Large-scale NDVI space-time prediction method considering multi-factor constraint

Abstract

The invention discloses a large-scale NDVI spatio-temporal prediction method considering multi-factor constraints. The method comprises the following steps: acquiring NDVI, meteorological data, earth surface parameters and other related variables from a plurality of data sources, and performing preprocessing to generate a spatio-temporal data set in a unified format; dividing the large-scale region into a plurality of sub-regions with similar spatial distribution characteristics by using a Voronoi diagram; for each sub-region, extracting feature information of NDVI and key variables, and constructing a space-time cube of each space-time point; flattening the space-time cube, and inputting the space-time cube into a trained multivariable LSTM model for space-time prediction; and performing spatial interpolation on a prediction result by using an ANUSPLIN interpolation method, and expanding the prediction result to the whole research area to generate a large-scale NDVI space-time prediction map. Through fusion of various technical means and comprehensive consideration of cross constraint effects of various environmental factors, accurate prediction of large-scale regional NDVI spatio-temporal changes is realized. The method has wide application prospects and great significance in ecological environment monitoring, carbon cycle evaluation and natural resource management.

Description

Technical Field

[0001] The present invention relates to the field of monitoring changes in the Normalized Difference Vegetation Index, and is related to, but not limited to, a large-scale NDVI spatiotemporal prediction method that considers multi-factor constraints. Background Art

[0002] In ecological environment monitoring and natural resource management, the Normalized Difference Vegetation Index (NDVI) is an important vegetation index indicator, which is widely used in vegetation cover change, carbon cycle assessment, drought monitoring and ecosystem health assessment. The spatiotemporal variation of NDVI not only reflects the growth status of surface vegetation, but is also closely related to environmental factors such as climate, soil conditions, and human activities. Therefore, accurately predicting the spatiotemporal variation of NDVI in large-scale regions is of great significance for understanding ecosystem dynamics, assessing carbon sink status, and formulating sustainable management strategies.

[0003] Existing NDVI prediction methods face many challenges when dealing with large-scale areas. First, large-scale areas usually have complex topography, diverse vegetation types, and changeable climatic conditions. The interaction between these factors makes the spatiotemporal variation of NDVI highly complex and uncertain. Secondly, existing prediction methods often only focus on the temporal characteristics of NDVI, ignoring its potential spatiotemporal connection with various environmental factors, resulting in limited accuracy and generalization of prediction results. Finally, due to the difficulty of data acquisition and processing, NDVI data in large-scale areas often have problems such as low temporal resolution and discontinuous spatial coverage, which further increases the difficulty of prediction.

[0004] In order to overcome these challenges, researchers have begun to explore NDVI prediction methods that integrate multi-source data and multi-technical means in recent years. Among them, Long Short-Term Memory (LSTM), as a special recurrent neural network, is widely used in NDVI prediction because it can effectively handle long-term dependencies in time series data. However, a single LSTM model is still insufficient when dealing with large-scale areas, and it needs to be combined with other technical means to further improve the prediction accuracy and generalization ability. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides a large-scale NDVI spatiotemporal prediction method considering multi-factor constraints, which integrates a variety of technical means such as Voronoi diagram, LSTM neural network and ANUSPLIN, aiming to achieve accurate prediction of the spatiotemporal changes of NDVI in large-scale areas by comprehensively considering the cross-constraints of various environmental factors on NDVI.

[0006] The technical solution of the embodiment of the present invention is achieved as follows:

[0007] An embodiment of the present invention provides a large-scale NDVI spatiotemporal prediction method considering multiple factor constraints, the method comprising:

[0008] NDVI, meteorological data, surface parameters and other related variables are obtained from multiple data sources and preprocessed to generate a spatiotemporal data set in a unified format; the Voronoi diagram is used to divide the large-scale area into multiple sub-areas with similar spatial distribution characteristics; the mean of each sub-area represents the value of the corresponding spatiotemporal point; for each sub-area, the characteristic information of NDVI and key variables is extracted to construct a spatiotemporal cube for each spatiotemporal point; wherein the key variables are environmental variables and vegetation variables related to NDVI; the spatiotemporal cube is flattened and input into a trained multivariate LSTM model for spatiotemporal prediction; wherein the multivariate LSTM model includes cascaded multi-layer bidirectional LSTM units, each of which includes a cell state, a forget gate, an input gate, and an output gate; the prediction results are spatially interpolated using the ANUSPLIN interpolation method and extended to the entire study area to generate a large-scale NDVI spatiotemporal prediction map.

