Agricultural meteorological forecasting method based on multi-scale convolutional network and diffusion system

Abstract

The present invention is applicable to the field of agricultural meteorological prediction, and is specifically an agricultural meteorological prediction method based on a multi-scale convolutional network and a diffusion system, including: using a trained multi-scale convolutional network to make a preliminary prediction of an agricultural meteorological image to obtain a preliminary prediction result; processing the preliminary prediction result through a diffusion model, and processing the preliminary prediction result includes: firstly increasing noise through a diffusion model to capture the connection between spatiotemporal features, and then gradually eliminating noise in the prediction result through a multi-step inverse diffusion process; training a prediction model based on a multi-scale convolutional network and a diffusion model to obtain a trained prediction model; for a given agricultural meteorological image, using the prediction model to generate and output a prediction image sequence of multiple time steps in the future. The present invention significantly improves the performance of spatiotemporal data for long-term prediction results by combining a spatiotemporal convolutional neural network with a diffusion system.

Description

Technical Field

[0001] The present invention belongs to the technical field of agricultural meteorological forecasting, and in particular relates to an agricultural meteorological forecasting method based on a multi-scale convolutional network and a diffusion system. Background Art

[0002] Spatiotemporal data mining is an important branch of data mining, which aims to extract valuable information and patterns from data with time and space dimensions. Spatiotemporal data are widely used in fields such as agricultural production, weather forecasting and environmental monitoring. With the development of remote sensing technology, the Internet of Things and sensor technology, the sources of agricultural meteorological data have become more abundant, including satellite images, meteorological data, geographic location information, etc.

[0003] Different from traditional data, agricultural meteorological data has complex spatiotemporal structure and dependencies, which not only include dynamic changes in time series, but also involve correlations between spatial locations. Therefore, spatiotemporal data mining in the field of agricultural meteorology faces the challenge of how to capture both temporal and spatial dependencies at the same time, which is specifically reflected in:

[0004] First, the complex dependencies of spatiotemporal data often lead to inaccurate model prediction capabilities when dealing with forecasting tasks. This inaccuracy is particularly evident when the spatiotemporal dependencies cannot be fully captured, which manifests itself as smoothing in images or forecast results, resulting in the blurring or neglect of key spatiotemporal information. Specifically, the model may tend to ignore certain subtle dynamic changes in the time dimension, or fail to capture the complex interactions between regions in the spatial dimension, ultimately leading to inaccurate predictions of the development trend of meteorological forecasts;

[0005] Second, agricultural meteorological data not only involve dynamic changes in time series, but also include spatial interdependencies, such as the impact of factors such as topography, soil type and climate zone on meteorological conditions. How to effectively capture such complex spatiotemporal dependencies remains a difficult problem, and existing models are difficult to fully and accurately reflect these dependencies. Summary of the invention

[0006] The purpose of an embodiment of the present invention is to provide an agricultural meteorological forecasting method based on a multi-scale convolutional network and a diffusion system. By combining a multi-scale convolution module and a diffusion module, it is possible to extract multi-level feature information from spatiotemporal data, and optimize the feature representation through a diffusion process, thereby enhancing the robustness and accuracy of long-term predictions. Through the collaborative work of an encoder, a time evolution module, and a decoder, spatiotemporal features are extracted and converted, and feature representations containing complex spatiotemporal dynamics are mutually transmitted through a diffusion process, thereby effectively improving the accuracy of the model in long-term predictions.

[0007] To achieve the above object, the present invention provides the following technical solutions.

[0008] An embodiment of the present invention provides an agricultural meteorological forecasting method based on a multi-scale convolutional network and a diffusion system, and the agricultural meteorological forecasting method includes:

[0009] S101: Encoding the agricultural meteorological image input sequences from different time steps or spatial positions to obtain corresponding encoding matrices;

[0010] S102: construct a multi-scale convolutional network, use the encoding matrix to train the multi-scale convolutional network, extract the spatiotemporal features of the input sequence, use the trained multi-scale convolutional network to make a preliminary prediction of the agricultural meteorological image, and obtain a preliminary prediction result;

[0011] S103: Processing the preliminary prediction results through a diffusion model, wherein the processing of the preliminary prediction results includes: firstly, adding noise through a diffusion model to capture the connection between spatiotemporal features, and then gradually eliminating the noise in the prediction results through a multi-step inverse diffusion process;

[0012] S104: training a prediction model based on a multi-scale convolutional network and a diffusion model. During the training process, the mean square error between the predicted image and the true image is minimized by iteratively updating the model parameters until the loss function converges, thereby obtaining a trained prediction model.

