Winter wheat yield prediction method based on XLSTM and assimilation crop growth model

Abstract

The invention relates to the technical field of agricultural big data processing, and discloses a winter wheat yield prediction method based on XLSTM and assimilation crop growth model, comprising the following steps: data acquisition and data preprocessing; extracting a winter wheat planting area; performing leaf area index inversion; correlation analysis and feature extraction; establishing a crop growth model based on a WOFOST model; performing data assimilation based on ensemble Kalman filtering; establishing a time sequence data set; establishing a yield prediction model based on the extended long-short-term memory neural network; and model prediction: inputting the current remote sensing image and the meteorological data of the future time period into a yield prediction model, and outputting a yield prediction value of the winter wheat in the target area. According to the method, by integrating and fusing multi-source data and combining complex models and algorithms, the problem that fusion of the multi-source data and the complex relation between the multi-source data and the yield are not fully considered due to the fact that only a single feature is concerned or the models are not fully considered is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural big data processing, and more specifically, to a winter wheat yield prediction method based on XLSTM and an assimilated crop growth model. Background Art

[0002] Crop yield prediction has always been a key link in agricultural production and management. Accurate yield prediction is of great significance for ensuring food security, guiding farmers to arrange planting rationally, regulating the balance of supply and demand in the agricultural product market, formulating agricultural policies, and carrying out agricultural insurance. With the global population growth and the intensification of climate change, it is particularly important to improve the accuracy and efficiency of crop yield prediction. Traditional yield prediction methods mainly rely on field sampling and statistical models. This method has the disadvantages of being time-consuming and labor-intensive, highly destructive, limited in coverage, and low in prediction accuracy, which makes it difficult to meet the needs of modern agricultural production and management. In recent years, the rise of remote sensing technology and machine learning methods has provided a new, efficient and accurate means for crop yield prediction.

[0003] Remote sensing technology has been widely used in agricultural production due to its non-destructive, real-time and large-scale monitoring characteristics. Satellite remote sensing images can be used to obtain real-time information on the growth of crops over a large area, thus providing rich data support for crop yield prediction. By calculating the normalized vegetation index enhanced vegetation index (EVI) in remote sensing images, the growth status of vegetation can be quantified. EVI introduces a blue light band to correct atmospheric effects and soil background noise, and exhibits better dynamic range and accuracy in high-density vegetation areas, which is particularly suitable for reflecting the growth dynamics and health status of vegetation. In addition, remote sensing images can also be combined with meteorological data such as precipitation, temperature and net surface radiation to further analyze the impact of meteorological factors on crop growth and yield. However, it is difficult to obtain accurate results for yield prediction by relying solely on remote sensing data, because crop yields are affected by a variety of factors, including soil conditions, climate change, pests and diseases, and management measures.

[0004] In recent years, the integration of crop growth mechanism models and data assimilation technology has provided new breakthroughs in yield prediction. The mechanism model can simulate the growth process of crops from sowing to maturity based on environmental data such as meteorology and soil, and generate time series data of key growth indicators such as leaf area index (LAI). An example of a mechanism model is the WOFOST (WOrld FOodSTudies) model. Data assimilation technology (such as ensemble Kalman filtering, EnKF) can effectively integrate mechanism model simulation data with remote sensing observation data to dynamically correct and optimize crop growth status. This combination can make up for the limitations of a single data source, improve the accuracy and timeliness of yield prediction, and provide more scientific and comprehensive decision support for agricultural production.

[0005] In order to improve the accuracy of crop yield prediction, machine learning methods have been widely used in the field of agricultural remote sensing in recent years. Machine learning algorithms, especially ensemble learning methods such as LSTM, have the ability to process high-dimensional and diversified data, and can make full use of remote sensing images and multi-source data for yield prediction. XLSTM (Extended Long Short-Term Memory Network) is an improvement on the traditional LSTM. By introducing exponential gating and enhanced memory structure, it effectively solves the inherent limitations of LSTM. XLSTM includes two new variants: sLSTM (scalar LSTM) and mLSTM (matrix LSTM). sLSTM uses sigmoid activation function, which enables input gates and forget gates to control information flow more flexibly, while introducing standardized states to maintain numerical stability; mLSTM uses matrix memory to replace traditional scalar units, greatly improving storage capacity, and can more efficiently store and retrieve complex data patterns. In addition, XLSTM achieves efficient calculation of a large number of parameters through residual stacking and parallelization technology, and can handle ultra-long sequences and large-scale context information. These improvements enable XLSTM to excel in capturing time series features, dynamic relationships, and long-range dependencies, surpassing traditional LSTM and some Transformer models, and providing more powerful support for natural language processing, time series analysis, and other tasks that require long-term memory.