[0009] In some possible embodiments, the extracting characteristic information of NDVI and key variables for each sub-region includes: using a geographic detector and a Gini index method to screen key variables related to NDVI; using mean data in each sub-region to extract characteristic information of NDVI and the key variables; the characteristic information includes statistical features in time series data and geometric features in spatial distribution.

[0010] In some possible embodiments, the cell state is used to update the information at the current time t, the forget gate determines how much information to discard or retain from time t-1, the input gate determines whether to update the input information at time t to the cell state, and the output gate controls how the cell state is output.

[0011] In some possible embodiments, in the process of constructing the multivariate LSTM model, a random dropout layer and a fully connected layer are added after the last layer of LSTM units in the cascade to prevent the model from overfitting and realize nonlinear feature mapping.

[0012] In some possible embodiments, the method further includes: training the multivariate LSTM model using the preprocessed data, performing parameter tuning operations through GridSearchCV in scikit-learn to improve prediction performance, specifying a range of candidate hyperparameter values ​​and traversing combinations of these parameters to implement a parameter tuning process.

[0013] In some possible embodiments, the method further includes: evaluating the performance of the multivariate LSTM model using the following multiple evaluation indicators: determination coefficient, adjusted determination coefficient, root mean square error, mean absolute percentage error, explained variance, and mean directional accuracy.

[0014] The beneficial effects brought by the technical solution provided by the embodiment of the present application include at least:

[0015] In an embodiment of the present invention, by integrating remote sensing technology, GIS spatial analysis technology and deep learning algorithms, comprehensive consideration and accurate modeling of various factors affecting NDVI changes are achieved. On the one hand, by integrating multi-factor feature information, the potential spatiotemporal connection between NDVI and various factors is fully considered to improve the prediction accuracy. On the other hand, through the combination of Voronoi diagram and LSTM neural network, large-scale NDVI spatiotemporal prediction from point to surface is achieved, and spatial generalization ability is enhanced. The present invention helps to more accurately evaluate the regional carbon sink status, reveal the dynamic change process of vegetation coverage, and provide scientific basis and decision-making support for ecological environmental protection, climate change response, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0017] Figure 1 A schematic diagram of a flow chart of a large-scale NDVI spatiotemporal prediction method considering multiple factor constraints provided by an embodiment of the present invention;

[0018] Figure 2 A schematic diagram of a process for generating a Voronoi diagram from spatiotemporal point data provided by an embodiment of the present invention;

[0019] Figure 3 The logical framework diagram of the large-scale NDVI spatiotemporal prediction method considering multi-factor constraints provided by the embodiment of the present invention, wherein part (a) is the multi-factor feature information extraction process, part (b) is the multivariate LSTM neural network model architecture and prediction process, and part (c) is the schematic process of NDVI prediction value interpolation;

[0020] Figure 4 A schematic diagram of the input and output process of the multivariate LSTM model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0022] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0023] It should be pointed out that the terms "first\second\third" involved in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described here can be implemented in an order other than that illustrated or described here.

[0024] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the field to which the embodiments of the present invention belong. It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with those in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless specifically defined as here.

[0025] Voronoi cell is a method of dividing space into multiple regions so that all points in each region are closest to a generating point (or seed point). This division is often used in computational geometry, scientific computing, and spatial analysis.

[0026] In Chinese, Voronoi cell can be called "Voronoi unit" or "subregion of Voronoi diagram". All points in each Voronoi subregion are closer to the generating point of the subregion than to any other generating point. This division can be used to represent the neighboring area of ​​each point in space, or to analyze and process spatial data.

