Sun-induced chlorophyll fluorescence space-time prediction method and device

Abstract

The embodiment of the invention discloses a time-space prediction method and device for sun-induced chlorophyll fluorescence. The method comprises the following steps: acquiring a time-space data set; based on the spatio-temporal data set, taking a county administrative unit as a reference unit to perform area aggregation, dividing a large-scale area into a plurality of homogenized sub-areas, and selecting a representative spatial point from each sub-area; extracting the SIF of the corresponding representative spatial point and variable feature information of the key covariable, and constructing a four-dimensional space-time cube based on the variable feature information; sTL decomposition is carried out on the four-dimensional space-time cube to obtain a trend term, a seasonal term and a residual term, and flattening processing is carried out on the trend term, the seasonal term and the residual term to obtain a time sequence feature vector; the time sequence feature vector is input into a trained LSTM-Transform model for prediction, and a trend-season point value prediction result is obtained; and jointly interpolating the point value prediction result and a preset covariable to a preset target prediction area by using an ANUSPLIN thin plate smooth spline interpolation method, and generating an SIF large-scale space-time prediction map.

Description

Technical Field

[0001] This application relates to the field of monitoring changes in solar-induced chlorophyll fluorescence, and relates to, but is not limited to, a method and apparatus for spatiotemporal prediction of solar-induced chlorophyll fluorescence. Background Technology

[0002] In ecological environment monitoring and natural resource management, solar-induced fluorescence (SIF), as a spectral signal that can directly characterize photosynthetic efficiency, is increasingly being used in areas such as vegetation growth dynamics, carbon sink assessment, drought monitoring, and ecosystem health diagnosis. Compared to the Normalized Difference Vegetation Index (NDVI), SIF is more sensitive to changes in photochemical efficiency and less prone to saturation. It can reflect the physiological responses of vegetation to water stress and climate fluctuations in real time. Therefore, accurate prediction of the spatiotemporal variations of SIF over large-scale regions is of great significance for understanding the carbon cycle of terrestrial ecosystems and formulating sustainable management strategies.

[0003] However, large-scale regional SIF prediction still faces multiple challenges: First, the complex topography, vegetation diversity, and climate gradients result in highly heterogeneous and non-stationary SIF spatiotemporal distributions; second, the SIF signal itself is weak and greatly affected by cloud cover and atmospheric correction errors, leading to low signal-to-noise ratio (NDVI) and significant data gaps in the observed data; finally, existing methods often focus on the inherent patterns of time series data, neglecting the potential coupling relationship between SIF and environmental factors such as meteorology, soil moisture, vegetation structure, and human activities, thus limiting prediction accuracy and generalization ability.

[0004] Therefore, related technologies typically integrate multi-source remote sensing and meteorological data and introduce deep learning models to improve prediction accuracy. However, these technologies usually only perform point-to-point modeling or simple point-to-surface interpolation, limiting their ability to collaboratively model cross-scale spatially related features. Furthermore, single models are insufficient in handling long-sequence dependencies, suppressing overfitting, and maintaining spatial continuity. This reduces the prediction accuracy of spatiotemporal variations in SIF over large-scale regions and lowers the accuracy and reliability of vegetation photosynthetic efficiency monitoring and carbon sink assessment. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method and apparatus for predicting the spatiotemporal variation of solar-induced chlorophyll fluorescence (SIF). By comprehensively considering the cross-constraint effects of meteorological, vegetation physiological structure, soil and human activity data on SIF, and integrating county-level surface-point-surface cross-scale mapping, LSTM-Transformer model and ANUSPLIN thin plate smoothing spline interpolation method, the prediction accuracy of large-scale SIF spatiotemporal changes is improved.

[0006] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide a method for spatiotemporal prediction of solar-induced chlorophyll fluorescence, comprising: Raw datasets were obtained from multiple data sources. These raw datasets included first meteorological variables, first vegetation physiological state variables, first soil variables, and first human activity variables in a large-scale region. The raw datasets were preprocessed to generate a spatiotemporal dataset in a unified format. Based on this spatiotemporal dataset, surface aggregation was performed using county-level administrative units as the baseline unit in the large-scale region. Empirical wavelet decomposition and attention mechanisms were combined to divide the large-scale region into multiple homogeneous sub-regions. Each homogeneous sub-region exhibited similar environmental characteristics, and representative spatial points were selected from each sub-region. For each sub-region, the variable feature information of solar-induced chlorophyll fluorescence (SIF) and key covariates of the corresponding representative spatial points was extracted. Based on the variable feature information of each representative spatial point, a time-domain... A four-dimensional spatiotemporal cube is constructed in the spatial domain, where each representative spatial point corresponds to variable feature information at a specific spatial location and time point. The temporal data of each spatiotemporal unit in the four-dimensional spatiotemporal cube is decomposed using STL to obtain a trend term, a seasonal term, and a residual term. These terms are then flattened to obtain corresponding temporal feature vectors. These temporal feature vectors are input into a trained LSTM-Transformer model for spatiotemporal prediction, yielding trend-seasonal point value predictions for each representative spatial point. Using the ANUSPLIN thin-plate smoothing spline interpolation method, the point value predictions are interpolated along with preset covariates to a preset target prediction region, generating a large-scale SIF spatiotemporal prediction map with a smooth, continuous surface.

[0007] Secondly, embodiments of this application provide a solar-induced chlorophyll fluorescence spatiotemporal prediction device, comprising: The system comprises a data acquisition module, a partitioning module, a feature construction module, a prediction module, and a spatial interpolation module. Specifically: the data acquisition module acquires raw datasets from multiple data sources, including a first meteorological variable, a first vegetation physiological state variable, a first soil variable, and a first human activity variable within a large-scale region; it preprocesses the raw datasets to generate a unified spatiotemporal dataset; the partitioning module, based on the spatiotemporal dataset, performs surface aggregation using county-level administrative units within the large-scale region as the baseline unit, and divides the large-scale region into multiple homogeneous sub-regions using empirical wavelet decomposition and attention mechanisms. Each homogeneous sub-region has similar environmental characteristics, and representative spatial points are selected from each sub-region; the feature construction module extracts variable feature information of the solar-induced chlorophyll fluorescence (SIF) and key covariates from the corresponding representative spatial points within each sub-region. Based on the variable feature information of each representative spatial point, a four-dimensional spatiotemporal cube containing both time and spatial domains is constructed. Each representative spatial point in the four-dimensional spatiotemporal cube corresponds to variable feature information at a specific spatial location and a specific time point. The prediction module is used to perform STL decomposition on the temporal data of each spatiotemporal unit in the four-dimensional spatiotemporal cube to obtain a trend term, a seasonal term, and a residual term. The trend term, the seasonal term, and the residual term are flattened to obtain a corresponding temporal feature vector. The temporal feature vector is input into a trained LSTM-Transformer model for spatiotemporal prediction to obtain the trend-seasonal point value prediction result for each representative spatial point. The spatial interpolation module is used to interpolate the point value prediction result and a preset covariate to a preset target prediction region using the ANUSPLIN thin-plate smooth spline interpolation method to generate a large-scale spatiotemporal prediction map of SIF with a smooth and continuous surface.