[0006] At present, some studies have used remote sensing technology and machine learning methods to estimate crop yields. For example, using remote sensing image data such as MODIS and Landsat, combined with vegetation indices such as the Normalized Difference Vegetation Index (NDVI) and the Ratio Vegetation Index (RVI), yield prediction is performed through linear regression, support vector machine and other methods. These methods have improved the accuracy of prediction to a certain extent, but there are still some limitations. First, existing studies often only focus on a single feature or model, and fail to fully consider the fusion of multi-source data and its complex relationship with yield; second, some methods have high requirements for data processing and model construction, and are complex to operate, making them difficult to promote and apply in actual agricultural production. Summary of the invention

[0007] The present invention provides a winter wheat yield prediction method based on XLSTM and assimilated crop growth model, which solves the technical problem that the related technology only focuses on a single feature or model and fails to fully consider the fusion of multi-source data and its complex relationship with yield.

[0008] The present invention provides a winter wheat yield prediction method based on XLSTM and assimilation crop growth model, comprising the following steps:

[0009] The present invention provides a winter wheat yield prediction method based on XLSTM and assimilation crop growth model, comprising the following steps:

[0010] Step 100, data acquisition and preprocessing: using a multispectral remote sensing satellite to acquire remote sensing image data of the winter wheat planting area of ​​the target area, and simultaneously acquiring meteorological data, soil data, yield data and crop parameter data; acquiring remote sensing image data of the winter wheat planting area of ​​the target area, selecting remote sensing images by season according to phenological information, performing radiation calibration, atmospheric correction and orthorectification, generating high-precision remote sensing image data, and cutting and mosaicking the remote sensing image data by vector boundaries;

[0011] Step 200, winter wheat planting area extraction: visual interpretation to obtain winter wheat planting area image samples; calculate sample separability; use random forest classification model to supervise the remote sensing image of winter wheat planting area, realize feature layered extraction, and obtain winter wheat planting area vector; perform vector clipping on the remote sensing image of winter wheat planting area; confusion matrix analysis;

[0012] Step 300, leaf area index inversion: calculate the enhanced vegetation index based on the pre-processed remote sensing image data of the winter wheat planting area; obtain the leaf area index by using the leaf area empirical inversion model in different seasons;

[0013] Step 400, correlation analysis and feature extraction: combining the acquired meteorological data and yield data, performing classical correlation analysis and principal component analysis on the multi-dimensional feature data, calculating the correlation between different features and yield, and retaining the main features related to yield;

[0014] Step 500, establishing a crop growth model based on the WOFOST model: combining the acquired meteorological data and soil data, referring to the crop parameter settings in the same region, establishing a crop growth model based on the WOFOST model, and outputting the LAI time series data of the entire growth period of winter wheat;

[0015] Step 600, performing data assimilation based on ensemble Kalman filtering: Based on ensemble Kalman filtering, using the LAI time series data obtained from the crop growth model, data assimilation is performed on the LAI time series data obtained from remote sensing image processing of the winter wheat planting area to obtain accurate LAI time series data;

[0016] Step 700, establishing a time series data set: sorting the LAI and meteorological principal component data corresponding to the pixels with corresponding station yield data in the remote sensing image of the winter wheat planting area in chronological order to form a time series input;

[0017] Step 800, establishing a production prediction model based on an extended long short-term memory neural network: establishing an extended long short-term memory neural network as a production prediction model based on a time series data set and actual site production data;

[0018] Step 900, yield prediction: input the remote sensing image of the current winter wheat planting area and the meteorological data of the future period into the yield prediction model, and output the yield prediction value of winter wheat in the target area.

[0019] Furthermore, the remote sensing image data comes from the Gaofen satellite, and the input features include four bands, namely the blue light band, the green light band, the red light band and the near-infrared band.

[0020] Furthermore, the calculation formula of the enhanced vegetation index is as follows:

[0021]

[0022] in, represents the enhanced vegetation index, NIR represents the reflectance of the near-infrared band, RED represents the reflectance of the red band, and BLUE represents the reflectance of the blue band.

[0023] Furthermore, the empirical inversion model of leaf area is as follows:

[0024] Jointing period:

[0025]

[0026] The grouting period is:

[0027]

[0028] Among them, EVI stands for Enhanced Vegetation Index, is the leaf area index.

[0029] Furthermore, the characteristics of the principal component analysis in step 400 include the daily average 2-meter-high temperature, the daily average precipitation, and the daily average surface net thermal radiation.

[0030] Furthermore, the specific steps of constructing the crop growth model include:

[0031] Step 201, data collation: data cleaning is performed on the collected meteorological data, soil data, and crop physiological parameter data, and the data sets are collated and compiled into a YAML format;

[0032] Step 202, initial parameter setting: setting the initial parameters of the WOFOST model according to the characteristics of the winter wheat variety, and setting the parameters of the meteorological, soil and crop management measures;

[0033] Step 203, crop growth simulation: based on the WOFOST model, the crop growth process is simulated, by dividing the crop growth process into multiple stages, dynamically adjusting the growth rate and biomass accumulation of each stage according to the input meteorological data and soil data, and simulating and calculating the biomass accumulation of crops at different growth stages;

[0034] Step 204, data output: obtain LAI time series data based on WOFOST model simulation;

[0035] The WOFOST model is used as a crop growth model.

