Method and system for predicting non-grain spatial distribution of cultivated land

Abstract

The invention relates to the technical field of remote sensing land prediction, in particular to a farmland non-grain spatial distribution prediction method and system, and the method comprises the steps: abstracting a farmland plot into a graph node, defining an edge weight as a weighted combination of an inter-plot Euclidean distance, crop type similarity and irrigation system connectivity, constructing a three-layer dynamic graph structure comprising land parcels, neighborhoods and administrative units; taking the three-layer dynamic graph structure as input, and performing node feature extraction by adopting a graph attention long-short-term memory network to obtain a non-grain ratio preliminary predicted value; non-stationary state space-variational Kalman filtering is adopted to correct the preliminary prediction value of the non-grain ratio, and a plot-level prediction value is obtained; and performing spatial reconstruction on the plot-level predicted value by using policy-sensitive Kriging interpolation. According to the invention, spatial continuity and time serialization cultivated land non-grain trend prediction is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of remote sensing land prediction, in particular to a cultivated land non-grain spatial distribution prediction method and system. BACKGROUND

[0002] Cultivated land resources are facing multiple development pressures, and the non-grain trend is accelerating, that is, part of the cultivated land is converted to plant economic crops, fruits, energy crops or is in a fallow and idle state. The intensification of non-grain phenomenon not only affects food security, but also brings challenges to land use planning and ecological protection, and puts forward higher requirements for regulating agricultural structure and formulating land use policies.

[0003] The non-grain monitoring and its monitoring method are currently less studied. Traditional non-grain monitoring methods, such as land use map patch change comparison and transfer matrix analysis, are difficult to capture the nonlinear dynamic characteristics existing in the time series evolution. In recent years, the development of time series remote sensing data and the development of deep learning technology have provided new possibilities for agricultural land class evolution time series modeling. Among them, the long short-term memory network (LSTM) shows good ability in processing non-stationary time series, and the integrated graph attention network can replace the sequence input of the traditional bidirectional LSTM network, which is suitable for modeling the historical evolution and trend prediction of cultivated land use state. However, deep neural networks are sensitive to noise and have the risk of overfitting. The output is difficult to reflect the prediction uncertainty, resulting in poor result stability. SUMMARY

[0004] The embodiment of the application provides a cultivated land non-grain spatial distribution prediction method, which solves the problem of poor result stability caused by prediction uncertainty.

[0005] The embodiment of the application also provides a cultivated land non-grain spatial distribution prediction system.

[0006] The embodiment of the first aspect of the application provides a cultivated land non-grain spatial distribution prediction method, which comprises: The cultivated land plot is abstracted as a graph node, the edge weight is defined as the weighted combination of the Euclidean distance between plots, the crop type similarity and the irrigation system connectivity, and a three-layer dynamic graph structure including plots, neighborhoods and administrative units is constructed; Taking the three-layer dynamic graph structure as input, a graph attention long short-term memory network is used for node feature extraction to obtain a non-grain ratio preliminary prediction value; A non-stationary state space-varying Kalman filter is used to correct the non-grain ratio preliminary prediction value to obtain a plot-level prediction value; The plot-level prediction value is spatially reconstructed by using a policy-sensitive Kriging interpolation.

[0007] Further, the cultivated land plot is abstracted as a graph node, which comprises: Obtain crop distribution data of existing resolution; Obtain mask data of cultivated land plots; Upsample the crop distribution data to 30-meter resolution, align with the mask data pixel level; Extract non-grainization plots as graph nodes by the difference between cultivated land pixels and crop pixels.

[0008] Further, the node attributes of the graph nodes include non-grainization ratio, crop competition index and policy intervention intensity.

[0009] Further, taking the three-layer dynamic graph structure as input, the node feature extraction is performed by using the graph attention long short-term memory network to obtain the preliminary prediction value of the non-grainization ratio, including: The node feature propagates along the three-layer dynamic graph structure, and the neighborhood information is aggregated by using the multi-head attention weight, and the multi-head attention weight is dynamically generated by the crop competition index and the policy intervention intensity; The long-term spatio-temporal dependence is extracted through the forget gate, the input gate, the output gate and the cell state update, and a spatial constraint term is added in the hidden state update to force the non-grainization difference between adjacent plots to be less than the policy threshold; The graph attention long short-term memory network is connected to the fully connected layer and the Dropout layer at the end, and outputs the preliminary prediction value of the non-grainization ratio of each node at the next time.