[0008] The beneficial effects of the technical solutions provided in this application include at least the following: This application obtains raw datasets from multiple data sources, dividing a large-scale region into homogeneous sub-regions with similar environmental characteristics, thus reconciling the differences in spatial resolution among raw datasets from different sources. The large-scale region is divided using county-level administrative units, and the variable feature information of key covariates related to SIF in each sub-region is extracted. These key covariates consider the cross-constraints of multiple factors on SIF, including meteorology, vegetation physiological structure, soil, and human activities. The variable feature information is transformed into a temporal feature vector suitable for input to the LSTM-Transformer model, which outputs trend-seasonal point value predictions for each representative spatial point. The point value predictions reflect the future direction and rate of change and periodic fluctuations of SIF. Based on the point value predictions, a large-scale spatiotemporal prediction map of SIF is output using the ANUSPLIN thin-plate smoothing spline interpolation method, improving the accuracy and generalization ability of large-scale regional SIF spatiotemporal prediction. The large-scale spatiotemporal prediction map of SIF can serve as an important reference for monitoring vegetation photosynthesis, providing reliable technical support for regional-global carbon cycle research and ecological management. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A schematic flowchart of a solar-induced chlorophyll fluorescence spatiotemporal prediction method provided in an embodiment of this application; Figure 2 A schematic diagram of a historical dataset modeling framework for a solar-induced chlorophyll fluorescence spatiotemporal prediction method provided in this application embodiment; Figure 3 A schematic diagram of the input and output process of an LSTM-Transformer model for a solar-induced chlorophyll fluorescence spatiotemporal prediction method provided in this application embodiment; Figure 4 This is a schematic diagram of a solar-induced chlorophyll fluorescence spatiotemporal prediction device provided in an embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is 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.

[0012] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0013] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0014] This application provides a method for spatiotemporal prediction of solar-induced chlorophyll fluorescence. Figure 1 A flowchart illustrating a solar-induced chlorophyll fluorescence spatiotemporal prediction method provided in this application embodiment is shown below. Figure 1 As shown, the method includes at least the following steps: Step S110: Obtain the original dataset from multiple data sources. The original dataset includes the first meteorological variable, the first vegetation physiological state variable, the first soil variable, and the first human activity variable in a large-scale region. Preprocess the original dataset to generate a spatiotemporal dataset in a unified format. The original datasets were acquired using the European Centre for Medium-Range Weather Forecasts Reanalysis 5-Land (ERA5-Land), the Global Land Data Assimilation System Version 2.1 (GLDAS-2.1), and the Moderate Resolution Imaging Spectroradiometer (MODIS).

[0015] The primary meteorological variables include: surface solar radiation (SSR), monthly average temperature (MAT), and monthly cumulative precipitation (MCP). The primary vegetation physiological state variables include: leaf area index (LAI), fraction of absorbed photosynthetically active radiation (FPAR), and aboveground biomass (AGB). The primary soil variable is surface soil moisture (SSM). The primary human activity variable is population density (PD).

[0016] The original dataset is preprocessed, and datasets from different sources are standardized to generate a spatiotemporal dataset with a unified format. This addresses the issues of spatiotemporal resolution differences and missing values ​​in data from different sources, ensuring data integrity and uniform format.

[0017] Step S120: Based on the spatiotemporal dataset, surface aggregation is performed using county-level administrative units in the large-scale region as the reference unit. The large-scale region is divided into multiple homogeneous sub-regions by combining empirical wavelet decomposition and attention mechanism. Each homogeneous sub-region has similar environmental features. Representative spatial points are selected from each sub-region. The spatiotemporal datasets of large-scale regions are statistically summarized according to county-level administrative boundaries. Pixel-level data are transformed into area data of county-level administrative units. The average pixel value within a county-level administrative unit can approximate the value of that spatiotemporal point. Each county-level administrative unit consists of its administrative boundary and all its internal raster cells.

[0018] By performing area aggregation on the data in the spatiotemporal dataset, the problem of differences in spatial resolution between different datasets can be solved, generating mean, sum, or quantile statistics at the county scale. Dividing a large-scale region into numerous county-level administrative units can avoid unnecessary redundant information.

[0019] Empirical Wavelet Decomposition (EWD) is an adaptive signal decomposition method that separates multi-scale environmental features from surface domain data to obtain sub-data of different frequencies.

[0020] The attention mechanism is a weighted allocation model that automatically learns the contribution of different county or environmental characteristics to homogeneous regionalization. By weighting the features of the surface region sub-data, key environmental variables are highlighted, and the sub-data is clustered into homogeneous sub-regions with similar environmental characteristics based on the attention weights. For example, attention weights reveal that soil moisture and evaporation are the dominant factors in dividing arid or humid sub-regions. The county environmental characteristics within each sub-region are similar, and representative spatial points are selected from each sub-region, using the coordinates of the county government's location as the generation point.

[0021] Dividing a large-scale region into homogeneous sub-regions with similar environmental characteristics can reconcile the differences in spatial resolution between data from different sources.

[0022] This invention utilizes county-level administrative units to divide large-scale regions and extracts feature information, fully considering the cross-constraints of multiple factors such as meteorology, vegetation physiological structure, soil, and human activities on SIF, thereby improving the adaptability of the prediction model to complex ecosystems.

[0023] Step S130: For each sub-region, extract the variable feature information of key covariates related to the change of solar-induced chlorophyll fluorescence (SIF) at the corresponding representative spatial points. Based on the variable feature information of each representative spatial point, construct a four-dimensional spatiotemporal cube containing time and space domains. Each representative spatial point in the four-dimensional spatiotemporal cube corresponds to the variable feature information of a specific spatial location and a specific time point. Key covariates include the primary meteorological variable, primary vegetation physiological state variable, primary soil variable, and primary human activity variable in the large-scale region. Variable characteristic information consists of multi-dimensional environmental parameters corresponding to each representative spatial point at a specific time and space.

[0024] A four-dimensional spacetime cube is a structured data container that organizes variable feature information according to spatial, temporal, and variable dimensions to form a multidimensional array, and then stores all variables in a structured manner in the form of tables or tensors. In a four-dimensional spacetime cube, each representative spatial point (such as a key grid point within a county) corresponds to a variable feature information, which describes the specific attributes of each representative spatial point.

[0025] Step S140: Perform STL decomposition on the temporal data of each spatiotemporal unit in the four-dimensional spatiotemporal cube to obtain a trend term, a seasonal term, and a residual term. Flatten the trend term, the seasonal term, and the residual term to obtain the corresponding temporal feature vector. Input the temporal feature vector into the trained LSTM-Transformer model for spatiotemporal prediction to obtain the trend-seasonal point value prediction result for each representative spatial point. The Seasonal and Trend Decomposition using Loess (STL) method, based on Local Weighted Regression, is suitable for decomposing high-dimensional data and can simultaneously capture the long-term trends, periodic patterns, and sudden anomalies in the spatiotemporal dynamics of time-series data.