[0013] S105: For a given agricultural meteorological image, a prediction model is used to generate and output a prediction image sequence for multiple future time steps, wherein the prediction image sequence is used to reflect the spatiotemporal changes of meteorological conditions.

[0014] Furthermore, the step of encoding the agricultural meteorological image input sequence from different time steps or spatial positions includes:

[0015] Input sequence of agricultural meteorological images Spatiotemporal data generated by hybrid operations , spatiotemporal data Represents the N+1th sample of the input sequence ;

[0016] Uniformly resize all input image sequences The size of , and normalize it, and finally get the corresponding encoding matrix .

[0017] Furthermore, in the construction of the multi-scale convolutional network, the multi-scale convolutional network includes an encoder module, a translator module and a decoder module. The multi-scale convolutional network extracts spatiotemporal features and performs time evolution learning through a multi-layer convolutional structure to reconstruct spatiotemporal features.

[0018] Furthermore, in the process of spatiotemporal feature extraction, the encoder module uses stacked The ConvNormReLU blocks extract the spatial features of the input data. The operation of each convolution block is to convolve C channels on the spatial dimension (H, W). The specific formula is as follows:

[0019] ;

[0020] In the formula, input The shape is (T, C, H, W), output The shape is ;

[0021] Through these convolutional blocks, the encoder module extracts multi-scale spatial features of the input sequence.

[0022] Furthermore, in the temporal evolution learning of spatiotemporal features, the translator module uses Inception modules are used to learn the time evolution of input data; the Inception module contains 4 parallel Conv2d and kernel convolutions of 3, 5, 7, and 11 respectively, and parallel GroupConv2d operators; in the time evolution part, the change of spatiotemporal features is generated by the following formula: , where the input The shape is (T×C, H, W), and the output The shape is ;

[0023] The translator module captures the spatiotemporal dynamic characteristics of the input data.

[0024] Furthermore, in the reconstruction of spatiotemporal features, the decoder module uses The unConvNormReLU blocks reconstruct the ground truth frame; each deconvolution operator is used to restore the original resolution and perform convolution on the time dimension (T, C, H, W). The specific formula is:

[0025] ;

[0026] Among them, input The shape is , output The shape is (T, C, H, W);

[0027] Through the deconvolution operation, the decoder module recovers the spatiotemporal features of the input data and finally generates the output of multi-scale convolutional network reconstruction.

[0028] Furthermore, the diffusion module of the embodiment of the present invention further processes the preliminary prediction results of the multi-scale convolutional network through a two-stage operation of adding noise and removing noise, thereby improving the prediction accuracy of the model and the expression ability of spatiotemporal features; specifically, each diffusion iteration adds noise to the preliminary prediction results, and removes the noise through a reverse process to obtain the final prediction results. This process effectively reduces the noise interference in the prediction and improves the spatiotemporal prediction ability of the model.

[0029] The step of increasing noise through a diffusion model to capture the connection between spatiotemporal features includes:

[0030] In the diffusion model, preliminary prediction results After the noise addition step, samples with different noise levels are obtained , the noise adding process is performed in the spatial dimension (H, W), and the formula is as follows: ;

[0031] Where: t represents the time step of the diffusion process; represents the weight coefficient associated with time step t, which is used to control the amplitude of the noise; represents the noise sampled from a Gaussian noise distribution; The input dimension is (T, C, H, W), and the obtained The output dimension is .