[0010] Further, the non-stationary state space-varying Kalman filter is used to correct the preliminary prediction value of the non-grainization ratio to obtain the plot-level prediction value, including: The preliminary prediction value of the non-grainization ratio is decomposed into a three-dimensional state vector including a climate-driven component, an economic-driven component and a policy intervention component; The three-dimensional state vector is updated through a state transition matrix; The posterior distribution is approximated by using the variational Bayesian inference, and the process noise covariance is dynamically updated by maximizing the lower bound of the evidence; In the updating process, a policy correction term is introduced, and if the plot is located in the policy red line area, the observation noise covariance is forced to tend to zero.

[0011] Further, the plot-level prediction value is spatially reconstructed by using the policy-sensitive Kriging interpolation, including: The plot-level prediction value is combined with the night light data to correct the Kriging semi-variogram function; The prior variance is set to a very small positive number in the policy red line area; The policy-sensitive Kriging interpolation generates a continuous spatial surface by using the corrected Kriging semi-variogram function.

[0012] Further, the plot-level prediction value is combined with the night light data to correct the Kriging semi-variogram function, which is expressed by the formula as: ; wherein, denotes the modified Kriging semi-variogram function, denotes a forced modification value where the observation noise covariance tends to zero within the policy red line area, denotes the location of a plot belongs to the policy red line area , denotes the location of a plot belongs to the policy red line area , and is a constant, is the distance between two locations, is a decay coefficient used to control the Kriging semi-variogram function.

[0013] The second aspect embodiment of the present application provides a cultivated land non-grain spatial distribution prediction system, comprising: a graph structure construction module, configured to abstract the cultivated land plots into graph nodes, define the edge weight as a weighted combination of the Euclidean distance between plots, crop type similarity and irrigation system connectivity, and construct a three-layer dynamic graph structure including plots, neighborhoods and administrative units; a non-grain ratio preliminary prediction module, configured to take the three-layer dynamic graph structure as input, extract node features using a graph attention long short-term memory network, and obtain a non-grain ratio preliminary prediction value; a plot-level prediction module, configured to correct the non-grain ratio preliminary prediction value using a non-stationary state space-variance Kalman filter to obtain a plot-level prediction value; a reconstruction module, configured to perform spatial reconstruction on the plot-level prediction value using a policy-sensitive Kriging interpolation.

[0014] The embodiments of the present application have at least the following beneficial effects: The method of the present application can accurately depict nonlinear and non-stationary spatio-temporal variation processes, and realize spatial continuity, time series and cultivated land non-grain trend prediction. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 a flowchart of a cultivated land non-grain spatial distribution prediction method provided by the embodiments of the present application; Figure 2 a module block diagram of a cultivated land non-grain spatial distribution prediction system provided by the embodiments of the present application. DETAILED DESCRIPTION

[0016] ​​In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not used to limit the present application.

[0017] The cultivated land non-grain space distribution prediction method of the present application takes "graph-grid coupling" data fusion, non-stationary spatio-temporal modeling and policy-sensitive filtering as the core, and realizes the accurate description and scenario prediction of the evolution process of cultivated land non-grain through the cooperation of high-resolution remote sensing, dynamic graph attention long short-term memory network and non-stationary state space-variance Kalman filtering.

[0018] Referring to Figure 1 The cultivated land non-grain space distribution prediction method of the present application, comprising: S1 abstracts the cultivated land plot as a graph node, defines the edge weight as the weighted combination of the Euclidean distance between plots, the crop type similarity and the irrigation system connectivity, and constructs a three-layer dynamic graph structure including plots, neighborhoods and administrative units; S2 takes the three-layer dynamic graph structure as input, and uses graph attention long short-term memory network to extract node features to obtain a preliminary prediction value of the non-grain ratio; S3 uses non-stationary state space-variance Kalman filtering to correct the preliminary prediction value of the non-grain ratio to obtain a plot-level prediction value; S4 uses policy-sensitive Kriging interpolation to reconstruct the plot-level prediction value in space.

[0019] In step S1, in the data acquisition and preprocessing stage, in an embodiment, the cultivated land plot is abstracted as a graph node, including: Obtain crop distribution data of existing resolution; Obtain mask data of cultivated land plots; Upsample the crop distribution data to 30m resolution and align it with the mask data at the pixel level; Extract non-grain plots as graph nodes by the difference between cultivated land pixels and crop pixels.