[0026] Using STL, a complete time series is extracted from each representative spatial point in a four-dimensional spatiotemporal cube. The time series of each variable at each representative spatial point is independently decomposed to obtain a trend term, a seasonal term, and a residual term. The trend term reflects the long-term, slow change of the variable, such as the slow increase in vegetation productivity due to global warming. The seasonal term extracts fluctuations with fixed periods; for example, it characterizes the vegetation growing season and precipitation cycles. The residual term removes anomalous events and noise after removing the trend and seasonal signals; for example, a sudden decrease in the SIF value caused by abnormal times such as drought or fire.

[0027] The time series data of the three components—trend, seasonality, and residual—are concatenated into a one-dimensional vector to form a fixed-length time series feature vector.

[0028] The LSTM-Transformer model is a hybrid deep learning model that combines a Long Short-Term Memory (LSTM) network and a Transformer module to process temporal feature vectors. LSTM excels at handling long-term dependencies in temporal feature vectors, enabling time series prediction and natural language processing. The LSTM module addresses the vanishing gradient problem through a gating mechanism.

[0029] The time-series feature vector is input into the LSTM-Transformer model for prediction, yielding the trend-seasonal point value prediction results for each representative spatial point. The trend-seasonal point value prediction result is the SIF value at a specific time point, which is the superposition result of trend and seasonal components. The SIF value reflects the long-term direction and rate of change and periodic fluctuation characteristics in the future.

[0030] Step S150: Using the ANUSPLIN thin-plate smooth spline interpolation method, the point value prediction results and preset covariates are interpolated together to the preset target prediction region to generate a large-scale spatiotemporal prediction map of SIF with a smooth and continuous surface.

[0031] The Australian National University Thin Plate Smoothing Spline (ANUSPLIN) is a spatial interpolation technique based on spline functions. By introducing covariates (such as elevation and latitude / longitude), it extends discrete point-value predictions to a continuous spatial surface with high accuracy. Regularization parameters balance fitting accuracy and surface smoothness, avoiding overfitting. The final result is a high-resolution, large-scale spatiotemporal prediction map covering the target area. For example, the resolution can reach 1 kilometer (km).

[0032] By using the ANUSPLIN thin-plate smoothing spline interpolation method, spatial interpolation is performed on the trend-season point value prediction results output by the LSTM-Transformer model to fill the gaps between adjacent county-level administrative units, thereby achieving SIF prediction from point to area and realizing comprehensive prediction of SIF spatiotemporal changes in large-scale regions.

[0033] The ANUSPLIN thin-plate smoothing spline interpolation method improves interpolation accuracy by minimizing the bending energy function, while maintaining the smoothness of the predicted surface. This is achieved by adaptively incorporating the influence of environmental covariates such as terrain. The thin-plate smoothing spline interpolation method is suitable for areas with complex terrain or sparse data, such as mountainous regions and deserts.

[0034] The ANUSPLIN thin-plate smooth spline interpolation method models the spatial autocorrelation and variability of the data, interpolating discrete points into a smooth, continuous surface. The theoretical statistical model of the ANUSPLIN thin-plate smooth spline interpolation method is shown in equation (1): Formula (1); In formula (1), Z i Let x be the variable to be interpolated located at spatial position i; i It is the spline independent variable; y i e is a predefined covariate in the P-dimensional dimension; iLet b be the random error term of the independent variable with an expected value of 0, and let b be the coefficient vector of the covariates.

[0035] In the ANUSPLIN thin-plate smooth spline interpolation method, the predefined covariates include: Digital Elevation Model (DEM), slope, and land use. Among them, DEM is the core predefined covariate. A DEM is a spatial dataset that digitally represents the surface elevation in the form of a regular grid; essentially, it is a digital map of the surface elevation. A DEM can represent the spatial distribution of surface elevation in raster or vector form.

[0036] This application obtains raw datasets from multiple data sources, dividing a large-scale region into homogeneous sub-regions with similar environmental characteristics, thus reconciling the differences in spatial resolution among raw datasets from different sources. The large-scale region is divided using county-level administrative units, and the variable feature information of key covariates related to SIF in each sub-region is extracted. These key covariates consider the cross-constraints of multiple factors on SIF, including meteorology, vegetation physiological structure, soil, and human activities. The variable feature information is transformed into a temporal feature vector suitable for input to the LSTM-Transformer model, which outputs trend-seasonal point value predictions for each representative spatial point. The point value predictions reflect the future direction and rate of change and periodic fluctuations of SIF. Based on the point value predictions, a large-scale spatiotemporal prediction map of SIF is output using the ANUSPLIN thin-plate smoothing spline interpolation method, improving the accuracy and generalization ability of large-scale regional SIF spatiotemporal prediction. The large-scale spatiotemporal prediction map of SIF can serve as an important reference for monitoring vegetation photosynthesis, providing reliable technical support for regional-global carbon cycle research and ecological management.

[0037] This invention acquires raw datasets using hyperspectral remote sensing technology, performs area aggregation, homogenized sub-region division, and constructs a four-dimensional spatiotemporal cube using GIS (Geographic Information System) spatial analysis technology, and builds an LSTM-Transformer model to achieve comprehensive integration and accurate modeling of multiple factors influencing SIF changes, including meteorological conditions, vegetation physiological states, soil conditions, and human activities. On one hand, the joint input of pre-defined covariates fully characterizes the potential spatiotemporal coupling relationship between SIF and environmental factors, significantly improving prediction accuracy. On the other hand, the combination of area-point-area cross-scale mapping and the LSTM-Transformer model enables large-scale spatiotemporal prediction of SIF from representative points at the county level to continuous pixel surfaces, enhancing the model's spatial generalization ability. This invention achieves high-precision spatiotemporal prediction of solar-induced chlorophyll fluorescence, which helps to more accurately assess regional carbon sink potential, reveal the dynamic evolution of photosynthetic efficiency, and provide scientific basis and decision support for ecological environmental protection and climate change response. This invention has significant application value in agricultural monitoring, carbon cycle assessment, and ecological disaster early warning, providing a scientific basis for precise environmental management.

[0038] In some embodiments, step S110, "preprocessing the original dataset," includes: Step S1101: Perform geometric correction, projection stitching, resampling, maximum value synthesis, and missing value imputation on the original dataset.

[0039] The original dataset is preprocessed to generate a unified spatiotemporal dataset. Specifically, projection stitching unifies data from different sources to the same coordinate system, achieving seamless stitching; resampling adjusts the spatial resolution to a uniform scale, for example, resampling remote sensing images of different resolutions to a 1km grid; maximum value synthesis eliminates noise and preserves the true features of the original dataset; and missing value imputation is performed through spatiotemporal interpolation or neighbor-to-neighbor filling to repair data gaps. The resulting unified spatiotemporal dataset has a consistent spatiotemporal reference and a standardized storage structure. For example, the storage structure is in tensor form.