[0032] Furthermore, the step of gradually eliminating noise in the prediction result through a multi-step reverse diffusion process includes:

[0033] The denoising process gradually removes the noise added to the prediction results and finally restores the prediction results close to the actual ones. , the process is expressed by the formula: ;

[0034] in: Represents parameters related to the time step, which are used to control the speed of denoising; represents a neural network for predicting noise; is the noise sampled from a Gaussian distribution and used as the random perturbation in the denoising step; The input dimension is , The output dimensions of is (T, C, H, W).

[0035] Furthermore, during the training process, the model parameters are iteratively updated to minimize the loss function to ensure that the model can achieve accurate prediction under different meteorological conditions; wherein, the step of iteratively updating the model parameters to minimize the mean square error between the predicted image and the real image until the loss function converges includes: mean square error MSELoss, which is used to measure the difference in pixel values ​​between the predicted image and the real image; the overall loss is calculated by mean square error MSE and mean absolute error MAE, and the performance of the prediction model is evaluated by combining the structural similarity index SSIM and peak signal-to-noise ratio PSNR as auxiliary evaluation indicators; wherein:

[0036] The overall loss is expressed as:

[0037] ;

[0038] The mean absolute error MAE is expressed as:

[0039] ;

[0040] The structural similarity SSIM is expressed as:

[0041] ;

[0042] The peak signal-to-noise ratio PSNR is expressed as:

[0043] ;

[0044] In the formula, pred represents the model prediction value, true represents the true value, and N represents the number of samples. and are the means of the predicted image and the real image respectively, and are the variances of the predicted image and the true image, respectively. and It is a constant, a constant and This is set according to the dynamic range of SSIM to stabilize the calculation of SSIM.

[0045] Compared with the prior art, the agricultural meteorological forecasting method based on the multi-scale convolutional network and diffusion system of the present invention has the following beneficial effects:

[0046] First, the present invention makes full use of the multi-level feature information of spatiotemporal data through the combination of multi-scale convolution modules and diffusion modules, and optimizes the feature representation through the diffusion process, thereby improving the generalization ability and long-term prediction accuracy of the model in dynamic spatiotemporal data;

[0047] Second, the present invention adopts layer-by-layer convolution and diffusion modules of multiple convolutional neural networks. In the first stage, multi-level spatiotemporal features are extracted through a multi-scale convolution module. In the second stage, the feature information is mutually transferred in the time and space dimensions through a diffusion module to optimize the feature representation and improve the stability of the prediction.

[0048] Third, in the diffusion optimization process based on feature enhancement, the present invention performs diffusion optimization after convolution feature extraction, propagates feature information in complex spatiotemporal dynamics through the diffusion process, enhances the complementarity and correlation of spatiotemporal features, and ensures the robustness and accuracy of the prediction model to long-term spatiotemporal dynamic changes.

[0049] In summary, the present invention significantly improves the performance of long-term prediction results of spatiotemporal data through the innovative spatiotemporal convolutional neural network combined with the diffusion system. Compared with the current most advanced models, it has obvious improvements in various evaluation indicators, especially in image quality and the capture of key spatiotemporal information, which solves the problem that the existing technology cannot fully capture spatiotemporal dependencies. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a system framework diagram of the agricultural meteorological forecasting method based on a multi-scale convolutional network and a diffusion system of the present invention;

[0051] Figure 2 This is a flow chart for implementing the agricultural meteorological forecasting method based on a multi-scale convolutional network and a diffusion system of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] At present, agricultural meteorological spatiotemporal data mining methods of related technologies can be mainly divided into two categories: one is based on statistical models, and the other is based on deep learning. Methods based on statistical models usually rely on the distribution characteristics of assumed meteorological data to estimate climate trends and periodic changes through models. Such methods include spatiotemporal autoregressive model (STAR), Kriging interpolation (Kriging), etc. Although these models perform well when the data is sparse or the structure is simple, when dealing with complex and high-dimensional spatiotemporal data, it is often difficult to effectively capture the nonlinear characteristics in the data, especially sudden changes and extreme weather events in meteorological data; methods based on deep learning use neural networks to automatically learn complex patterns in spatiotemporal data.

