large value during the training process, and then decreased to the set final value, forming a single-cycle learning rate change curve. The OneCycleLR strategy allows the use of a larger learning rate in the early stage of training to prompt the model to quickly jump out of the local optimal solution, accelerate the convergence speed, and reduce the training time required. At the same time, using a smaller learning rate in the later stage can help the model make fine adjustments in the region close to the optimal solution, which is helpful to improve the generalization ability of the model.
[0097] (6) Deploy the UniTCN model to achieve real-time prediction and early warning.
[0098] This example is placed on a dataset with the tomato leaf miner as the prediction object, the time window is 72h, and the verification metrics are selected as MAE, MSE, Memory, and Latency. When used as a pest prediction task in this embodiment, MAE represents the absolute error between the predicted value and the true value. The lower the MAE, the more accurate the prediction of the pest occurrence amount. MSE represents the squared error between the predicted value and the true value. The lower the MSE value, the more sensitive the prediction of the outliers of the pest occurrence amount. Memory represents the memory occupancy during training, reflecting the model deployment cost; the lower the Memory, the smaller the memory occupied by the model, the fewer the number of parameters, and the lower the computational complexity. Latency represents the single inference latency, and the average value is taken through multiple runs, reflecting the real-time performance of the model; the lower the Latency, the faster the model inference speed, the better the model performance, and it is more in line with the application requirements in the actual scenario.
[0099] In this embodiment, the existing LSTM is selected as the comparison model, and the test results of the two in the scenario reconstruction task are shown in the following table:
[0100] Table 1: Test results on the tomato leaf miner occurrence amount dataset
[0101]
[0102] The above experiments show that the method of the present invention has better pest prediction performance compared with the existing general methods, better robustness to abnormal data in the pest occurrence amount dataset, and better performance and efficiency in terms of memory occupation and inference efficiency. It can well meet the pest infestation prevention and control requirements in agricultural production scenarios and has good application value.
[0103] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A pest time series prediction method based on hierarchical convolution, characterized in that, The pest time series prediction method includes the following specific steps: S1. Collect multivariate time series data, including the occurrence quantity of pests and environmental parameters; S2. Preprocess the data, including filling missing values, cleaning outliers, and standardizing; S3. Divide the preprocessed data set into three subsets: a training set, a validation set, and a test set; S4. Construct a hierarchical convolutional model UniTCN, which includes a RevIN layer, a time series decomposition module, and a hierarchical convolutional attention module; S5. Use the Huber loss function and the OneCycleLR learning rate strategy for model evaluation and update. If the preset number of iterations is reached, end the training; S6. Deploy the hierarchical convolutional model UniTCN to achieve real-time prediction and early warning of the occurrence quantity of pests.
2. The pest time series prediction method based on hierarchical convolution according to claim 1, wherein In step S1, the data on the occurrence quantity of pests is collected by deploying environmental monitoring devices, and the environmental parameters are recorded synchronously. The time resolution is 1 hour and it covers 1 year. Samples are generated using a sliding window, where the window length = 24 hours and the step size = 1 hour.
3. A pest time series prediction method based on hierarchical convolution according to claim 1, characterized in that In step S2, the missing data is filled with the sliding window mean of adjacent time series data. For outliers, a statistical method based on the 3σ principle is used for identification and replaced with locally weighted smoothed values. The obtained multivariate time series data is respectively standardized using Z-score to eliminate the dimensional difference between different features and make the data comparable. The formula is: where x is a single sample, μ is the mean of all samples, and σ is the standard deviation of all samples.
4. A pest time series prediction method based on hierarchical convolution according to claim 1, characterized in that, In step S3, the processed data set is divided into three subsets: a training set, a validation set, and a test set, in the ratio of 8:1:
1.