[0040] By preprocessing, the spatial and temporal resolution differences between the original datasets from different data sources are unified into a form suitable for the input of the LSTM-Transformer model.

[0041] In some embodiments, the "variable feature information" in step S130 includes statistical features in the time domain and geometric features in the spatial domain. The "extraction of variable feature information of key covariates related to solar-induced chlorophyll fluorescence (SIF) changes at representative spatial points for each sub-region" in step S130 includes: Step S1301: For each sub-region, the spatiotemporal dataset is subjected to correlation and importance analysis using at least one of the following methods: geographic detector method, feature arrangement method, and sensitivity analysis experiment method, to screen out key covariates related to SIF changes. The geographic detector method identifies environmental factors significantly correlated with the spatial distribution of SIF, such as climate and soil, through spatial stratification heterogeneity testing. The feature ranking method quantifies the importance of each key covariate to the prediction results based on changes in model prediction performance. Sensitivity analysis experiments assess the impact of perturbation variables on the SIF output. Finally, key covariates with strong correlation to SIF changes or high explanatory power are selected, providing key covariates with a high signal-to-noise ratio.

[0042] This application can use three methods—geographic detector method, feature ranking method, and sensitivity analysis experiment method—to screen key covariates related to SIF changes. Alternatively, any one or any two of these three methods can be used to screen key covariates related to SIF changes.

[0043] Step S1302: For each sub-region, extract the statistical characteristics of the key covariates of the corresponding representative spatial points in the time domain and their geometric characteristics in the spatial domain. For each sub-region, temporal statistical features and spatial geometric features are extracted from key covariates of representative spatial points. Temporal statistical features quantify the dynamic changes of covariates, while spatial geometric features characterize the spatial distribution patterns of covariates. Temporal statistical features include mean, standard deviation, maximum value, minimum value, and trend slope. Spatial geometric features include adjacency similarity and slope variance.

[0044] We extract the SIF and key covariate features from the mean data within each sub-region. By fusing temporal statistical features and spatial geometric features, we construct a multidimensional feature set that combines temporal dynamics and spatial heterogeneity.

[0045] Step S1303, the construction of a four-dimensional spatiotemporal cube containing time and space domains based on the variable feature information of each representative spatial point, includes: constructing a four-dimensional spatiotemporal cube containing time and space domains based on the statistical features of each representative spatial point in the time domain and the geometric features in the spatial domain.

[0046] The four-dimensional spatiotemporal cube comprises spatial longitude, spatial latitude, time series data, and variables. It unifies multi-scale data, integrating point, area, and time series data into structured tensor data, facilitating input into LSTM-Transformer models. By capturing spatiotemporal dynamics through temporal statistical features and spatial geometric features, such as driving factors of vegetation mutations, it enhances feature representation capabilities. The four-dimensional spatiotemporal cube supports causal analysis, for example, by combining it with the geographic detector method to analyze the spatial coupling relationship between SIF and key covariates.

[0047] In some embodiments, the "LSTM-Transformer model" in step S140 includes an input module, an LSTM module, a Transformer module, and an output module; the "inputting the temporal feature vector into the trained LSTM-Transformer model for spatiotemporal prediction to obtain the trend-seasonal point value prediction result for each representative spatial point" in step S140 includes: Step S1401: Receive the time-series feature vector through the input module, preprocess the time-series feature vector to obtain a unified dimension vector; The temporal feature vectors undergo standardization, dimension alignment, tensor reshaping, and masking. Specifically, each feature dimension is standardized or scaled to eliminate dimensional differences. Temporal features of different lengths or sources are padded or truncated to a fixed length to ensure consistent input tensor shape. Masks are generated for missing values ​​or padded portions to prevent the model from focusing on invalid data. A uniform-dimensional temporal tensor is output, which can be directly input into subsequent LSTM modules for temporal feature extraction. This ensures consistency of input to the LSTM-Transformer model, improving its generalization ability to spatiotemporal data.

[0048] Step S1402: The LSTM module is used to extract features from the unified dimension vector to analyze the long-term dependencies in the unified dimension vector. The transmission and forgetting of information are dynamically controlled through a gating mechanism to obtain multivariate time-series features. The LSTM module includes cascaded multi-layer bidirectional LSTM layers, each of which includes multiple LSTM units; each LSTM unit includes a cell state, a forget gate, an input gate, and an output gate. 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 cell state with the input information at time t. The output gate controls the output of the cell state.

[0049] In some embodiments, an LSTM cell contains a cell state. The three neural network layers are the forget gate. Input gate and output gate The calculation process of the LSTM unit at time t is shown in formulas (2) to (7).

[0050] Forgotten Gate The decision is based on the previous cell state. Discarded information: Formula (2); Input gate : Formula (3); Formula (4); Update cell state : Formula (5); Output gate : Formula (6); Formula (7); In formulas (2) to (7), This represents the uniform dimension vector input to the LSTM unit at time t; , Let represent the hidden state at time t-1 and the hidden state at time t, respectively. Use the sigmoid activation function; This represents the candidate cell state at time t; is the activation function corresponding to the candidate cell state; , These represent the cell state at time t-1 and the cell state at time t, respectively. These represent the weight vectors for the forget gate, input gate, output gate, and candidate cell state, respectively. These represent the bias vectors for the forget gate, input gate, output gate, and candidate cell state, respectively.

[0051] Each LSTM layer not only processes the SIF variable, but also processes multiple key covariates that affect the SIF, enhancing the model's ability to extract multi-factor features.

[0052] Step S1403: Through the Transformer module, feature extraction is performed on the multivariate temporal features using a multi-head attention mechanism to analyze the correlation between different spatial locations and transform the multivariate temporal features into a spatiotemporal fusion code with high-order spatiotemporal features. Step S1404: Through the output module, the spatiotemporal fusion code is subjected to nonlinear transformation and dimensional adjustment to obtain the trend-seasonal point value prediction results for each representative spatial point.