[0036] Furthermore, the specific steps of data assimilation in step 600 include:

[0037] Step 301, initialization set: construct an initial set containing multiple samples, each sample represents a possible estimated value of LAI, and the initial set is composed of LAI time series data generated by the WOFOST model and LAI data obtained by remote sensing image processing as the starting state;

[0038] Step 302, calculating observation residuals: comparing the LAI observation data obtained by remote sensing image processing with the LAI time series data obtained by model prediction, and calculating observation residuals;

[0039] Step 303, Kalman gain calculation: Calculate the Kalman gain according to the residual covariance of the predicted value and the error covariance of the observed data;

[0040] Step 304, update the prediction state: adjust the prediction value of each sample according to the Kalman gain, integrate the observed data into the model prediction, and thus update the prediction value of LAI;

[0041] Step 305, updating the covariance matrix: calculating the updated covariance matrix to provide uncertainty measurement for the next prediction;

[0042] Step 306, iterative update: repeat the prediction and update steps, and each time new remote sensing observation data enters, use the ensemble Kalman filter algorithm to adjust and optimize the LAI time series data until the error reaches the allowable value;

[0043] Step 307, pixel-by-pixel processing: performing the data assimilation of the above steps on the LAI data of each pixel in the remote sensing image;

[0044] Step 308, output data: output the assimilated accurate LAI time series data.

[0045] Furthermore, in step 700, a time series input is formed:

[0046] , ;in represents the time step, , , , Respectively represent the 1st, 2nd, t, The feature vector at time, represents the leaf area index at time t, Represents the principal component analysis data at time t.

[0047] Furthermore, the specific steps of establishing the production prediction model based on the extended long short-term memory neural network include:

[0048] Step 401: Dataset division: The data set is divided into a training set and a validation set, where the training set accounts for 70% of the total data and the validation set accounts for 30%;

[0049] Step 402: Module establishment: establish sLSTM and mLSTM modules, use sigmoid function as activation function, and adjust the number of neurons through multiple trainings;

[0050] Step 403: residual stacking: connect the multi-layer sLSTM module and the mLSTM module through residual connection, combine the output and input of the xLSTM module, perform nonlinear processing and then feed back to the extended long short-term memory neural network;

[0051] Step 404: Model training: The time series data in the training set is input into the extended long short-term memory neural network, and is gradually added to the training set through five independent predictions and the average value is taken. The mean square error is used as the loss function, and the model is repeatedly trained and the parameters are adjusted.

[0052] Loss Function The calculation formula is as follows:

[0053]

[0054] in, is the true value, is the predicted value, Represents the total number of training samples;

[0055] Step 405: Model evaluation: by making predictions on the validation set, calculating the root mean square error, mean square error, mean absolute error and coefficient of determination;

[0056] The calculation formula of the model evaluation index is as follows:

[0057]

[0058]

[0059]

[0060]

[0061] in is the root mean square error, is the mean square error, is the mean absolute error, is the determination coefficient.

[0062] The present invention provides a computer storage medium, wherein the computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the processor executes the aforementioned image drawing method based on a generation system technology of big data and artificial intelligence.

[0063] The beneficial effects of the present invention are:

[0064] The present invention integrates and fuses multi-source data, such as remote sensing images, meteorological data, soil data and crop parameters, and combines complex models and algorithms, such as the WOFOST crop growth model, ensemble Kalman filter and extended long short-term memory neural network (XLSTM), to solve the problem of focusing only on a single feature or the model failing to fully consider the fusion of multi-source data and its complex relationship with yield;

[0065] Accurate monitoring of key parameters (such as LAI) during crop growth can help achieve precision agriculture, reduce the use of fertilizers and pesticides, protect the environment, and promote sustainable agricultural development. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flow chart of the winter wheat yield prediction method based on XLSTM and crop growth model assimilation of the present invention. DETAILED DESCRIPTION

[0067] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the present specification. Various examples may omit, replace, or add various processes or components as needed. In addition, the features described in some examples may also be combined in other examples.

[0068] At least one embodiment of the present invention discloses a method for predicting winter wheat yield based on XLSTM and assimilation crop growth model, such as Figure 1 As shown, the following steps are included:

[0069] Step 100, data acquisition and preprocessing: use multispectral remote sensing satellites (such as Gaofen-1 satellite) to obtain remote sensing image data of winter wheat planting areas in the target area, including spectral data during the winter wheat growth cycle; obtain meteorological data, including daily average 2-meter high temperature, daily average precipitation, daily average surface net thermal radiation, etc.; obtain actual yield data from local agricultural departments.

[0070] In some embodiments of the present invention, the remote sensing image data comes from a high-resolution satellite (GF1-WFV wide-angle camera), and the input features include four bands, namely, a blue light band, a green light band, a red light band, and a near-infrared band.

[0071] The central wavelength of the blue light band is 450nm, the wavelength width is 70nm, and the spatial resolution is 16m;

[0072] The central wavelength of the green light band is 520nm, the wavelength width is 80nm, and the spatial resolution is 16m;

[0073] The central wavelength of the red light band is 660nm, the wavelength width is 60nm, and the spatial resolution is 16m;

[0074] The central wavelength of the near-infrared band is 830nm, the wavelength width is 100nm, and the spatial resolution is 16m.