[0020] For example, obtain 30m land use classification data for 2005, 2010, 2015, and 2020, and the corresponding 1 km main grain crop distribution layer; use weighted Gaussian kernel joint geographic weighted regression to upsample the crop distribution data to 30m resolution and align it with the cultivated land parcel mask data at the pixel level. Then, extract non-grain land blocks as graph nodes by calculating the difference between cultivated land pixels and crop pixels. A threshold can be set to determine which differences represent non-grain land blocks. Collect multi-source factors such as annual precipitation, annual average temperature, regional GDP, population density, elevation, nighttime light, and cultivated land subsidies, and protected area levels, and unify them to 30m resolution and complete node attribute alignment. All attributes are standardized by Z-Score: , where is the original value, , are the mean and standard deviation, respectively. Then, use Daubechies wavelet-soft threshold joint denoising to eliminate temporal noise and ensure graph node attribute consistency.

[0021] 30m resolution refers to the pixel size of a remote sensing image or geographic spatial data, i.e. each pixel represents a 30m x 30m area on the ground.

[0022] Cultivated land pixels refer to pixels (pixels) in remote sensing images that represent cultivated land. These pixels are usually identified through remote sensing image classification or land cover maps. Cultivated land refers to land used for planting crops, including farmland, orchards, vegetable gardens, etc. Crop pixels are more specific to pixels in remote sensing images that represent specific crops. These pixels are distinguished by crop recognition technology and can identify different crop types such as wheat, corn, rice, etc.

[0023] In an embodiment, the node attributes of the graph nodes include non-grain ratio, crop competition index, and policy intervention intensity. The non-grain ratio refers to the proportion of non-grain crop planting area to total planting area in a specific land block, which can be calculated by comparing the planting area of grain crops and non-grain crops identified in remote sensing images. The crop competition index is a quantitative indicator that measures the degree of competition between different crops in resource utilization (such as water resources, soil fertility, light, etc.), which can be calculated by analyzing the growth cycle, water requirement, soil nutrient demand, etc. of different crops. Policy intervention intensity refers to the strength or strictness of government policies implemented in a specific land block, such as cultivated land protection policies, agricultural subsidy policies, etc., which can be quantified by analyzing policy documents, regulations, subsidy amounts, etc.

[0024] The edge weights between plots can be defined first based on the Euclidean distance between the plot center points. This reflects the physical proximity between plots. The edge weights can also include crop type similarity, which can be calculated by comparing the crop types of adjacent plots. The higher the similarity, the greater the weight, indicating that agricultural activities between plots can be more similar. In embodiments of the present application, if plots are connected by an irrigation system, this can also be included as part of the edge weight. The higher the connectivity, the greater the weight, indicating that water resources sharing and distribution can be more closely linked. A three-layer dynamic graph structure is constructed including plots, neighborhoods, and administrative units, where the plot layer is the bottom layer, which is a single cultivated plot, and each plot is a node, and the connection between nodes is based on the edge weight defined above. The neighborhood layer, which is the middle layer, is a neighborhood composed of multiple plots with similar geographical locations. These neighborhoods can be defined based on specific geographical or socio-economic criteria, such as villages or farms. The administrative unit layer is the top layer, which is a larger administrative division, such as a township or county. This layer can be used to analyze broader land use patterns and policy impacts.

[0025] Over time, the use of plots, crop types, and irrigation systems may change. Therefore, the graph structure needs to be updated regularly to reflect these changes. Various analyses can be performed, such as shortest path analysis, clustering analysis, etc., to identify interactions and influences between plots. The graph structure can also be visualized using GIS software to help understand and interpret the data. In an embodiment, comprising: Node features propagate along the three-layer dynamic graph structure, and multi-head attention weights are used to aggregate neighborhood information, which are dynamically generated from crop competition indices and policy intervention strengths; Long-term spatio-temporal dependencies are extracted through forget gates, input gates, output gates, and cell state updates, and a spatial constraint term is added to the hidden state update to force the non-grain difference between adjacent plots to be no more than the policy threshold; The graph attention long short-term memory network is connected to a fully connected layer and a Dropout layer at the end, and outputs the preliminary prediction value of the non-grain ratio of each node at the next time.