[0053] In some embodiments, the "Transformer module" in step S1403 includes: a location encoding layer, a multi-head self-attention layer, and a feedforward network layer; step S1403, through the Transformer module, uses a multi-head attention mechanism to extract features from the multivariate temporal features to analyze the correlation between different spatial locations, and transforms the multivariate temporal features into a spatiotemporal fusion encoding with high-order spatiotemporal features, includes: Step S14031: Through the location encoding layer, the temporal location information and spatial location information are encoded into vectors and embedded into the multivariate temporal features to obtain sequential encoding features; Step S14032: Through the multi-head self-attention layer, the sequential encoding features are dynamically weighted and fused in multiple modes. The residual connection is used to preserve the original information and the layer normalization is used to optimize the gradient propagation, thereby generating global association features that represent global spatiotemporal dependence. The multi-head self-attention layer first maps sequential encoded features into query (Q), key (K), and value (V) vectors. Then, it calculates scaled dot product correlations within each attention head to generate weights, and then performs a weighted summation of V according to the weights. After concatenating and linearly transforming the results of each attention head, it can capture the dependencies between different time or spatial locations in the global scope. The output global correlation features provide rich feature representations for the subsequent model. The implementation of the multi-head self-attention layer is shown in Equations (8) to (10): Linear mapping (h-th head) ): Formula (8); Scaling Dot Product Attention : Formula (9); Multi-head splicing and output mapping : Formula (10); In formulas 8 to 10, H represents the number of attention heads, indicating that the attention heads are concatenated, and X represents the sequential encoding features of the input. For the dimensions of Q and K in a single head, For the query, key, and value matrix of the h-th head, For learnable linear mapping weights, This represents the weight vector of the output gate. This is a scaling factor to prevent large gradients from being caused by large-dimensional dot products.

[0054] Step S14033: Through the feedforward network layer, the global correlation features are subjected to two fully connected projections and nonlinear activations of higher order transformation, and the original input and the layer normalized stable feature distribution are fused through residual connections to output the spatiotemporal fusion code.

[0055] Two fully connected layers linearly transform the globally correlated features, mapping them to a higher-dimensional space to obtain higher-order features and enhance the model's expressive power. A non-linear activation function is introduced between the two projections to capture complex feature interactions. Residual connections alleviate the vanishing gradient problem in deep networks by element-wise adding the globally correlated features to the higher-order features, while preserving low-level feature information. Layer normalization is applied to the features after residual connections, outputting a spatiotemporal fusion encoding that stabilizes feature distribution, accelerates model convergence, and improves training stability.

[0056] In some embodiments, the method further includes: Step S101: Construct the framework of the initial LSTM-Transformer model; The initial LSTM-Transformer model was built using the Keras neural network API in Python: first, multiple LSTM units were stacked to form a bidirectional LSTM layer; then, a Transformer module with positional encoding, scaled dot product multi-head self-attention layers (including residual connections and layer normalization) was inserted; finally, a fully connected layer was added; the entire process was accelerated using GPU training. This model, through its cascaded structure, gradually integrates the influence of multiple factors on the target variable, providing more accurate and comprehensive point value prediction results, and can automatically extract key features from the input temporal feature vector during training.

[0057] Step S102: Adaptive gradient adjustment is performed using the Adam optimizer, and the update step size of the model parameters is dynamically adjusted based on the preset learning rate to optimize the learning process of the initial model. The Adaptive Moment Estimation (Adam) optimizer, based on a preset initial learning rate, assigns an independent adaptive learning rate to each parameter by calculating the first moment (mean) and second moment (variance) of the gradient in real time. The first moment tracks the exponential moving average of the gradient, accelerating convergence and suppressing oscillations. The second moment estimates the exponential moving average of the squared gradient, adaptively adjusting the parameter update step size. By adaptively adjusting the parameter update process in the LSTM-Transformer model, the Adam optimizer significantly improves the model's learning efficiency for spatiotemporal fusion encoding.

[0058] Step S103: Obtain historical datasets from multiple data sources, including SIF variables, second meteorological variables, second vegetation physiological state variables, second soil variables, and second human activity variables. The historical dataset consists of five categories of historical observation variables. The SIF variable was obtained from the Global OCO-2 Solar-Induced Chlorophyll Fluorescence dataset (GOSIF v2). The second meteorological variable, the second vegetation physiological state variable, the second soil variable, and the second human activity variable were obtained from the European Centre for Medium-Range Weather Forecasts Reanalysis 5-Land (ERA5-Land), the Global Land Data Assimilation System Version 2.1 (GLDAS-2.1), and the Moderate Resolution Imaging Spectroradiometer (MODIS).

[0059] The second meteorological variable includes surface solar shortwave radiation, monthly average temperature, and monthly cumulative precipitation; the second vegetation physiological state variable includes leaf area index, photosynthetically active radiation absorption ratio, and aboveground biomass; the second soil variable includes topsoil moisture content; and the second human activity variable includes population density. The SIF variable includes historical SIF observations for the prediction area.

[0060] Among them, the second meteorological variable, the second vegetation physiological state variable, the second soil variable, and the second human activity variable are used as the second key covariates and input into the LSTM-Transformer model. The key covariates are the independent variables of the LSTM-Transformer model, and the SIF variable is the dependent variable of the LSTM-Transformer model.

[0061] Figure 2 This is a schematic diagram of the historical dataset modeling framework for a solar-induced chlorophyll fluorescence spatiotemporal prediction method provided in an embodiment of this application.

[0062] like Figure 2As shown, this diagram illustrates the entire process from microscopic observation of the plant canopy to the construction of a macroscopic spatiotemporal cube. Text box 201 on the left marks the four core levels of data abstraction: 1) acquiring the plant canopy, generating canopy information, and correspondingly acquiring and preprocessing historical datasets; 2) extracting key variables (key covariates, Cvpsbh+) related to SIF from the canopy information; 3) obtaining the actual SIF change sequence from the canopy information; and 4) structurally storing the key covariates (i.e., a four-dimensional spatiotemporal cube, SpaPoCvpsbh+) in the form of a spatial point cube. On the right, 202 dynamically displays the evolution of data at each level through a horizontal time series panel (t-1, t, t+1). The legend 203 at the bottom explains that (P(i,j)) represents spatial location, dashed circles represent spatial relationships, and bidirectional arrows represent temporal relationships. A small schematic diagram 204 illustrates the process from Cvpsbh+ to SpaPoCvpsbh+.

[0063] Based on the variable feature information of each representative spatial point, a four-dimensional spatiotemporal cube is constructed, as shown in formula (11): Formula (11); In formula (11), This represents the predicted SIF point value at time t. ψ represents the spatiotemporal features in a four-dimensional cube at time t, and ψ represents the mapping relationship from the spatiotemporal features to the SIF point value prediction results.

[0064] Step S104: The historical dataset is cyclically split into three subsets through triple cross-validation, and three rounds of training-validation are performed. In each round, the initial model is trained with two subsets and the generalization of the initial model is validated with one subset. The historical dataset was randomly and evenly divided into three subsets to ensure that each subset covered all spatiotemporal ranges and avoided seasonal or regional biases. The original correspondences between key covariates were maintained, and three rounds of training-validation were performed. In each round, a different combination of two subsets was used to train the initial LSTM-Transformer model. In each round, one subset was used to validate the generalization ability of the model after this round of training, and the validation metric values ​​for each round were recorded. The average of the three validation metrics was calculated as the final evaluation of generalization ability. The best-performing model obtained from the three rounds of training-validation was retained and ensembled using a weighted average.