[0075] The acquired remote sensing image data is imported into ENVI (The Environment for Visualizing Images) software for radiation calibration, atmospheric correction and orthorectification to generate high-precision remote sensing image data; the image data is cropped and mosaicked through vector boundaries to retain the crop planting area within the target area.

[0076] In some embodiments of the present invention, data preprocessing includes performing radiometric calibration, Flash atmospheric correction and orthorectification on the remote sensing image data, and cropping and mosaicking the image data using the vector boundaries of the target area.

[0077] In some embodiments of the present invention, the meteorological data is downloaded in batches by downloading ERA5 data from the European Space Agency, including the daily average 2-meter-high temperature, daily average precipitation, and daily average surface net thermal radiation during the growing period of winter wheat.

[0078] In some embodiments of the present invention, actual site yield data is obtained by applying online to the agricultural management department in the winter wheat planting area.

[0079] Step 200, winter wheat planting area extraction: visual interpretation to obtain winter wheat planting area image samples; by calculating the Bhattacharyya distance between various types of ground feature samples to verify the separability, the calculation formula is as follows:

[0080]

[0081] Among them, P(x) and Q(x) are the probability density functions of the two types of samples. The larger the Bhattacharyya distance, the greater the difference between the two types of distributions, and the easier it is to distinguish the samples.

[0082] Then the random forest classification algorithm is used for classification. At the same time, the classification results are verified by confusion matrix to ensure the accuracy of the extracted planting area. The calculation formula of the relevant indicators is:

[0083] Overall Accuracy:

[0084]

[0085] in, Represents the overall accuracy, and are the true positives, false positives, false negatives, and true negatives of class i. The overall accuracy reflects the overall performance of the classifier on all categories.

[0086] Kappa coefficient:

[0087]

[0088] in, is the observed accuracy, is the expected accuracy, and the value range of the Kappa coefficient is , the closer the value is to 1, the better the classification effect is.

[0089] The Kappa coefficient is a statistical indicator used to measure classification consistency, and its value is between -1 and 1. The calculation of the Kappa coefficient is based on the confusion matrix and is mainly used to evaluate the consistency between the model prediction results and the actual classification results.

[0090] In some embodiments of the present invention, the labeled samples include: planting areas, woodlands, grasslands, water areas, construction land and areas to be utilized. After verifying the separability of the samples, the remote sensing images are classified using a random forest classification algorithm, and then the classification effect is verified through confusion matrix analysis, the planting area vector file is extracted, and the remote sensing image is cropped.

[0091] Step 300, leaf area index inversion: calculate the enhanced vegetation index based on the pre-processed remote sensing image data; obtain the leaf area index by using the leaf area empirical inversion model in different time periods;

[0092] Vegetation indices include various spectral indices such as the Enhanced Vegetation Index (EVI), and the leaf area index LAI is inverted using an empirical model;

[0093] Calculate the Enhanced Vegetation Index (EVI) using the following formula:

[0094]

[0095] Among them, NIR represents the reflectivity of the near infrared band, RED represents the reflectivity of the red light band, and BLUE represents the reflectivity of the blue light band. By calculating EVI from the remote sensing image of the target area, the growth status index of the crop is obtained.

[0096] The leaf area index (LAI) is inverted using the empirical inversion model of the leaf area index in different seasons. The calculation formula is as follows:

[0097] Jointing period:

[0098]

[0099] The grouting period is:

[0100]

[0101] Among them, EVI stands for Enhanced Vegetation Index.

[0102] Step 400, correlation analysis and feature extraction: Combine the acquired meteorological data and yield data, perform classical correlation analysis (CCA) and principal component analysis (PCA) on the multi-dimensional feature data, calculate the correlation between different features and yield, and extract the characteristic principal components that have the greatest impact on yield, including the monthly average 2-meter high temperature, monthly average precipitation, monthly average surface net thermal radiation, etc. The principal component analysis steps include:

[0103] Suppose the multidimensional feature data matrix is ​​X, and calculate its covariance matrix S:

[0104]

[0105] in, is the i-th sample, is the feature mean vector, is the total number of samples, represents transpose;

[0106] By solving the eigenvalues ​​and eigenvectors of the covariance matrix, the original data is projected into the principal component space to obtain the principal component matrix :

[0107]

[0108] in, is the eigenvector matrix of the covariance matrix S.

[0109] The first principal component with the highest contribution rate is obtained and recorded as the meteorological principal component PC1.

[0110] Step 500, establishing a crop growth model based on the WOFOST model: combining the acquired meteorological data and soil data, referring to the crop parameter settings in the same region, establishing a crop growth model based on the WOFOST model; outputting the LAI time series data of the entire growth period of winter wheat;

[0111] This step includes the following sub-steps:

[0112] Step 201: Data collation: clean the collected meteorological data (including daily average temperature, maximum and minimum temperature, precipitation, solar radiation, wind speed, humidity), soil data (soil pH value, soil organic matter content, soil moisture status), and crop physiological parameter data, and compile a YAML format data set;

[0113] In some embodiments of the present invention, meteorological data are from NOAA meteorological products, crop physiological parameter data are from previous research settings in the same region, and soil data are from measured data.