[0027] The generation of Voronoi cells usually follows the following steps: Select a set of generating points that can represent certain features or data points in space. For each generating point, determine its neighborhood, that is, the set of all spatial points that are closer to this point than other generating points. The boundaries of these neighborhoods, that is, the dividing lines between different sub-regions, are usually formed by the minimum distance line segments or curves between the generating points.

[0028] The Voronoi diagram has a wide range of applications in geographic information systems (GIS), material science, biodiversity research, and other fields. It can be used to analyze spatial distribution, optimize facility layout, and simulate natural phenomena.

[0029] LSTM neural network: Long short-term memory network (LSTM) has been widely studied and applied due to its advantages in processing sequence data. LSTM solves the gradient vanishing problem of traditional RNN through the gating mechanism and can capture long-term dependencies, which makes it perform well in areas such as time series prediction.

[0030] Voronoi diagram: Voronoi diagram is a method to divide space into multiple regions, and the distance between the points in each region and the corresponding generating point (seed point) is the shortest. It has a wide range of applications in computational geometry, scientific computing, and spatial analysis.

[0031] ANUSPLIN interpolation method: ANUSPLIN is an interpolation method based on data spatial autocorrelation and variability. It can interpolate discrete points into a smooth continuous surface. It is suitable for processing long time series data and allows the introduction of linear covariate submodels.

[0032] Figure 1 A schematic diagram of a large-scale NDVI spatiotemporal prediction method considering multiple constraints provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method comprises at least the following steps:

[0033] Step S110, obtaining NDVI, meteorological data, surface parameters and other related variables from multiple data sources and preprocessing them to generate a spatiotemporal dataset in a unified format.

[0034] Here, the multiple data sources include but are not limited to LADDS (Land Data Assimilation System), LOESS (Land Surface and Ocean Energy and Water Cycle Studies), TPDC (The Tibetan Plateau Data Center) and GLASS (Global Land Surface Satellite), etc. The data time range covers 2000 to 2021, and the spatial resolution varies according to different data sources. The meteorological data include average temperature MAT, precipitation MCP, surface temperature LST, etc., and the surface parameters include potential evapotranspiration ET0, albedo Albedo, leaf area index LAI, photosynthetically active radiation absorption ratio FAPAR, etc.

[0035] Preprocessing operations include geometric correction, projection stitching, resampling, maximum synthesis, etc. to reduce cloud contamination and soil background effects. Through preprocessing, the differences in spatial and temporal resolutions between different data sources are unified into a form suitable for model input.

[0036] Step S120: using the Voronoi diagram to divide the large-scale region into a plurality of sub-regions having similar spatial distribution characteristics.

[0037] Here, the mean of each sub-region represents the value of the corresponding time-space point. In order to effectively extract the characteristic information of NDVI and its influencing factors in a large-scale area, the embodiment of the present invention uses the Voronoi diagram, that is, the Voronoi diagram, to divide the study area into multiple sub-regions (Voronoi cells) with similar spatial distribution and containing characteristic information. Each Voronoi cell contains specific time-space information to capture the local variation characteristics of NDVI. Each Voronoi cell consists of its generating point and the nearest points around it, ensuring that the data contained is similar and representative.

[0038] The present invention utilizes the Voronoi diagram to divide the region and extract characteristic information, fully considers the cross-constraint effect of different environmental factors on NDVI, and improves the adaptability of the prediction model to complex ecosystems.

[0039] Step S130: for each sub-region, extract the characteristic information of NDVI and key variables, and construct a space-time cube for each space-time point.

[0040] Here, the key variables are environmental variables and vegetation variables related to NDVI; the feature information includes statistical features in time series data and geometric features in spatial distribution. These feature information are constructed into a space-time cube and flattened as input data for the neural network.

[0041] Step S140, flattening the space-time cube and inputting it into the trained multivariate LSTM model for space-time prediction.