[0065] Step S105: Optimize the hyperparameter combinations of the initial model by traversing the hyperparameter combinations using GridSearchCV; the parameter combinations include the number of hidden units, the number of layers, the learning rate, and the Dropout ratio.

[0066] Grid Search Cross-Validation (GridSearchCV) automatically optimizes the hyperparameters of the model obtained through triple cross-validation by using grid search and cross-validation. Specifically, grid search systematically searches for the optimal model configuration by traversing predefined parameter combinations.

[0067] Dropout Regularization Technique (Dropout) is a regularization technique in deep learning that prevents overfitting by randomly zeroing out some neurons in a neural network during model training.

[0068] The model obtained through hyperparameter optimization can simultaneously characterize the long-term temporal dependencies and spatial heterogeneity in spatiotemporal datasets. By using data from multiple data sources as joint inputs, it can effectively improve the accuracy of SIF point value prediction results.

[0069] Figure 3 This diagram illustrates the input and output process of an LSTM-Transformer model for a solar-induced chlorophyll fluorescence spatiotemporal prediction method provided in this application embodiment.

[0070] like Figure 3 As shown, the data processing from historical time (t-1) to future prediction (t+1) is illustrated. Cvpsbh+SIF represents the historical dataset, which includes SIF variables and a second key covariate. The SpaPoCvpsbh+SIF module gradually transforms the historical dataset into spatiotemporal fusion features, and finally outputs the prediction result SpaPoSIF.

[0071] The SIF point prediction results (point value prediction results) are output by cascading multiple bidirectional LSTM units and multi-head self-attention layers. Normalization (BN) is used to normalize the input data. After the last layer of the Transformer module, a Dropout layer and a fully connected Dense layer are added to suppress overfitting and achieve non-linear feature mapping.

[0072] Figure 3 In this context, DN represents the input layer, and DDN represents the output layer. to This represents a uniform dimension vector input to the LSTM unit. This represents the propagation relationship from the m-th layer to the n-th layer of an LSTM. This indicates that the cell state is transferred from the a-th LSTM unit to the n-th LSTM unit. The trend-season point value prediction result of the future SIF obtained by the LSTM-Transformer model is shown in formula (12): Formula (12); In formula (12), Let δ be the temporal feature vector from the initial time step to the current time step, and let δ be the LSTM-Transformer model. This represents the predicted point value of SIF at time t+1.

[0073] In some embodiments, the method further includes: Step S160: Obtain the actual SIF change sequence of the target prediction region, and compare the predicted SIF value in the large-scale spatiotemporal prediction map with the true SIF value in the actual SIF change sequence point by point to obtain the evaluation index of the trained LSTM-Transformer model. The evaluation index includes: coefficient of determination, adjusted coefficient of determination, root mean square error, mean absolute percentage error, explained variance, and mean directional accuracy. Based on the evaluation index, evaluate the performance of the trained LSTM-Transformer model to obtain the evaluation result. Based on the evaluation result, update the trained LSTM-Transformer model.

[0074] The actual SIF values ​​are obtained from the measured SIF variables of the prediction region using GOSIF v2 data. This ensures that the predicted values ​​match the actual values ​​in both time and space.

[0075] The model's ability to explain SIF changes is quantified by the coefficient of determination and the adjusted coefficient of determination. The absolute or relative deviation between the predicted SIF value and the true SIF value is assessed by the root mean square error and the mean absolute percentage error. The ability of the model to capture SIF fluctuations is measured by the explained variance. The accuracy of trend prediction is judged by the mean directional accuracy, thus forming a comprehensive performance evaluation result.

[0076] Based on the evaluation results, if systematic errors are found in a specific spatiotemporal region or combination of variables, the model is optimized based on the systematic errors, ultimately achieving iterative improvement in the model's prediction accuracy and generalization ability.

[0077] By calculating the Population Stability Index, we can check whether the distribution of the prediction results of the LSTM-Transformer model is consistent across different time periods or spatial regions, thus verifying the model's stability and generalization ability and avoiding performance degradation due to data offset.

[0078] The following describes a solar-induced chlorophyll fluorescence spatiotemporal prediction method according to a specific embodiment. This specific embodiment is only for better illustrating the invention and does not constitute an undue limitation of the invention. The specific implementation process includes data preparation, feature extraction, model construction, prediction, and result interpolation in sequence.

[0079] To conduct spatiotemporal prediction of SIF (Self-Induced Irradiation Factor), various relevant data need to be collected, including SIF data, meteorological data, vegetation physiological structure data, and soil and human activity data. The data time range covers 2000 to 2022, and the spatial resolution varies depending on the data source.

[0080] In the data preprocessing stage, data cleaning, interpolation, and resampling are required to ensure data integrity and consistency. In particular, it is necessary to address the differences in spatial and temporal resolution between different data sources and unify them into a format suitable for model input. For missing values, methods such as nearest-neighbor interpolation or time-series interpolation are used to impute them.

[0081] This invention divides a large-scale areal region into county-level administrative units with similar ecological functions and preserved administrative interpretability, based on county boundaries. It assumes that the average pixel value within a county-level administrative unit can approximately represent the value at that spatiotemporal point. Each county-level administrative unit consists of its administrative boundary and all internal grid cells, with the coordinates of the county government's location used as the generation point.

[0082] Dividing large-scale areal regions into numerous county-level administrative units avoids unnecessary redundant information and significantly enhances the LSTM-Transformer model's ability to capture multi-scale features from local details to macro-trends. For each pixel within a county-level administrative unit, empirical wavelet decomposition is first performed for denoising, followed by the calculation of attention weights.

[0083] Within each county-level administrative unit, feature information of the SIF and its highly important covariates is extracted, including statistical features such as mean, standard deviation, maximum, minimum, and trend slope in time-series data, as well as geometric features such as adjacency similarity and slope variance in spatial distribution. The feature information of each county-level administrative unit is used to construct a corresponding spatiotemporal cube, which is then flattened to obtain feature vectors suitable for model learning and training, providing complete input for subsequent network prediction.

[0084] An LSTM-Transformer model is introduced, which leverages the strength of LSTM in capturing long-term temporal dependencies by incorporating a self-attention mechanism to learn cross-county spatial relationships. Compared to a single LSTM module or traditional machine learning models, this structure avoids gradient explosion and can simultaneously handle long sequences and multi-point correlations, thus exhibiting higher stability and accuracy in SIF prediction.

[0085] The invention utilizes Keras, a neural network API in Python, to construct an initial LSTM-Transformer model. First, two layers of bidirectional LSTM units are stacked to form a deep temporal encoder. Input and output layers are then built simultaneously, and GPU acceleration is used to train the model. This hybrid model leverages the gated memory of LSTM and the global dependency capture function of the Transformer module to automatically extract spatiotemporal features from the input data. Simultaneously, its parallel structure can process multiple input sequences concurrently and share weights and biases among memory units corresponding to different variables. This allows it to learn the patterns of multiple factors interacting with the SIF (Sequential Indicator Function), while preserving the unique information of each variable.