[0054] Different from the existing dynamic spatiotemporal prediction methods, the present invention provides an agricultural meteorological prediction method based on a multi-scale convolutional network and a diffusion system. By combining a multi-scale convolution module and a diffusion module, the method can extract multi-level feature information in spatiotemporal data, and optimize the feature representation through the diffusion process, thereby enhancing the robustness and accuracy of long-term predictions. Furthermore, the present invention extracts and converts spatiotemporal features through the collaborative work of an encoder, a time evolution module, and a decoder, and mutually transmits feature representations containing complex spatiotemporal dynamics through a diffusion process, thereby effectively improving the accuracy of the model in long-term predictions.

[0055] Specifically, the present invention uses a multi-scale convolutional neural network to extract features from agricultural meteorological image data from different time steps and spatial positions. Specifically, the input image sequence is normalized and converted into a unified image coding matrix. The multi-scale convolutional network captures complex patterns in agricultural meteorological images from different spatiotemporal scales and positions, thereby extracting spatiotemporal features. Then, the preliminary prediction results are denoised by a diffusion model. The diffusion model gradually reduces the noise in the prediction through a multi-step inverse diffusion process to ensure that the generated output image is clearer and more accurate. On this basis, the prediction results of different time steps are fused into a comprehensive prediction output. Finally, the fused results are further optimized by the diffusion model to generate the final accurate spatiotemporal prediction. This process is designed to ensure that the model can accurately capture the dynamic changes of agricultural meteorological images and provide farmers and decision makers with agricultural meteorological forecasts for multiple time steps in the future, covering the spatiotemporal changes of meteorological conditions such as temperature and precipitation.

[0056] The specific implementation of the agricultural meteorological prediction method based on a multi-scale convolutional network and a diffusion system of the present invention is described in detail below in conjunction with specific embodiments.

[0057] like Figure 1 As shown, a system framework diagram of the present invention is shown, in which Inception represents a multi-scale feature extraction module; Conv2d represents a two-dimensional convolution; GroupNorm represents group normalization; LeakyReLU represents a leaky linear rectified unit (activation function); and GroupConv2d represents a two-dimensional group convolution.

[0058] Please refer to Figure 2 In one embodiment of the present invention, an agricultural meteorological forecasting method based on a multi-scale convolutional network and a diffusion system is provided, and the agricultural meteorological forecasting method includes:

[0059] S101: Encoding the agricultural meteorological image input sequences from different time steps or spatial positions to obtain corresponding encoding matrices;

[0060] S102: construct a multi-scale convolutional network, use the encoding matrix to train the multi-scale convolutional network, extract the spatiotemporal features of the input sequence, use the trained multi-scale convolutional network to make a preliminary prediction of the agricultural meteorological image, and obtain a preliminary prediction result;

[0061] S103: Processing the preliminary prediction results through a diffusion model, wherein the processing of the preliminary prediction results includes: firstly, adding noise through a diffusion model to capture the connection between spatiotemporal features, and then gradually eliminating the noise in the prediction results through a multi-step inverse diffusion process;

[0062] S104: training a prediction model based on a multi-scale convolutional network and a diffusion model. During the training process, the mean square error between the predicted image and the true image is minimized by iteratively updating the model parameters until the loss function converges, thereby obtaining a trained prediction model.

[0063] S105: For a given agricultural meteorological image, a prediction model is used to generate and output a prediction image sequence for multiple future time steps, wherein the prediction image sequence is used to reflect the spatiotemporal changes of meteorological conditions.

[0064] The present invention significantly improves the performance of long-term prediction results of spatiotemporal data by combining an innovative spatiotemporal convolutional neural network with a diffusion system. Compared with the current most advanced models, it has obvious improvements in various evaluation indicators, especially in image quality and the capture of key spatiotemporal information, which solves the problem that existing technologies cannot fully capture spatiotemporal dependencies.