5. A pest time series prediction method based on hierarchical convolution according to claim 1, characterized in that, In step S4, the RevIN layer of the hierarchical convolutional model UniTCN normalizes and denormalizes for a single sample; The time series decomposition module uses the moving average method to decompose the time series into a seasonal term and a trend term to suppress the interference of non-stationary noise on the prediction result. The hierarchical convolutional attention module includes a point-level attention sub-module, a block-level attention sub-module, and a variable-level attention sub-module. The point-level attention sub-module extracts local time series features based on small kernel convolution and parameter sharing mechanism; The block-level attention sub-module models cross-block time series patterns through multi-scale depth convolution to capture long-term dependencies. The variable-level attention sub-module explicitly learns the dynamic association between multi-source variables and fuses cross-variable information through grouped convolution.
6. The pest time series prediction method based on hierarchical convolution according to claim 5, wherein, The specific content of constructing the hierarchical convolutional model UniTCN is as follows: ① Let the historical time step be T and the input variable dimension be M, then the input matrix is expressed as: X = [x1, x2, …, x T ∈ R T*M , where x t ∈R M represents the multivariate observation at the t-th time step; ② Predict the occurrence quantity of pests in the next S time steps, and the output matrix is: Y = [y T+1 , y T+2 , …, y T+S} ∈ R T*1 , Among them, y T+s represents the predicted value of the pest occurrence amount at the s-th future time step; ③ The input sequence first undergoes normalization and denormalization for a single sample via the RevIN layer to solve the distribution shift problem. The formula is: x = x norm *σ norm +μ norm , where x norm is a single normalized sample, μ norm is the mean of a single sample, and σ norm is the standard deviation of a single sample; ④ The time series decomposition module uses the moving average method to decompose the input sequence into a seasonal term and a trend term: T t = MA(X t ; kernel size = K), S t = X t - T t , where K is the moving average window size, and X t ∈R T*M is the input time series data. The decomposed seasonal component S t is used to model periodic variations, and the trend component T t is used to capture long-term trends; ⑤ The hierarchical convolutional model UniTCN extracts multi-scale time series features through a hierarchical convolutional module.
7. A pest time series prediction method based on hierarchical convolution according to claim 6, characterized in that, The specific content of the hierarchical convolutional model UniTCN extracting multi-scale time series features through a hierarchical convolutional module is as follows: First, in the point-level attention sub-module, local general temporal features are extracted based on the parameter sharing mechanism using small convolutional kernels, that is: Among them, represents the output of the block-level attention sub-module in the previous layer, P represents the stacking layer number of the point-level attention sub-module and the block-level attention sub-module, PtWB(·) represents the point-level attention sub-module, Conv(·) represents the convolution operation, and GELU(·) represents the activation function based on the Gaussian distribution; Then, in the block-level attention sub-module, depthwise separable convolutions are combined with multi-scale convolutions to model cross-block long-term dependencies and fuse temporal patterns at different time spans, that is: Among them, represents the output of the previous layer of point-level attention sub-module, PhWB(·) represents the block-level attention sub-module, Conv1(·) and Conv2(·) represent two multi-scale convolution operations based on depthwise separable convolution, and Drop(·) represents a dropout operation; Finally, the variable-level attention sub-module uses grouped convolutions to learn the dynamic associations of multiple variables and explicitly model the non-linear coupling relationship between environmental parameters and pest occurrence amounts. The formula is as follows: O l = VWB(O l-1 ), l = 1, …, Q, VWB(O l-1 ) = Drop(GELU(Conv(O l-1 ))) Among them, O l-1 represents the output of the previous-layer variable-level attention sub-module, Q represents the stacking layer number of the variable-level attention sub-module, and VWB(·) represents the variable-level attention sub-module.
8. A pest time series prediction method based on hierarchical convolution according to claim 1, characterized in that The Huber loss function in step S5 is a loss function that combines MSE and MAE. When dealing with regression problems, the Huber loss function can both smooth the training process and reduce the influence of outliers. Its formula is: where y is the actual occurrence amount of pests, is the predicted occurrence amount of pests, and δ is the parameter that controls the contributions of MSE and MAE.
9. A method for identifying images of invasive pest sticky boards based on deep learning according to claim 8, characterized in that, In step S5, OneCycleLR first linearly increases the learning rate from a small value to a large value during the training process, and then linearly decreases it to the set final value, forming a single-cycle learning rate change curve.