[0114] Step 202: Initial parameter setting: setting the initial parameters of the WOFOST model according to the characteristics of the winter wheat variety, and setting the relevant parameters of the meteorological, soil and crop management measures;

[0115] In some embodiments of the present invention, the growing period of winter wheat is determined through actual investigation, and the soil setting is set according to the characteristics of the planting area.

[0116] Step 203: Crop growth simulation: Based on the WOFOST model, the crop growth process is simulated. The crop growth process is divided into multiple stages. The growth rate and biomass accumulation of each stage are dynamically adjusted according to the input meteorological data and soil data, and the biomass accumulation of crops at different growth stages is simulated and calculated.

[0117] Step 204: Data output: obtaining LAI time series data with high time resolution and excellent accuracy based on WOFOST model simulation;

[0118] The WOFOST model in step 204 is used as a crop growth model.

[0119] Step 600, data assimilation based on ensemble Kalman filtering: based on ensemble Kalman filtering, using the LAI time series data obtained from the crop growth model, pixel-by-pixel data assimilation is performed on the LAI time series data obtained from remote sensing image processing to obtain more accurate LAI time series data with higher temporal resolution;

[0120] This step includes the following sub-steps:

[0121] Step 301: Initialize the set: construct an initial set containing multiple samples, each sample represents a possible estimated value of LAI. The initial set is composed of LAI time series data generated by the WOFOST model and LAI data obtained by remote sensing image processing as the starting state.

[0122] In some embodiments of the present invention, the high temporal resolution LAI time series data obtained from the crop growth model is used as the prediction data of the model, and the low temporal resolution LAI time series data obtained from the remote sensing image is used as the observation data. The set is initialized in the following way:

[0123]

[0124] in, is the high temporal resolution LAI time series data obtained from the crop growth model, and N is the number of samples.

[0125] At the same time, the LAI time series data inverted by remote sensing can be expressed as:

[0126]

[0127] in, is the LAI time series data inverted by remote sensing, and M is the number of time steps of the remote sensing image.

[0128] Step 302: Calculate observation residuals: compare the LAI observation data obtained by remote sensing image processing with the LAI time series data obtained by model prediction to calculate the observation residuals.

[0129] In some embodiments of the present invention, for each time step, the observation residual is calculated, that is, the difference between the LAI observation data obtained by the remote sensing image and the LAI prediction data output by the crop growth model. The specific calculation formula is as follows:

[0130]

[0131] in, is the LAI time series data inverted by remote sensing, is the LAI time series data obtained from the crop growth model. is the observed residual.

[0132] Step 303: Kalman gain calculation: Calculate the Kalman gain based on the residual covariance of the predicted value and the error covariance of the observed data. The specific calculation formula is:

[0133]

[0134] in, is the forecast covariance matrix, is the observation matrix, is the observation error covariance matrix, and T represents the matrix transpose.

[0135] Step 304: Update the prediction state: adjust the prediction value of each sample according to the Kalman gain, effectively integrate the observed data into the model prediction, and thus update the prediction value of LAI. The specific calculation formula is:

[0136]

[0137] in, is the prediction data of remote sensing images, is the observation residual, is the Kalman gain, It is the updated LAI timing data.

[0138] Step 305: Update the covariance matrix: Calculate the updated covariance matrix to provide uncertainty measurement for the next prediction. The specific calculation formula is:

[0139]

[0140] in, is the identity matrix, is the updated covariance matrix.

[0141] Step 306: Iterative update: Repeat the prediction and update steps. Each time new observation data enters, the ensemble Kalman filter algorithm is used to adjust and optimize the LAI time series data.

[0142] Step 307: Pixel-by-pixel processing: perform the data assimilation of the above steps on the LAI data of each pixel in the remote sensing image.

[0143] Step 308: Output data: output more accurate LAI timing data.

[0144] Step 700, establishment of a time series data set: sort the LAI and meteorological principal component data of the site production data pixels in chronological order to form a time series input , ,in represents the time step, , , , Respectively represent the 1st, 2nd, t, The feature vector at time, represents the leaf area index at time t, Represents the principal component analysis data at time t.

[0145] In some embodiments of the present invention, the time step is set to 3 days, and a time series data set of 40 time steps is constructed.