[0042] Here, the multivariate LSTM model includes cascaded multi-layer bidirectional LSTM units, each of which includes a cell state, a forget gate, an input gate, and an output gate; the model can capture complex time dependencies and long-term trends in the data. Using multiple environmental factors as inputs to the multivariate LSTM model for spatiotemporal prediction effectively handles the long-term dependencies and spatial heterogeneity in the NDVI time series data. Before prediction, the multivariate LSTM model is trained using preprocessed multi-factor feature data.

[0043] In the implementation, the neural network API Keras in Python is used to build a multivariate LSTM neural network model. Multiple LSTM cells are stacked to form an LSTM structure, the input layer and the output layer are constructed, and the GPU is called to accelerate the model training. This model gradually integrates the impact of multiple factors on the target variable through the LSTM cell cascade structure, providing more accurate and comprehensive prediction results. Moreover, the model can automatically extract useful features from the input data through training. The specific information of each layer is as follows:

[0044] The input layer receives the NDVI and multi-factor feature information extracted from the small-scale area after the Voronoi diagram is divided. These feature information are constructed into a spatiotemporal cube and flattened as the input of the neural network. Unlike the traditional LSTM that only processes single time series data, the input of the present invention is spatiotemporal data that integrates multivariate feature information, which enhances the model's ability to handle complex spatiotemporal relationships.

[0045] Each layer of LSTM cell contains cell state, forget gate, input gate and output gate. The gating mechanism controls the flow of information, avoids gradient explosion and gradient diffusion problems, and effectively captures long-term dependencies in the data. Compared with single-layer LSTM, multi-layer structures can better learn the abstract representation of data and capture feature information at different levels. Each layer of LSTM not only processes the time series data of NDVI, but also processes multiple variables that affect NDVI at the same time, enhancing the model's multi-factor feature extraction capabilities.

[0046] The output layer receives the hidden state output by the LSTM layer and converts it into a predicted value through a fully connected layer. The predicted value includes not only the numerical value of NDVI, but also the spatial distribution information related to it.

[0047] Step S150, use the ANUSPLIN interpolation method to perform spatial interpolation on the prediction results, and expand them to the entire study area to generate a large-scale NDVI spatiotemporal prediction map.

[0048] Here, in addition to the basic NDVI numerical prediction, the multivariate LSTM model also combines spatial information, and uses the ANUSPLIN interpolation method to spatially interpolate the NDVI prediction values ​​output by the LSTM model to fill the blank areas between the Voronoi cells, realize the NDVI prediction from point to surface, and realize the comprehensive prediction of the spatiotemporal changes of NDVI in large-scale regions. Spatial information fusion: Through ANUSPLIN interpolation, the model converts discrete point prediction values ​​into smooth continuous surfaces, further improving the spatial continuity and accuracy of the prediction results.

[0049] The present invention first divides the large-scale planar area into multiple Voronoi cells, which can avoid unnecessary redundant information and enhance the ability of the multivariate LSTM neural network to capture trends from smaller-scale details to larger-scale trends. Secondly, the feature information of NDVI and its high-importance variables is extracted for each Voronoi cell. Then, a multivariate LSTM neural network model is constructed, and the extracted feature information is used as input for spatiotemporal prediction. Finally, interpolation methods such as ANUSPLIN are used to expand the prediction results from points and surfaces to the entire area, thereby realizing a large-scale NDVI spatiotemporal prediction method that considers multi-factor constraints, which to a certain extent explains the complex time dependence and spatial heterogeneity in NDVI prediction.

[0050] In some possible embodiments, the extracting characteristic information of NDVI and key variables for each sub-region includes: using a geographic detector and a Gini index method to screen key variables related to NDVI; using mean data in each sub-region to extract characteristic information of NDVI and the key variables; the characteristic information includes statistical features in time series data and geometric features in spatial distribution.

[0051] Here, the key variables are variables that are highly important to NDVI, including statistical characteristics such as mean, standard deviation, maximum value, minimum value, and change trend in time series data, as well as geometric characteristics such as texture and shape in spatial distribution.