[0026] The graph attention long short-term memory network is a two-layer graph attention long short-term memory network (GA-LSTM). Node features propagate along a "plot-neighborhood-administration" three-layer dynamic graph structure, and neighborhood information is aggregated using multi-head attention weights, which are dynamically generated from the crop competition index and policy intervention strength. The crop competition index measures the degree of competition between different crops in terms of resources such as water resources, soil fertility, and light. It can be calculated based on historical data, crop growth cycle, water demand, and soil nutrient demand. Policy intervention strength refers to the strength or strictness of government policies implemented on a specific plot, such as land protection policies, agricultural subsidy policies, etc. Policy intervention strength can be quantified by analyzing policy documents, regulations, and subsidy amounts. Based on the crop competition index and policy intervention strength, a correlation score is calculated for each pair of nodes in the graph. This score reflects the similarity of two plots in terms of crop planting and policy impact. Subsequently, long-term spatio-temporal dependencies are extracted through the forget gate, input gate, output gate, and cell state update, and a spatial constraint term is added to the hidden state update to force the non-grain difference between adjacent plots to be within the policy threshold. The policy threshold is based on historical data, agricultural research, and expert opinions to determine a reasonable non-grain difference threshold. This threshold is incorporated into relevant agricultural policies and land management regulations. The network is connected to a fully connected layer and a Dropout layer (p=0.3) at the end, which outputs the preliminary prediction of the non-grain ratio of each node at the next time step.

[0027] The spatial constraint term is represented as: , where is the policy constraint strength coefficient, is the maximum non-grain ratio difference threshold allowed by the policy, i.e., the policy threshold (such as 0.05), At time step , the hidden state of the th plot, represents the hyperbolic tangent function, represents the weight matrix of the hidden layer, represents the hidden state of the th plot at time step , represents the influence of the th plot on it, where is the neighbor set of the th plot, is the attention weight between the th plot and the th plot, is the weight matrix of the neighbor hidden layer, represents the calculation of the th plot and the land parcel at time step the difference between the non-grainification ratio and the policy threshold .

[0028] The multi-head attention weights are dynamically generated by the crop competition index and the policy intervention intensity: , where, represents the activation function, represents the learnable parameter variable, represents the transpose, represents the weight matrix of the hidden state of the land parcel at time step , represents the weight matrix of the hidden state of the land parcel at time step , represents the weight matrix of the hidden state of the land parcel at time step , represents the feature vector of the edge between land parcel and land parcel . represents the weight matrix of the hidden state of the land parcel at time step , represents the feature vector of the edge between land parcel and land parcel .

[0029] In an embodiment, the non-grainification ratio preliminary prediction value is corrected by using a non-stationary state space-Variational Kalman filter to obtain a land parcel level prediction value, including: decomposing the non-grainification ratio preliminary prediction value into a three-dimensional state vector including a climate driven component, an economic driven component and a policy intervention component; updating the three-dimensional state vector through a state transition matrix; dynamically updating the process noise covariance through evidence lower bound maximization by using variational Bayesian inference to approximate the posterior distribution; introducing a policy correction term in the updating process, if the land parcel is located in the policy red line area, then the observation noise covariance tends to zero.

[0030] The formula is expressed as: , represents the evidence lower bound (Evidence Lower Bound), which is the objective function in variational Bayesian inference, used to approximate the posterior distribution, represents the expectation under the approximate posterior distribution , represents the observation data given the state ​the likelihood function of denotes the Kullback-Leibler divergence between the posterior distribution and the prior distribution, which measures the difference between two distributions, denotes the policy regularization coefficient, which controls the strength of the policy regularization term. denotes the policy red line region, denotes the non-grain ratio observation value of the land parcel, denotes the non-grain ratio target value specified by the policy, which is usually the maximum non-grain ratio allowed by the policy. denotes the policy regularization term.

[0031] In the non-stationary state space-varying Kalman filter stage, the preliminary prediction value of the non-grain ratio is decomposed into three components: climate-driven component, economic-driven component, and policy intervention component, and a three-dimensional state vector is constructed; the state transition matrix is in block diagonal form, and the state of the climate-driven component, economic-driven component, and policy intervention component is updated respectively, and each sub-matrix is automatically learned by Granger causality test.