[0146] Step 800, establishing a production prediction model based on an extended long short-term memory neural network (XLSTM): establishing an extended long short-term memory neural network based on a time series data set and actual site production data;

[0147] In one embodiment of the present invention, based on the extracted feature data, the winter wheat yield is predicted using an extended long short-term memory network, and the specific steps include:

[0148] Step 401: Dataset division: divide the data set into a training set and a validation set, with the training set accounting for 70% of the total data and the validation set accounting for 30%;

[0149] The input in the training process is the site production and its corresponding time series data in the training set;

[0150] Step 402: Module establishment: XLSTM modules are mainly divided into sLSTM modules and mLSTM modules. sLSTM and mLSTM modules are established, the activation function is the sigmoid function, and the number of neurons is adjusted through multiple trainings to determine the optimal number;

[0151] Among them, sLSTM is a model based on scalar storage, and its core formula is as follows:

[0152]

[0153]

[0154]

[0155]

[0156]

[0157]

[0158]

[0159] in, is the current cell state of sLSTM, representing the core part of the model memory. is the previous cell state of sLSTM, is the regularization state of sLSTM, which is used to stabilize the output and gradient of the model. is the current hidden state of sLSTM, is the previous hidden state of the sLSTM, Describe the input features. It is the input gate of sLSTM and controls the addition of new information. It is the forget gate of sLSTM, which determines how much historical information is forgotten. It is the output gate of sLSTM, which is used to adjust the output state. Candidate gates for sLSTM, controlling the amount of input and to be added The previous hidden state of The amount. , , , They represent the first, second, third, and fourth weight matrices of sLSTM respectively, and T represents transpose. , , , They represent the first, second, third, and fourth recurrent weight matrices of sLSTM respectively. , , , They represent the first, second, third, and fourth bias parameters of sLSTM respectively. is a non-linear activation function.

[0160] Furthermore, in order to prevent the overflow of the exponential gate, a stabilization state is added

[0161]

[0162] is a logarithmic function, is the previous stable state.

[0163] And adjust the input gate and forget gate to:

[0164]

[0165]

[0166] represents the exponential function, is the new input gate, For the new forget gate.

[0167] mLSTM uses matrix storage to increase storage capacity. The core formula is as follows:

[0168]

[0169]

[0170]

[0171]

[0172]

[0173]

[0174]

[0175]

[0176]

[0177] in, is the cell state of mLSTM, is the regularization state of mLSTM, is the hidden state of mLSTM, is the query vector of mLSTM, is the key vector of mLSTM, is the value vector of mLSTM. is the forget gate of mLSTM, is the input gate of mLSTM, is the output gate of mLSTM, , , , , , They represent the first, second, third, fourth, fifth, and sixth weight matrices of mLSTM respectively. , , , , They represent the first, second, third, and fourth bias parameters of sLSTM respectively. is a nonlinear activation function, Indicates or.

[0178] Step 403: Residual stacking: Multi-layer sLSTM modules are connected to mLSTM modules through residual connections, the output of the xLSTM module is combined with the input, non-linearly processed and then fed back to the network. The residual block of the xLSTM combines the output of the xLSTM module with the input, performs non-linear processing and then feeds back to the network. Specifically, there are two architectures:

[0179] In the sLSTM module, a post-projection structure is used to directly add the input and output, as follows:

[0180] Input feature vector After that, the output is:

[0181]

[0182] is the hidden state after the feature vector enters the sLSTM module.

[0183] Then the output is nonlinearly changed through a multi-layer perceptron (MLP):

[0184]

[0185] for The result after nonlinear transformation through MLP.

[0186] Then make a residual connection and add the input directly to the MLP output:

[0187]

[0188] The final output of the residual connection of the sLSTM module;

[0189] mLSTM adopts a front-up projection structure, first performing nonlinear processing in a high-dimensional space, then mapping it back to the original space and adding the input, as follows:

[0190] First, input features Mapping to high-dimensional space:

[0191]

[0192] is the projection matrix, which is responsible for mapping the input feature variables to a high-dimensional space. is the bias term of the upward projection, is the feature vector after mapping to high-dimensional space.

[0193] Perform matrix operations in high-dimensional space:

[0194]

[0195] is the hidden state obtained after mLSTM operation.

[0196] Then map the output of mLSTM back to the original feature space:

[0197]

[0198] is the output feature after mapping back to the original space, is the projection matrix, which is responsible for converting the high-dimensional Mapping back to the original low-dimensional space, is the bias term for the downward projection.

[0199] Then perform residual connection and add the original input directly to the mapping result:

[0200]

[0201] The residual connection is the final output of the mLSTM module.

[0202] Then, residual stacking is performed between multi-layer modules: By stacking multiple residual blocks to build a deep network, each layer in xLSTM can extract high-dimensional features from the previous layer while retaining the shallow input information, as follows:

[0203] No. Output of the layer

[0204]

[0205] is the input of this layer, represent The output after being processed by the module of this layer.

[0206] Inter-layer transfer:

[0207]

[0208] For the The input of the layer.

[0209] Step 404: Model training: The time series data in the training set is input into the network, and is gradually added to the training set through five independent predictions and taking the average value. The mean square error (MSE) is used as the loss function, and the model is repeatedly trained and the parameters are adjusted to ultimately minimize the loss function.

[0210] Loss Function The calculation formula is as follows

[0211]

[0212] in, is the true value, is the predicted value, Represents the total number of training samples. By adjusting the loss function, the optimal model is obtained.