[0065] Among them, step S101 in the embodiment of the present invention is a multi-scale convolutional network coding process, specifically, the agricultural meteorological image input sequence Spatiotemporal data generated by hybrid operations , spatiotemporal data Represents the N+1th sample of the input sequence ; Uniformly adjust all input image sequences The size of , and normalize it, and finally get the corresponding encoding matrix ; Wherein, the unified image size includes: using a bilinear interpolation algorithm to All are scaled to size The image is used to make the input image meet the input specifications of the multi-scale convolutional network; the image normalization includes: dividing all pixel values ​​in the image x of the same size by 255, and then normalizing the original values ​​(val) of the three channels of RGB of the image to obtain the normalized value X, wherein the mean (mean) of the three channels of RGB is 0, and the standard deviation (std) is 1. Through the normalization process, the gradient explosion of the neural network model is prevented during the training process; wherein, the original values ​​(val) of the three channels of RGB of the image are normalized using the following formula: .

[0066] Step S102 of the embodiment of the present invention is a process of preliminary training of a multi-scale convolutional network. Specifically, the encoding matrix And the corresponding multi-scale convolutional network training. Through this model, the spatiotemporal features in the input sequence are extracted, and the hybrid coding matrix is ​​used Further improve the model's predictive ability. The model achieves accurate prediction of agricultural meteorological images through multi-scale convolution and spatiotemporal feature extraction;

[0067] Step S103 of the embodiment of the present invention is a process of further training the diffusion system, and the preliminary prediction results are processed by adding noise and denoising through the diffusion model. The diffusion model first captures the specific connection between time and space by adding noise, and then gradually eliminates the noise in the prediction to improve the accuracy and precision of the spatiotemporal prediction, ensuring that the model accurately captures future meteorological changes;

[0068] Step S104 of the embodiment of the present invention is a process of minimizing the overall loss. Specifically, during the training process, the mean square error (MSELoss) between the predicted image and the real image is minimized by iteratively updating the model parameters, so as to ensure accurate prediction under various meteorological conditions until the loss function converges.

[0069] Step S105 of the embodiment of the present invention is a process of predicting target agrometeorological data. For a given agrometeorological image, the model generates a sequence of predicted images for multiple time steps in the future, reflecting the spatiotemporal changes of meteorological conditions such as temperature and precipitation. The final prediction results provide a reliable basis for agrometeorological forecasts for farmers and decision makers.

[0070] Furthermore, in the construction of the multi-scale convolutional network, the multi-scale convolutional network includes an encoder module, a translator module and a decoder module. The multi-scale convolutional network extracts spatiotemporal features and performs time evolution learning through a multi-layer convolutional structure to reconstruct spatiotemporal features.

[0071] In the process of extracting spatiotemporal features, the encoder module uses stacked ConvNormReLU blocks To extract the spatial features of the input data, the operation of each convolution block is to convolve C channels on the spatial dimension (H, W). The specific formula is as follows:

[0072] ;

[0073] In the formula, input The shape is (T, C, H, W), output The shape is ;

[0074] Through these convolutional blocks, the encoder module extracts multi-scale spatial features of the input sequence.

[0075] In the temporal evolution learning of spatiotemporal features, the translator module uses Inception modules are used to learn the time evolution of input data; the Inception module contains 4 parallel Conv2d and kernel convolutions of 3, 5, 7, and 11 respectively, and parallel GroupConv2d operators; in the time evolution part, the change of spatiotemporal features is generated by the following formula: , where the input The shape is (T×C, H, W), and the output The shape is ;

[0076] The translator module captures the spatiotemporal dynamic characteristics of the input data.

[0077] In the reconstruction of spatiotemporal features, the decoder module uses unConvNormReLU blocks Reconstruct the ground truth frame; each deconvolution operator is used to restore the original resolution and perform convolution on the time dimension (T, C, H, W). The specific formula is: ;

[0078] Among them, input The shape is , output The shape is (T, C, H, W);

[0079] Through the deconvolution operation, the decoder module recovers the spatiotemporal features of the input data and finally generates the output of multi-scale convolutional network reconstruction.