[0052] The above embodiment introduces methods such as geographic detectors and Gini index to screen key variables, ensuring that the selected variables can fully explain the spatiotemporal variation of NDVI, thereby reducing redundant information and reducing the complexity and uncertainty of the prediction results.

[0053] In some possible embodiments, the cell state is used to update the information at the current time t, the forget gate determines how much information to discard or retain from time t-1, the input gate determines whether to update the input information at time t to the cell state, and the output gate controls how the cell state is output.

[0054] Here, each LSTM unit contains a cell state C t , the three neural network layers are respectively the forget gate f t , input gate i t and output gate o t , the calculation process of LSTM at time t is as follows:

[0055] Deciding to discard information:

[0056] f t =σ(W f ·[h t-1 ,x t ]+b f ) formula (1);

[0057] Confirm update information:

[0058] i t =σ(W f ·[h t-1 ,x t ]+b i ) formula (2);

[0059]

[0060] Cell Status:

[0061]

[0062] Output information:

[0063] o t =σ(W f ·[h t-1 ,x t ]+b o ) formula (5);

[0064] h t =o t tanh(C t ) formula (6);

[0065] In the formula, the symbol '·' represents the dot product operation of two vectors; x t represents the input vector; h t-1 、h t Respectively represent the hidden state at time t-1 and the hidden state at time t; represents the candidate cell state at time t; C t-1 , C t Respectively represent the cell state at time t-1 and the cell state at time t; W f , W i , W o , Wc Represent the weight vectors of the forget gate, input gate, output gate and cell state respectively; b f 、b i 、b o 、b c They represent the bias vector of forget gate, input gate, output gate and cell state respectively; σ and tanh are activation functions.

[0066] In some possible embodiments, in the process of constructing the multivariate LSTM model, a random dropout layer and a fully connected layer are added after the last layer of LSTM units in the cascade to prevent the model from overfitting and realize nonlinear feature mapping.

[0067] In some possible embodiments, the method further includes: training the multivariate LSTM model using the preprocessed data, performing parameter tuning operations through GridSearchCV in scikit-learn to improve prediction performance, specifying a range of candidate hyperparameter values ​​and traversing combinations of these parameters to implement a parameter tuning process.

[0068] Here, during the model training process, the input sequence is divided into a training set and a test set, and GridSearchCV is used to tune the parameters, hyperparameters such as learning rate, random inactivation rate, iteration cycle, batch size, etc., and 2-fold cross validation is used to verify the reliability of the parameters.

[0069] In some possible embodiments, the method further includes: evaluating the performance of the multivariate LSTM model using the following multiple evaluation indicators: determination coefficient, adjusted determination coefficient, root mean square error, mean absolute percentage error, explained variance, and model determination coefficient.

[0070] Here, the coefficient of determination R is calculated 2 , root mean square error RMSE, mean absolute percentage error MAPE and other evaluation indicators are used to evaluate the accuracy and reliability of the model prediction results. The stability and generalization ability of the model are verified by calculating indicators such as PopulationStability Index.

[0071] The large-scale NDVI spatiotemporal prediction method considering multiple constraints is described below in conjunction with a specific embodiment. However, it is worth noting that this specific embodiment is only for better illustrating the present invention and does not constitute an improper limitation to the present invention.

[0072] An embodiment of the present invention provides a large-scale NDVI spatiotemporal prediction method considering multi-factor constraints, which consists of multi-factor feature information extraction, multivariate LSTM neural network prediction and NDVI prediction value interpolation. The method takes environmental variables and vegetation variables of high importance to NDVI as input, and outputs the future spatial distribution of NDVI. First, the feature information of each input variable is extracted by Voronoi diagram partitioning, and then it is introduced into a three-layer multivariate LSTM neural network. Finally, the NDVI prediction value is output and expanded to the entire study area in combination with ANUSPLIN. The specific implementation process includes several major steps including data preparation, feature extraction, model construction, prediction and result interpolation, which are described as follows:

[0073] S1. Data preparation

[0074] S11, obtain data source: In order to make NDVI spatiotemporal prediction, it is necessary to collect a variety of relevant data, including NDVI data, meteorological data (such as average temperature MAT, precipitation MCP, surface temperature LST, etc.), ecological data (such as potential evapotranspiration ET0, albedo Albedo, leaf area index LAI, photosynthetically active radiation absorption ratio FAPAR, etc.). The data sources mainly include LADDS, LOESS, TPDC and GLASS, etc. The data time range covers 2000 to 2021, and the spatial resolution varies according to different data sources.