[0213] Step 405: Model evaluation: By making predictions on the validation set, the root mean square error, mean square error, mean absolute error and coefficient of determination are calculated to evaluate the accuracy of the model prediction. The calculation formula of the model evaluation index is as follows:

[0214]

[0215]

[0216]

[0217]

[0218] in is the root mean square error, is the mean square error, is the mean absolute error, is the determination coefficient.

[0219] Step 900, yield prediction: input the current remote sensing image and the meteorological data of the future period (derived from the existing meteorological forecast system) into the yield prediction model, and output the yield prediction value of winter wheat in the target area. According to the results of multiple model evaluations, the yield prediction model is selected and combined with the results of principal component analysis to predict future yields.

[0220] In some embodiments of the present invention, a computer storage medium is provided, wherein the computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the processor executes the aforementioned image drawing method based on big data and artificial intelligence generation system technology.

[0221] The following content is an explanation of some professional terms in this embodiment.

[0222] The Enhanced Vegetation Index (EVI) is an improved vegetation index used to more accurately quantify the growth and health of vegetation. Based on the traditional Normalized Difference Vegetation Index (NDVI), EVI adds correction items for atmospheric and soil background effects, enhances sensitivity to areas with high vegetation coverage, and reduces the impact of air particles (such as aerosols) on the results.

[0223]

[0224] in, represents the reflectivity in the near-infrared band, represents the reflectivity of the red light band, Indicates the reflectivity of the blue light band.

[0225] The WOFOST crop growth model (World Food Studies) is a crop growth simulation model widely used in the agricultural field. It is used to simulate the growth, development and yield of crops under different environmental conditions. Based on the basic principles of plant physiology and ecology, it simulates the growth process of crops under different climate, soil and management conditions. The model estimates the dry matter accumulation and final yield of crops by simulating photosynthesis, respiration, water use and crop development stages.

[0226] Ensemble Kalman Filter (EnKF) is a data assimilation technique based on the Monte Carlo method, which is particularly effective in dealing with high-dimensional, nonlinear systems. It avoids the high computational cost of explicitly calculating the covariance matrix in the traditional Kalman filter method by using a set of samples (i.e., a collection) to represent the probability distribution of the system state.

[0227] The Extended Long Short-Term Memory (xLSTM) is an improved recurrent neural network (RNN) model that focuses on overcoming the limitations of traditional LSTM. By introducing exponential gating and extended memory structure, xLSTM can more efficiently capture short-term and long-term dependencies in time series and is widely used in time series modeling and regression prediction tasks.

[0228] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are within the protection of this embodiment.

Claims

1. A winter wheat yield prediction method based on XLSTM and assimilation crop growth model, characterized in that: The following steps are involved: Step 100, data acquisition and preprocessing: using a multispectral remote sensing satellite to acquire remote sensing image data of the winter wheat planting area of ​​the target area, and simultaneously acquiring meteorological data, soil data, yield data and crop parameter data; acquiring remote sensing image data of the winter wheat planting area of ​​the target area, selecting remote sensing images by season according to phenological information, performing radiation calibration, atmospheric correction and orthorectification, generating high-precision remote sensing image data, and cutting and mosaicking the remote sensing image data by vector boundaries; Step 200, winter wheat planting area extraction: visual interpretation to obtain winter wheat planting area image samples; calculate sample separability; use random forest classification model to supervise the remote sensing image of winter wheat planting area, realize feature layered extraction, and obtain winter wheat planting area vector; perform vector clipping on the remote sensing image of winter wheat planting area; confusion matrix analysis; Step 300, leaf area index inversion: calculate the enhanced vegetation index based on the pre-processed remote sensing image data of the winter wheat planting area; obtain the leaf area index by using the leaf area empirical inversion model in different seasons; Step 400, correlation analysis and feature extraction: combining the acquired meteorological data and yield data, performing classical correlation analysis and principal component analysis on the multi-dimensional feature data, calculating the correlation between different features and yield, and retaining the main features related to yield; Step 500, establishing a crop growth model based on the WOFOST model: combining the acquired meteorological data and soil data, referring to the crop parameter settings in the same region, establishing a crop growth model based on the WOFOST model, and outputting the LAI time series data of the entire growth period of winter wheat; Step 600, performing data assimilation based on ensemble Kalman filtering: Based on ensemble Kalman filtering, using the LAI time series data obtained from the crop growth model, data assimilation is performed on the LAI time series data obtained from remote sensing image processing of the winter wheat planting area to obtain accurate LAI time series data; Step 700, establishing a time series data set: sorting the LAI and meteorological principal component data corresponding to the pixels with corresponding station yield data in the remote sensing image of the winter wheat planting area in chronological order to form a time series input; Step 800, establishing a production prediction model based on an extended long short-term memory neural network: establishing an extended long short-term memory neural network as a production prediction model based on a time series data set and actual site production data; Step 900, yield prediction: input the remote sensing image of the current winter wheat planting area and the meteorological data of the future period into the yield prediction model, and output the yield prediction value of winter wheat in the target area.