[0080] Further, in an embodiment of the present invention, the step of increasing the noise by using a diffusion model to capture the connection between the temporal and spatial features includes: in the diffusion model, the preliminary prediction result After the noise addition step, samples with different noise levels are obtained , the noise adding process is performed in the spatial dimension (H, W), and the formula is as follows:

[0081] ;

[0082] Where: t represents the time step of the diffusion process; represents the weight coefficient associated with time step t, which is used to control the amplitude of the noise; represents the noise sampled from a Gaussian noise distribution; The input dimension is (T, C, H, W), and the obtained The output dimension is .

[0083] Furthermore, the step of gradually eliminating noise in the prediction result through a multi-step reverse diffusion process includes:

[0084] The denoising process gradually removes the noise added to the prediction results and finally restores the prediction results close to the actual ones. , the process is expressed by the formula: ;

[0085] in: Represents parameters related to the time step, which are used to control the speed of denoising; represents a neural network for predicting noise; is the noise sampled from a Gaussian distribution and used as the random perturbation in the denoising step; The input dimension is , The output dimensions of is (T, C, H, W).

[0086] Furthermore, during the training process, the model parameters are iteratively updated to minimize the loss function to ensure that the model can achieve accurate prediction under different meteorological conditions; wherein, the step of iteratively updating the model parameters to minimize the mean square error between the predicted image and the real image until the loss function converges includes: mean square error MSELoss, which is used to measure the difference in pixel values ​​between the predicted image and the real image; the overall loss is calculated by mean square error MSE and mean absolute error MAE, and the structural similarity index SSIM and peak signal-to-noise ratio PSNR are combined as auxiliary evaluation indicators to evaluate the performance of the prediction model.

[0087] Therefore, the present invention iteratively updates the model parameters to minimize the mean square error between the predicted image and the real image until the loss function converges, including: the overall loss is calculated by the mean square error MSE and the mean absolute error MAE, and the structural similarity index SSIM and the peak signal-to-noise ratio PSNR are combined as auxiliary evaluation indicators to evaluate the performance of the prediction model; wherein:

[0088] The overall loss is expressed as:

[0089] ;

[0090] The mean absolute error MAE is expressed as:

[0091] ;

[0092] The structural similarity SSIM is expressed as:

[0093] ;

[0094] The peak signal-to-noise ratio PSNR is expressed as:

[0095] ;

[0096] In the formula, pred represents the model prediction value, true represents the true value, and N represents the number of samples. and are the means of the predicted image and the real image respectively, and are the variances of the predicted image and the true image, respectively. and It is a constant, a constant and It is set according to the dynamic range of SSIM to stabilize the calculation of SSIM.

[0097] The present invention makes full use of the multi-level feature information of spatiotemporal data through the combination of multi-scale convolution modules and diffusion modules, optimizes feature representation through the diffusion process, and improves the generalization ability of the model in dynamic spatiotemporal data and the long-term prediction accuracy; the present invention adopts layer-by-layer convolution and diffusion modules of multiple convolutional neural networks, extracts multi-level spatiotemporal features through the multi-scale convolution module in the first stage, and transmits feature information to each other in time and space dimensions through the diffusion module in the second stage, optimizes feature representation, and improves prediction stability; in the diffusion optimization process based on feature enhancement, the present invention performs diffusion optimization after convolution feature extraction, propagates feature information in complex spatiotemporal dynamics through the diffusion process, enhances the complementarity and correlation of spatiotemporal features, and ensures the robustness and accuracy of the prediction model to long-term spatiotemporal dynamic changes.

[0098] In summary, the present invention significantly improves the performance of long-term prediction results of spatiotemporal data through the innovative spatiotemporal convolutional neural network combined with the diffusion system. Compared with the current most advanced models, it has obvious improvements in various evaluation indicators, especially in image quality and the capture of key spatiotemporal information, which solves the problem that the existing technology cannot fully capture spatiotemporal dependencies.

[0099] Another embodiment of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the agricultural meteorological forecasting method based on a multi-scale convolutional network and a diffusion system of any one of the above-mentioned embodiments is implemented.

[0100] The computer device includes a processor and a memory, and may also include an input system and an output system. The processor, memory, input system and output system may be connected via a bus or other means, and the input system may receive input digital or character information, and generate signal input related to agricultural meteorological forecasting based on a multi-scale convolutional network and a diffusion system. The output system may include display devices such as a display screen.