[0032] The variational Bayesian inference maximizes the evidence lower bound approximation of the posterior distribution, and adds a policy regularization term in the policy red line region, so that the observation noise covariance in the policy red line region is forced to zero, and robust estimation under policy disturbance is achieved. The policy red line region is the cultivated land protection region divided by the national spatial planning or other legal procedures.

[0033] In an embodiment, the land parcel level prediction value is spatially reconstructed by policy-sensitive Kriging interpolation, including: combining the land parcel level prediction value with the night light data to correct the Kriging semi-variogram function; setting the prior variance to a very small positive number in the policy red line region; the policy-sensitive Kriging interpolation generates a continuous spatial surface using the corrected Kriging semi-variogram function.

[0034] wherein the policy-sensitive Kriging interpolation refers to considering the impact of policy changes on spatial data during the interpolation process, and the Kriging semi-variogram function is also called semi-variance function. The semi-variogram function is usually calculated by the following steps: calculating the distance: determining the distance between all pairs of observation points in the study area. Calculate the difference: for each pair of observation points, calculate their observation value difference. Calculate the semi-variance: for each distance interval, calculate the average of the squares of the observation value differences of all point pairs, and divide by 2 to get the semi-variance.

[0035] In the spatial reconstruction stage, based on the filtered land parcel level prediction value, the semi-variogram function is corrected by combining the night light data: ; = , wherein, denotes the modified Kriging semi-variogram function, denotes a forced modification value where the observation noise covariance tends to zero within the policy red line area, denotes the location of a plot belongs to the policy red line area , denotes the location of a plot belongs to the policy red line area , and are constants, is the distance between two locations, is a decay coefficient used to control the Kriging semi-variogram function. With the policy-sensitive Kriging interpolation, the discrete predictions are reconstructed into 30 m continuous surfaces, and the zero variance is forced within the policy red line area. Finally, data in the format of GeoTIFF, GeoPackage, etc. can be derived.

[0036] On the other hand, referring to the module block diagram of the cultivated land non-grain spatial distribution prediction system shown in Figure 2 , a cultivated land non-grain spatial distribution prediction system can be understood with reference to the cultivated land non-grain spatial distribution prediction method described above. It includes: a graph structure construction module for abstracting cultivated land plots as graph nodes, defining the edge weight as a weighted combination of the Euclidean distance between plots, crop type similarity and irrigation system connectivity, and constructing a three-layer dynamic graph structure including plots, neighborhoods and administrative units; a non-grain ratio preliminary prediction module for taking the three-layer dynamic graph structure as input, using a graph attention long short-term memory network to extract node features, and obtaining a non-grain ratio preliminary prediction value; a plot-level prediction module for using a non-stationary state space-variational Kalman filter to modify the non-grain ratio preliminary prediction value to obtain a plot-level prediction value; a reconstruction module for using policy-sensitive Kriging interpolation to spatially reconstruct the plot-level prediction value.

[0037] The application constructs a three-layer dynamic graph structure, uses a graph attention long short-term memory network to replace the sequence input of a traditional bidirectional LSTM, and explicitly learns the spatial dependence between plots; a plot-level prediction module is designed, non-grain evolution is decomposed into a climate-driven component, an economic-driven component and a policy intervention component, the contribution of each component is dynamically estimated by a variational Kalman filter (VKF), and the defects of existing methods that cannot handle non-Gaussian policy disturbances are solved; through spatial reconstruction, the prior variance of Kriging interpolation is corrected by combining the vector boundary of the cultivated land protection zone and the night light data, so that the prediction result is forced to converge in the policy red line area.

[0038] Verification and application: Design spatio-temporal cross-validation: train with 2005+2010 graph data, predict 2015 distribution and compare with remote sensing inversion; train with 2005+2010+2015 graph data, predict 2020 distribution and compare with actual observation; calculate mean absolute error MAE and root mean square error RMSE, MAE≤0.06 and RMSE≤0.08, after verification, the prediction accuracy of the application method is high, suitable for the fusion scene of high-resolution remote sensing data and multi-source heterogeneous policy data, and can provide high-resolution decision support for regional agricultural policy making, land use optimization and food security guarantee.

[0039] The above only describes the preferred embodiment of the application, any modification, equivalent replacement and improvement based on the spirit of the application should be included in the protection scope of the application.