[0086] Before training, each input sequence is divided into two segments chronologically: the first 80% is used as the training set, and the latter 20% is used as the test set. Then, GridSearchCV from scikit-learn is used to perform a grid search on the model's parameters for each county-level administrative unit's feature vector to improve prediction performance. This study predetermines the learning rate, number of hidden units, and candidate values ​​for the Dropout ratio, iterates through all combinations, and records the validation error to determine the optimal parameters. All county-level administrative units share the same loss function, simultaneously optimizing the prediction results for all variables. Furthermore, triple cross-validation is introduced during training to verify the reliability of the optimal parameters. Table 1 shows the parameters and their corresponding values ​​in the LSTM-Transformer model proposed in this application.

[0087] Table 1. Parameters and their corresponding values In Table 1, the AdamW optimizer was used with an initial learning rate of 0.001 for parameter updates, and the batch size was set to 64 to balance memory efficiency and gradient stability. The LSTM-Transformer model was iteratively optimized through 100 training epochs. The GELU activation function was used to enhance the nonlinear feature representation ability, and a deactivation rate of 0.2 was introduced to suppress the risk of overfitting. The Huber loss (δ=0.1) was selected as the loss function to reduce outlier interference, while the Polyak averaging mechanism was turned off to simplify the training process.

[0088] The test set data is input into the trained LSTM-Transformer model to obtain the SIF prediction values ​​for each county. These prediction values ​​reflect the spatiotemporal variation characteristics of SIF under given multi-factor constraints. To extend the county-level prediction results to the entire prediction area, an improved ANUSPLIN thin-plate smoothing spline interpolation method is used to interpolate the prediction results. The interpolation results present a continuous large-scale spatiotemporal prediction map of SIF, facilitating subsequent analysis and application.

[0089] like Figure 3 As shown in the figure, the present invention reclassifies and encodes SIF predicted values ​​in combination with spatial information, and uses the ANUSPLIN thin-plate smoothing spline interpolation method to extend them to the regional level with a resolution of 1 km.

[0090] Although the improved ANUSPLIN thin-plate smoothing spline interpolation method may weaken local extremum information due to its smoothing process, it effectively reduces noise amplification and sample imbalance issues generated during feature extraction and deep network prediction. Compared with commonly used interpolation methods, the ANUSPLIN thin-plate smoothing spline interpolation method is more suitable for processing long-term series spanning multiple years and can complete multi-surface collaborative interpolation in one step. To ensure the accuracy of interpolation results for all seasons and even the whole year, this invention adopts a quadratic local thin-plate smoothing spline function model, using latitude and longitude as independent variables and incorporating the top 10% key covariates contributing to SIF as external covariates for fitting, thereby generating a smooth and realistic large-scale spatiotemporal prediction map of SIF with a resolution of 1km.

[0091] Multiple evaluation indicators are used to assess the prediction results. Furthermore, this method is compared with other mainstream prediction methods to demonstrate its superiority.

[0092] In summary, the large-scale SIF spatiotemporal prediction method proposed in this invention, which couples multivariate constraints with spatial expansion, significantly improves the accuracy and generalization ability of large-scale regional SIF spatiotemporal prediction by integrating county-level surface-to-point mapping, the LSTM-Transformer model, and the improved ANUSPLIN thin-plate smoothing spline interpolation method. This method has demonstrated excellent performance in practical applications in the grassland-agro-pastoral ecotone of northern China, validating its broad application prospects and promotional value in ecological carbon sink assessment and vegetation photosynthetic efficiency monitoring. Future work will continue to optimize the network structure, introduce more driving factors, and expand to global-scale scenarios, providing more reliable technical support for regional and global carbon cycle research and ecological management.

[0093] The embodiments of this application have the following advantages: Improved prediction accuracy: Integrating multi-source covariates fully couples the potential spatiotemporal relationships between SIF and the corresponding environmental factors, significantly reducing prediction errors. Achieving continuous SIF mapping from point to area, balancing high resolution and cross-regional interpretability. Reduced complexity and uncertainty: Employing geographic detector methods, feature ranking methods, and sensitivity analysis experiments to screen highly important variables, removing redundant information, and reducing model complexity and result uncertainty. Provides key technical support: Providing high-precision, automated spatiotemporal prediction solutions for large-scale regional carbon sink assessment, grassland productivity monitoring, and the achievement of carbon neutrality goals.

[0094] Figure 4 A schematic diagram of a solar-induced chlorophyll fluorescence spatiotemporal prediction device provided in this application embodiment is shown below. Figure 4 As shown, the device 400 includes a data acquisition module 410, a partitioning module 420, a feature construction module 430, a prediction module 440, and a spatial interpolation module 450; wherein: The data acquisition module 410 is used to acquire raw datasets from multiple data sources. The raw datasets include a first meteorological variable, a first vegetation physiological state variable, a first soil variable, and a first human activity variable in a large-scale region. The raw datasets are preprocessed to generate a spatiotemporal dataset in a unified format. The partitioning module 420 is used to perform surface aggregation based on the spatiotemporal dataset, taking the county-level administrative units in the large-scale region as the reference unit, and combining empirical wavelet decomposition and attention mechanism to divide the large-scale region into multiple homogeneous sub-regions. Each of the homogeneous sub-regions has similar environmental features, and representative spatial points are selected from each of the sub-regions. The feature construction module 430 is used to extract variable feature information of key covariates related to the change of solar-induced chlorophyll fluorescence (SIF) at representative spatial points for each sub-region, and to construct a four-dimensional spatiotemporal cube containing time and space domains based on the variable feature information of each representative spatial point. Each representative spatial point in the four-dimensional spatiotemporal cube corresponds to variable feature information at a specific spatial location and at a specific time point. The prediction module 440 is used to perform STL decomposition on the temporal data of each spatiotemporal unit in the four-dimensional spatiotemporal cube to obtain a trend term, a seasonal term, and a residual term; to flatten the trend term, the seasonal term, and the residual term to obtain a corresponding temporal feature vector; and to input the temporal feature vector into a trained LSTM-Transformer model for spatiotemporal prediction to obtain the trend-seasonal point value prediction result for each representative spatial point. The spatial interpolation module 450 is used to interpolate the point value prediction results and preset covariates to a preset target prediction region using the ANUSPLIN thin plate smooth spline interpolation method, thereby generating a large-scale SIF spatiotemporal prediction map with a smooth and continuous surface.

[0095] In some embodiments, the data acquisition module 410 further includes a preprocessing submodule.

[0096] The preprocessing submodule is used to preprocess the original dataset, including: geometric correction, projection stitching, resampling, maximum value synthesis, and missing value imputation of the original dataset.

[0097] In some embodiments, the feature construction module 430 includes a filtering submodule, a feature extraction submodule, and a construction submodule. The variable feature information includes statistical features in the time domain and geometric features in the spatial domain.