[0101] As a non-volatile computer-readable storage medium, the memory can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as the program instructions / modules corresponding to the agricultural meteorological prediction method based on the multi-scale convolutional network and the diffusion system in the embodiment of the present application. The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the use of the agricultural meteorological prediction method based on the multi-scale convolutional network and the diffusion system, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the local module via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0102] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is generally used to control the overall operation of the computer device. In this embodiment, the processor is used to run the program code stored in the memory or process data. The processors of the multiple computer devices of the computer device of this embodiment execute various functional applications and data processing of the server by running non-volatile software programs, instructions and modules stored in the memory, that is, implementing the steps of the agricultural meteorological prediction method based on the multi-scale convolutional network and diffusion system of the above method embodiment.

[0103] It will also be appreciated by those skilled in the art that various exemplary logic blocks, modules, circuits and algorithm steps described in conjunction with the disclosure herein can be implemented as electronic hardware, computer software or a combination of the two. In order to clearly illustrate this interchangeability of hardware and software, a general description has been given to the functions of various schematic components, blocks, modules, circuits and steps. Whether this function is implemented as software or hardware depends on specific applications and the design constraints imposed on the entire system. Those skilled in the art can implement the function in various ways for each specific application, but this implementation decision should not be interpreted as causing a departure from the disclosed scope of the embodiments of the present invention.

[0104] Finally, it should be noted that the computer-readable storage medium (e.g., memory) herein may be a volatile memory or a non-volatile memory, or may include both a volatile memory and a non-volatile memory. As an example and not by way of limitation, a non-volatile memory may include a read-only memory (ROM), a programmable ROM (PROM), an electrically programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), or a flash memory. A volatile memory may include a random access memory (RAM), which may act as an external cache memory. As an example and not by way of limitation, RAM may be obtained in a variety of forms, such as synchronous RAM (DRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The storage devices of the disclosed aspects are intended to include, but are not limited to, these and other suitable types of memory.

[0105] The various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure herein may be implemented or executed using the following components designed to perform such functions: a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination of these components. A general-purpose processor may be a microprocessor, but alternatively, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP, and / or any other such configuration.

[0106] The above are exemplary embodiments disclosed in the present invention, but it should be noted that various changes and modifications may be made without departing from the scope disclosed in the embodiments of the present invention as defined in the claims. The functions, steps and / or actions of the method claims according to the disclosed embodiments described herein do not need to be performed in any particular order. In addition, although the elements disclosed in the embodiments of the present invention may be described or required in individual form, they may also be understood as multiple unless explicitly limited to the singular.

[0107] It should be understood that, as used herein, the singular form "a" or "an" is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, "and / or" refers to any and all possible combinations of one or more of the items listed in association. The serial numbers of the embodiments disclosed in the above embodiments of the present invention are for description only and do not represent the advantages and disadvantages of the embodiments.

[0108] A person skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the disclosure of the embodiments of the present invention (including the claims) is limited to these examples; under the concept of the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other changes in different aspects of the above embodiments of the present invention, which are not provided in detail for the sake of simplicity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the protection scope of the embodiments of the present invention.