[0075] S12 Data preprocessing: In the data preprocessing stage, data cleaning, interpolation, resampling and other operations are required to ensure data integrity and consistency. In particular, it is necessary to deal with the differences in spatial resolution and temporal resolution between different data sources and unify them into a form suitable for model input. For missing values, neighboring point interpolation or time series interpolation methods are used to fill them.

[0076] S2. Feature extraction

[0077] S21 Voronoi diagram division: The present invention uses the Voronoi diagram to divide the large-scale surface area into Voronoi cells with similar spatial distribution and containing feature information. It is assumed that the mean of the Voronoi cell can approximately represent the value of the corresponding space-time point. Each Voronoi cell consists of its generating point and the nearest points around it, ensuring that the data contained is similar and representative. The Voronoi cell in the embodiment of the present invention is generated by 1000 space-time points that conform to the distribution of meteorological stations, such as Figure 2 shown.

[0078] Dividing a large-scale surface area into many sub-areas can avoid unnecessary redundant information and enhance the ability of the multivariate LSTM neural network to capture trends from smaller-scale details to larger-scale trends. For each spatiotemporal point in the Voronoi cell, the distance to the nearest spatiotemporal point is calculated and the weight value is calculated based on the distance. These weight values ​​are used to adjust the data value of the sample point to enhance or weaken the influence of the nearby Voronoi diagram. Especially in areas where the Voronoi cell area is small and the spatiotemporal points are densely distributed, the spatiotemporal changes of NDVI can be better reflected. In addition, all variables need to construct feature vectors that conform to model learning training before entering the multivariate LSTM model to help the model better predict the target variable data.

[0079] S22 Multi-factor feature information extraction: In each Voronoi cell, extract the feature information of NDVI and its high-importance variables, including statistical features such as mean, standard deviation, maximum, minimum, and change trend in time series data, as well as geometric features such as texture and shape in spatial distribution. Figure 3 As shown in part (a), the Voronoi diagram divides the large-scale area into small-scale areas, extracts NDVI and multi-factor feature information to construct specific features, and then constructs the feature information in each Voronoi cell into a corresponding space-time cube (also called a data cube). After flattening, the feature vector that meets the model learning training is obtained. The vector x represents the matrix composed of multiple factors and the number of samples at a certain moment.

[0080] S3. Model construction

[0081] S31 Multivariate LSTM Neural Network Model: The present invention introduces an LSTM neural network model, which can better capture complex time features from feature information. Compared with other models, LSTM can better solve the problems of gradient explosion and gradient diffusion, and has good stability and performance when processing long sequence data. LSTM can be trained to identify features in input data for classification, prediction or decision-making. LSTM avoids the problems of gradient explosion and gradient diffusion by introducing memory units. Memory units can store information from previous time steps and pass it to the current time step to better capture long-term dependencies when processing long sequence data. In addition, LSTM also introduces a gating mechanism to control the flow of information.

[0082] The present invention uses the neural network API Keras in Python to build a multivariate LSTM neural network model. Multiple LSTM units are stacked to form an LSTM structure, while the input layer and output layer are constructed, and the GPU is called to accelerate the model training. Figure 3Part (b) shows a three-layer multivariate LSTM neural network model architecture, which gradually integrates the impact of multiple factors on the target variable by cascading multiple layers of bidirectional LSTM units to provide more accurate and comprehensive prediction results. Figure 3 In the example, after LSTM3, dropout and dense layers are added to prevent the model from overfitting and realize nonlinear feature mapping. This model gradually integrates the influence of multiple factors on the target variable through the LSTM cell cascade structure, providing more accurate and comprehensive prediction results. Moreover, the multivariate LSTM model can automatically extract useful features from the input data through training. Since the multivariate LSTM model provided by the embodiment of the present invention has a parallel structure, it can process multiple input sequences at the same time and share weights and biases between memory cells corresponding to different variables. This means that the model can not only learn the common information and patterns of the target variable from multiple factors, but also retain the information unique to each multifactor.