Claims

1. A method for predicting the spatial distribution of cultivated land for non-grain use, characterized in that: include: The cultivated land parcels are abstracted as graph nodes, and the edge weights are defined as a weighted combination of the Euclidean distance between parcels, the similarity of crop types, and the connectivity of irrigation systems. A three-layer dynamic graph structure consisting of parcels, neighborhoods, and administrative units is constructed. Taking a three-layer dynamic graph structure as input, a graph attention long short-term memory network is used to extract node features and obtain a preliminary prediction value of the non-grain ratio. The non-stationary state space-variational Kalman filter is used to correct the initial prediction value of the non-grain conversion ratio and obtain the plot-level prediction value. The plot-level predictions are spatially reconstructed using policy-sensitive kriging interpolation, including: The kriging semivariogram was modified by combining the plot-level predictions with the nighttime light data; Set the prior variance to a very small positive number within the policy red line area; Policy-sensitive kriging interpolation produces a continuous spatial surface using a modified kriging semivariogram.

2. The method for predicting the spatial distribution of cultivated land for non-grain use according to claim 1, characterized in that: The farmland plots are abstracted into graph nodes, including: Obtain crop distribution data at existing resolutions; Get the mask data of the cultivated land plot; Upsample the crop distribution data to 30-meter resolution and align it with the mask data pixel level; The difference between cultivated land pixels and crop pixels is used to extract non-grain land as graph nodes.

3. The method for predicting the spatial distribution of cultivated land for non-grain use according to claim 2, characterized in that: The node attributes of the graph nodes include non-grain ratio, crop competition index and policy intervention intensity.

4. The method for predicting the spatial distribution of cultivated land for non-grain use according to claim 1, characterized in that: Taking the three-layer dynamic graph structure as input, the graph attention long short-term memory network is used to extract node features, and the preliminary prediction value of the non-grain ratio is obtained, including: Node features are propagated along a three-layer dynamic graph structure, and neighborhood information is aggregated using multi-head attention weights, which are dynamically generated by the crop competition index and policy intervention intensity. Long-term spatiotemporal dependencies are extracted through forget gates, input gates, output gates, and cell state updates. Spatial constraints are added to the hidden state updates to force the difference in non-grainization of adjacent plots to not exceed the policy threshold. The end of the attention long short-term memory network is connected to the fully connected layer and the Dropout layer, and outputs the preliminary predicted value of the non-grain ratio of each node at the next moment.

5. The method for predicting the spatial distribution of cultivated land for non-grain use according to claim 1, characterized in that: The non-stationary state space-variational Kalman filter is used to correct the initial prediction value of the non-grain ratio and obtain the plot-level prediction value, including: The preliminary forecast value of the non-grain utilization ratio is decomposed into a three-dimensional state vector including climate driving component, economic driving component and policy intervention component. Update the three-dimensional state vector through the state transfer matrix; Variational Bayesian inference is used to approximate the posterior distribution, and the noise covariance of the dynamic update process is maximized through the evidence lower bound; A policy correction term is introduced in the updating process. If the land parcel is located in the policy red line area, the observation noise covariance is forced to approach zero.

6. The method for predicting the spatial distribution of cultivated land for non-grain use according to claim 1, characterized in that: The plot-level prediction value is combined with the night light data to modify the Kriging semivariogram, which is expressed as follows: ; = , ,in, represents the modified kriging semivariogram, Indicates the forced correction value that the observation noise covariance tends to zero within the policy red line area, Indicates the location of the plot Belongs to the policy red line area Inside, Indicates the location of the plot Belongs to the policy red line area Inside, and is a constant, is the distance between the two locations, The decay factor used to control the kriging semivariogram.

7. A spatial distribution prediction system for non-grain land conversion, characterized by: include: A graph structure construction module is used to abstract farmland plots into graph nodes, define edge weights as a weighted combination of the Euclidean distance between plots, crop type similarity, and irrigation system connectivity, and construct a three-layer dynamic graph structure consisting of plots, neighborhoods, and administrative units; The module for preliminary prediction of the non-grain ratio takes a three-layer dynamic graph structure as input and uses a graph attention long short-term memory network to extract node features and obtain a preliminary prediction value of the non-grain ratio. The plot-level prediction module is used to modify the initial prediction value of the non-grain conversion ratio using non-stationary state space-variational Kalman filtering to obtain the plot-level prediction value; The reconstruction module is used to spatially reconstruct the plot-level prediction values ​​using policy-sensitive kriging interpolation.

Images