2. The winter wheat yield prediction method based on XLSTM and assimilation crop growth model according to claim 1, characterized in that: The remote sensing image data comes from the Gaofen satellite, and the input features include four bands, namely blue light band, green light band, red light band and near-infrared band.

3. The winter wheat yield prediction method based on XLSTM and assimilation crop growth model according to claim 1, characterized in that: The calculation formula of the enhanced vegetation index is as follows: ； in, represents the enhanced vegetation index, NIR represents the reflectance of the near-infrared band, RED represents the reflectance of the red band, and BLUE represents the reflectance of the blue band.

4. The winter wheat yield prediction method based on XLSTM and assimilation crop growth model according to claim 1, characterized in that: The empirical inversion model of leaf area is as follows: Jointing period: ； The grouting period is: ； Among them, EVI stands for Enhanced Vegetation Index, is the leaf area index.

5. The winter wheat yield prediction method based on XLSTM and assimilation crop growth model according to claim 1, characterized in that: The features of the principal component analysis in step 400 include the daily average 2-meter-high temperature, the daily average precipitation, and the daily average surface net thermal radiation.

6. The winter wheat yield prediction method based on XLSTM and assimilation crop growth model according to claim 1, characterized in that: The specific steps of constructing the crop growth model include: Step 201, data collation: data cleaning is performed on the collected meteorological data, soil data, and crop physiological parameter data, and the data sets are collated and compiled into a YAML format; Step 202, initial parameter setting: setting the initial parameters of the WOFOST model according to the characteristics of the winter wheat variety, and setting the parameters of the meteorological, soil and crop management measures; Step 203, crop growth simulation: based on the WOFOST model, the crop growth process is simulated, by dividing the crop growth process into multiple stages, dynamically adjusting the growth rate and biomass accumulation of each stage according to the input meteorological data and soil data, and simulating and calculating the biomass accumulation of crops at different growth stages; Step 204, data output: obtain LAI time series data based on WOFOST model simulation; The WOFOST model is used as a crop growth model.

7. The winter wheat yield prediction method based on XLSTM and assimilation crop growth model according to claim 1, characterized in that: The specific steps of data assimilation in step 600 include: Step 301, initialization set: construct an initial set containing multiple samples, each sample represents a possible estimated value of LAI, and the initial set is composed of LAI time series data generated by the WOFOST model and LAI data obtained by remote sensing image processing as the starting state; Step 302, calculating the observation residual: comparing the LAI observation data obtained by remote sensing image processing with the LAI time series data obtained by model prediction, and calculating the observation residual; Step 303, Kalman gain calculation: Calculate the Kalman gain according to the residual covariance of the predicted value and the error covariance of the observed data; Step 304, update the prediction state: adjust the prediction value of each sample according to the Kalman gain, integrate the observed data into the model prediction, and thus update the prediction value of LAI; Step 305, updating the covariance matrix: calculating the updated covariance matrix to provide uncertainty measurement for the next prediction; Step 306, iterative update: repeat the prediction and update steps, and each time new remote sensing observation data enters, use the ensemble Kalman filter algorithm to adjust and optimize the LAI time series data until the error reaches the allowable value; Step 307, pixel-by-pixel processing: performing the data assimilation of the above steps on the LAI data of each pixel in the remote sensing image; Step 308, output data: output the assimilated accurate LAI time series data.

8. The winter wheat yield prediction method based on XLSTM and assimilation crop growth model according to claim 1, characterized in that: In step 700, a time series input is formed , ; in represents the time step, , , , Respectively represent the 1st, 2nd, t, The feature vector at time, represents the leaf area index at time t, Represents the principal component analysis data at time t.

9. The winter wheat yield prediction method based on XLSTM and assimilation crop growth model according to claim 1, characterized in that: The specific steps of establishing the production prediction model based on the extended long short-term memory neural network include: Step 401: Dataset division: The data set is divided into a training set and a validation set, where the training set accounts for 70% of the total data and the validation set accounts for 30%; Step 402: Module establishment: establish sLSTM and mLSTM modules, use sigmoid function as activation function, and adjust the number of neurons through multiple trainings; Step 403: residual stacking: connect the multi-layer sLSTM module and the mLSTM module through residual connection, combine the output and input of the xLSTM module, perform nonlinear processing and then feed back to the extended long short-term memory neural network; Step 404: Model training: The time series data in the training set is input into the extended long short-term memory neural network, and is gradually added to the training set through five independent predictions and the average value is taken. The mean square error is used as the loss function, and the model is repeatedly trained and the parameters are adjusted. Loss Function The calculation formula is as follows: ； in, is the true value, is the predicted value, Represents the total number of training samples; Step 405: Model evaluation: by making predictions on the validation set, calculating the root mean square error, mean square error, mean absolute error and coefficient of determination; The calculation formula of the model evaluation index is as follows: ； ； ； ； in is the root mean square error, is the mean square error, is the mean absolute error, is the determination coefficient.

10. A computer storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes a winter wheat yield prediction method based on XLSTM and assimilated crop growth model as described in any one of claims 1-9.

Images