[0083] S32 Model training and verification: The specific process of input and output of the multivariate LSTM model provided by the embodiment of the present invention is as follows Figure 4 As shown, before training the multivariate LSTM model, each input sequence is divided into two parts, of which the first 80% is used as a training set and the last 20% is used as a test set. At the same time, the time step (TimeStep) is set to 3. Since it is necessary to predict the feature information represented by each Voronoi cell in multiple cycles, the parameter tuning operation is performed through GridSearchCV in scikit-learn to improve the prediction performance. The embodiment of the present invention specifies the range of candidate values ​​of hyperparameters and traverses the combination of these parameters to implement the parameter tuning process. Moreover, each Voronoi cell uses a common loss function to optimize the prediction results of the variables. In addition, 2-fold cross validation (CV) is also used to verify the reliability of the optimal parameters. The most common parameter combinations in this model are shown in Table 1.

[0084] Table 1 The most common parameter combinations in multivariate LSTM neural network models

[0085]

[0086] S4. Prediction and result interpolation: The test set data is input into the trained multivariate LSTM neural network model to obtain the NDVI prediction value of each Voronoi cell. These prediction values ​​reflect the spatiotemporal variation characteristics of NDVI under given multi-factor constraints. In order to extend the prediction results at the Voronoi cell level to the entire study area, the spatial interpolation method ANUSPLIN is used to interpolate the prediction values. The interpolation results will present a continuous spatiotemporal distribution map of NDVI, which is convenient for subsequent analysis and application.

[0087] The present invention reclassifies and encodes the NDVI prediction value in combination with spatial information, and uses the ANUSPLIN interpolation method to expand it to the regional level with a resolution of 1000 meters, such as Figure 3 As shown in part (c) of the figure. ANUSPLIN models the spatial autocorrelation and variability of the data and interpolates discrete points into a smooth continuous surface. In addition to spline independent variables, ANUSPLIN allows the introduction of linear covariate submodels. The theoretical statistical model is:

[0088] Z i =f(x i )+by i +e i (i=1,2,3,...,n) Formula (7);

[0089] In the formula, Z i is the dependent variable at point i in space; x i is the independent variable; i is the independent covariate of P dimension; e i is a random error of the independent variable with an expected value of 0.

[0090] Although ANUSPLIN's smoothing of data may result in the loss of local details, it can reduce the imbalance problem generated during feature extraction and neural network training. Compared with other interpolation methods, ANUSPLIN is more suitable for processing long time series data and can perform interpolation of multiple surface spaces at the same time. In order to ensure the accuracy of most seasonal interpolation results, the present invention adopts a quadratic local thin disk smooth spline function model. The independent variables of this model are longitude and latitude, and the covariates are variables with an importance to NDVI greater than 90%.

[0091] 5. Performance evaluation: In order to verify the performance of the CMFC-LSTM method, a variety of evaluation indicators are used to evaluate the prediction results, including the determination coefficient R 2 , Adjusted coefficient of determination Adj.R 2 , root mean square error RMSE, mean absolute percentage error MAPE, explained variance EVS and mean directional accuracy MDA, etc. At the same time, this method is compared with other mainstream prediction methods (such as ConvLSTM, random forest RF, etc.) to prove the superiority of the present invention.