Claims

1. The agricultural meteorological forecasting method based on multi-scale convolutional network and diffusion system is characterized by: The agricultural meteorological forecasting method includes: S101: Encoding the agricultural meteorological image input sequences from different time steps or spatial positions to obtain corresponding encoding matrices; S102: construct a multi-scale convolutional network, use the encoding matrix to train the multi-scale convolutional network, extract the spatiotemporal features of the input sequence, use the trained multi-scale convolutional network to make a preliminary prediction of the agricultural meteorological image, and obtain a preliminary prediction result; S103: Processing the preliminary prediction results through a diffusion model, wherein the processing of the preliminary prediction results includes: firstly, adding noise through the diffusion model to capture the connection between the spatiotemporal features, and then gradually eliminating the noise in the prediction results through a multi-step inverse diffusion process; S104: training a prediction model based on a multi-scale convolutional network and a diffusion model. During the training process, the mean square error between the predicted image and the true image is minimized by iteratively updating the model parameters until the loss function converges, thereby obtaining a trained prediction model. S105: for a given agricultural meteorological image, using a prediction model to generate and output a prediction image sequence of multiple future time steps, wherein the prediction image sequence is used to reflect the spatiotemporal changes of meteorological conditions; The steps of encoding the agricultural meteorological image input sequence from different time steps or spatial positions include: Input sequence of agricultural meteorological images Spatiotemporal data generated by hybrid operations , spatiotemporal data Represented as the N+1th sample of the input sequence ; Uniformly resize all input image sequences The size of , and normalize it, and finally get the corresponding encoding matrix ; In constructing a multi-scale convolutional network, the multi-scale convolutional network includes an encoder module, a translator module and a decoder module. The multi-scale convolutional network extracts spatiotemporal features and learns time evolution through a multi-layer convolutional structure to reconstruct spatiotemporal features. In the process of spatiotemporal feature extraction, the encoder module uses stacked The ConvNormReLU blocks extract the spatial features of the input data. The operation of each convolution block is to convolve C channels on the spatial dimension (H, W). The specific formula is as follows: ； In the formula, input The shape is (T, C, H, W), output The shape is ; Through these convolutional blocks, the encoder module extracts multi-scale spatial features of the input sequence; In the temporal evolution learning of spatiotemporal features, the translator module uses Inception modules are used to learn the time evolution of input data; the Inception module contains 4 parallel Conv2d and kernel convolutions of 3, 5, 7, and 11 respectively, and parallel GroupConv2d operators; in the time evolution part, the change of spatiotemporal features is generated by the following formula: , where the input The shape is (T×C, H, W), and the output The shape is ; The translator module captures the spatiotemporal dynamic characteristics of the input data; In the reconstruction of spatiotemporal features, the decoder module uses The unConvNormReLU blocks reconstruct the ground truth frame; each deconvolution operator is used to restore the original resolution and perform convolution on the time dimension (T, C, H, W). The specific formula is: ； Among them, input The shape is , output The shape is (T, C, H, W); Through deconvolution operations, the decoder module recovers the spatiotemporal features of the input data and finally generates the output of multi-scale convolutional network reconstruction; The steps of increasing noise to capture the connection between spatiotemporal features through diffusion model include: In the diffusion model, preliminary prediction results After the noise addition step, samples with different noise levels are obtained , the noise adding process is performed in the spatial dimension (H, W), and the formula is as follows: ; Where: t represents the time step of the diffusion process; represents the weight coefficient associated with time step t, which is used to control the amplitude of the noise; represents the noise sampled from a Gaussian noise distribution; The input dimension is (T, C, H, W), and the obtained The output dimension is ; The steps of gradually eliminating the noise in the prediction results through a multi-step inverse diffusion process include: The denoising process gradually removes the noise added to the prediction results and finally restores the prediction results close to the actual ones. , the process is expressed by the formula: ； in: Represents parameters related to the time step, which are used to control the speed of denoising; represents a neural network for predicting noise; is the noise sampled from a Gaussian distribution and used as the random perturbation in the denoising step; The input dimension is , The output dimensions of is (T, C, H, W).

2. The agricultural meteorological forecasting method based on multi-scale convolutional network and diffusion system according to claim 1 is characterized in that: The steps of iteratively updating the model parameters to minimize the mean square error between the predicted image and the true image until the loss function converges include: The overall loss is calculated by the mean square error (MSE) and the mean absolute error (MAE), and the structural similarity index (SSIM) and the peak signal-to-noise ratio (PSNR) are used as auxiliary evaluation indicators to evaluate the performance of the prediction model; The overall loss is expressed as: ； The mean absolute error MAE is expressed as: ； The structural similarity SSIM is expressed as: ； The peak signal-to-noise ratio PSNR is expressed as: ； In the formula, pred represents the model prediction value, true represents the true value, and N represents the number of samples. and are the means of the predicted image and the real image respectively, and are the variances of the predicted image and the true image, respectively. and is a constant.

Images