[0092] In summary, the large-scale NDVI spatiotemporal prediction method considering multiple constraints proposed in this invention effectively improves the accuracy and generalization ability of large-scale regional NDVI spatiotemporal prediction by integrating multiple technologies such as Voronoi diagram, LSTM neural network and ANUSPLIN. This method has achieved remarkable results in practical applications in the Yellow River Basin and has broad application prospects and promotion value. In the future, the model structure will be further optimized, more influencing factors will be introduced, and the application field will be expanded to provide more reliable technical support for regional ecological environment monitoring and carbon sink assessment.

[0093] Compared with the prior art, the large-scale NDVI spatiotemporal prediction method considering multi-factor constraints provided by the embodiment of the present invention has at least the following advantages: Improving prediction accuracy: integrating multi-factor feature information, fully considering the potential spatiotemporal connection between NDVI and multiple factors, and improving prediction accuracy. Enhanced generalization ability: through the combination of Voronoi diagram and LSTM neural network, large-scale NDVI spatiotemporal prediction from point to surface is realized, and spatial generalization ability is enhanced. Reduce complexity and uncertainty: Introduce geographic detectors to screen high-importance variables, reduce redundant information, and reduce the complexity and uncertainty of prediction results. Provide technical support: Provide key technical support for large-scale regional carbon sink assessment, which will help achieve the "carbon neutrality" goal.

[0094] It should be understood that "one embodiment" or "an embodiment" mentioned throughout the specification means that specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner. It should be understood that in various embodiments of the present invention, the size of the serial number of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention. The serial numbers of the above-mentioned embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments.

[0095] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0096] In the several embodiments provided by the present invention, it should be understood that the disclosed methods can be implemented in other ways. The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments. The features disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0097] The above is only an embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A large-scale NDVI spatiotemporal prediction method considering multiple constraints, It is characterized in that include: Obtain the Normalized Difference Vegetation Index (NDVI), meteorological data, surface parameters, and other related variables from multiple data sources and perform preprocessing to generate a spatiotemporal dataset in a unified format; The Voronoi diagram is used to divide a large-scale area into multiple sub-areas with similar spatial distribution characteristics; the mean value of each sub-area represents the value of the corresponding time and space point; For each sub-region, the characteristic information of NDVI and key variables is extracted to construct a space-time cube for each time-space point; wherein the key variables are environmental variables and vegetation variables related to NDVI; Flattening the space-time cube and inputting it into a trained multivariate LSTM model for space-time prediction; wherein the multivariate LSTM model includes cascaded multi-layer bidirectional LSTM units, each of which includes a cell state, a forget gate, an input gate, and an output gate; The prediction results were spatially interpolated using the ANUSPLIN interpolation method and extended to the entire study area to generate a large-scale NDVI spatiotemporal prediction map.

2. The method according to claim 1, It is characterized in that For each sub-region, the characteristic information of NDVI and key variables is extracted, including: The key variables related to NDVI were screened using the GeoDetector and Gini Index methods; The mean data in each sub-region is used to extract the characteristic information of NDVI and the key variables; the characteristic information includes the statistical characteristics in the time series data and the geometric characteristics in the spatial distribution.

3. The method according to claim 1, It is characterized in that The cell state is used to update the information at the current time t, the forget gate determines how much information to discard or retain from time t-1, the input gate determines whether to update the input information at time t to the cell state, and the output gate controls how the cell state is output.

4. The method according to claim 1, It is characterized in that In the process of constructing the multivariate LSTM model, a random dropout layer and a fully connected layer are added after the last layer of LSTM units in the cascade to prevent the model from overfitting and realize nonlinear feature mapping.

5. The method according to any one of claims 1 to 4, It is characterized in that The method further comprises: The multivariate LSTM model is trained using the preprocessed data, and parameter tuning is performed through GridSearchCV in scikit-learn to improve the prediction performance. The range of candidate hyperparameter values ​​is specified and the combinations of these parameters are traversed to achieve the parameter tuning process.

6. The method according to any one of claims 1 to 4, It is characterized in that The method further comprises: The performance of the multivariate LSTM model was evaluated using the following evaluation metrics: coefficient of determination, adjusted coefficient of determination, root mean square error, mean absolute percentage error, explained variance, and mean directional accuracy.

Images