[0098] The screening submodule is used to perform correlation and importance analysis on the spatiotemporal dataset for each sub-region using at least one of the following methods: geographic detector method, feature ranking method, and sensitivity analysis experiment method, to screen out key covariates related to SIF changes. The feature extraction submodule is used to extract the statistical features of key covariates of corresponding representative spatial points in the time domain and their geometric features in the spatial domain for each sub-region. A submodule is constructed to build a four-dimensional spatiotemporal cube containing both time and space domains, based on the statistical characteristics of each representative spatial point in the time domain and its geometric characteristics in the spatial domain.

[0099] In some embodiments, the prediction module 440 includes an input submodule, an LSTM submodule, a Transformer submodule, and an output submodule; The input submodule is used to receive the time-series feature vector, preprocess the time-series feature vector, and obtain a unified dimension vector. The LSTM submodule is used to extract features from the unified dimension vector to analyze the long-term dependencies in the unified dimension vector. It dynamically controls the transmission and forgetting of information through a gating mechanism to obtain multivariate temporal features. The Transformer submodule is used to extract features from the multivariate temporal features through a multi-head attention mechanism to analyze the correlation between different spatial locations and transform the multivariate temporal features into a spatiotemporal fusion code with high-order spatiotemporal features. The output submodule is used to perform nonlinear transformation and dimensional adjustment on the spatiotemporal fusion encoding to obtain the trend-seasonal point value prediction results for each representative spatial point.

[0100] In some embodiments, the LSTM submodule includes cascaded multi-layer bidirectional LSTM layers, each of which includes multiple LSTM units; each of the LSTM units includes a cell state, a forget gate, an input gate, and an output gate; 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 cell state with the input information at time t, and the output gate controls the output of the cell state.

[0101] In some embodiments, the Transformer submodule includes: a position encoding layer, a multi-head self-attention layer, and a feedforward network layer; The location encoding layer is used to encode temporal and spatial location information into vectors, which are then embedded into the multivariate temporal features to obtain sequential encoding features. The multi-head self-attention layer is used to perform multi-mode dynamic weighted fusion of the sequential encoding features, and combines residual connections to preserve the original information and layer normalization to optimize gradient propagation, thereby generating global correlation features that represent global spatiotemporal dependence. The feedforward network layer is used to perform two fully connected projections and nonlinear activations on the global correlation features, and to fuse the original input and the layer normalized stable feature distribution through residual connections to output the spatiotemporal fusion code.

[0102] In some embodiments, the device 400 further includes a training module.

[0103] The training module is used to construct the framework of the initial LSTM-Transformer model. Adaptive gradient adjustment is performed using the Adam optimizer, dynamically adjusting the update step size of the model parameters based on a preset learning rate to optimize the learning process of the initial model. Historical datasets are obtained from multiple data sources, including SIF variables, a second meteorological variable, a second vegetation physiological state variable, a second soil variable, and a second human activity variable. The historical datasets are cyclically split into three subsets using triple cross-validation, and three rounds of training-validation are performed. In each round, the initial model is trained with two subsets, and the generalization ability of the initial model is validated with one subset. The hyperparameter combinations of the initial model are traversed using GridSearchCV to optimize the hyperparameter combinations, which include the number of hidden units, the number of layers, the learning rate, and the Dropout ratio.

[0104] In some embodiments, the apparatus further includes an evaluation module.

[0105] The evaluation module is used to acquire the actual SIF change sequence of the target prediction region, compare the predicted SIF value in the large-scale spatiotemporal prediction map with the true SIF value in the actual SIF change sequence point by point, and obtain the evaluation metrics for the trained LSTM-Transformer model. The evaluation metrics include: coefficient of determination, adjusted coefficient of determination, root mean square error, mean absolute percentage error, explained variance, and mean directional accuracy. Based on the evaluation metrics, the performance of the trained LSTM-Transformer model is evaluated to obtain the evaluation result. Based on the evaluation result, the trained LSTM-Transformer model is updated.

[0106] It should be noted that the descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. For technical details not disclosed in the system embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0107] It should be noted that, in the embodiments of this application, if the above-mentioned solar-induced chlorophyll fluorescence spatiotemporal prediction method is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0108] Correspondingly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps in any of the solar-induced chlorophyll fluorescence spatiotemporal prediction methods described in the above embodiments. Correspondingly, embodiments of this application also provide a computer program product, which, when executed by a processor of an electronic device, is used to implement the steps in any of the solar-induced chlorophyll fluorescence spatiotemporal prediction methods described in the above embodiments.

[0109] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential 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 embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0110] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0111] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0112] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected to achieve the purpose of the embodiments of this application according to actual needs. In addition, each functional unit in the embodiments of this application may be fully integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the integrated unit may be implemented in hardware or in the form of hardware plus software functional units.

[0113] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0114] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict. The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined to obtain new method embodiments or device embodiments without conflict.

[0115] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for spatiotemporal prediction of solar-induced chlorophyll fluorescence, characterized in that, include: The original datasets are obtained from multiple data sources. The original datasets include the first meteorological variable, the first vegetation physiological state variable, the first soil variable, and the first human activity variable in a large-scale region. The original datasets are preprocessed to generate a spatiotemporal dataset in a unified format. Based on the spatiotemporal dataset, the county-level administrative units in the large-scale region are used as the reference units for surface aggregation. The large-scale region is divided into multiple homogeneous sub-regions by combining empirical wavelet decomposition and attention mechanism. The environmental features of each homogeneous sub-region are similar, and representative spatial points are selected from each sub-region. For each sub-region, the variable feature information of key covariates related to the change of solar-induced chlorophyll fluorescence (SIF) at the corresponding representative spatial points is extracted. Based on the variable feature information of each representative spatial point, a four-dimensional spatiotemporal cube containing time and space domains is constructed. Each representative spatial point in the four-dimensional spatiotemporal cube corresponds to the variable feature information of a specific spatial location and a specific time point. STL decomposition is performed on the time series data of each spatiotemporal unit in the four-dimensional spatiotemporal cube to obtain trend term, seasonal term and residual term. The trend term, seasonal term and residual term are flattened to obtain the corresponding time series feature vector. The temporal feature vector is input into the trained LSTM-Transformer model for spatiotemporal prediction to obtain the trend-seasonal point value prediction results for each representative spatial point. Using the ANUSPLIN thin-plate smooth spline interpolation method, the point value prediction results are interpolated together with preset covariates to a preset target prediction region to generate a large-scale SIF spatiotemporal prediction map with a smooth and continuous surface.

2. The method according to claim 1, characterized in that, The variable feature information includes statistical features in the time domain and geometric features in the spatial domain. For each sub-region, the variable feature information of key covariates related to changes in solar-induced chlorophyll fluorescence (SIF) at representative spatial points is extracted, including: For each sub-region, correlation and importance analysis are performed on the spatiotemporal dataset using at least one of the following methods: geographic detector method, feature arrangement method, and sensitivity analysis experiment method, to screen out key covariates related to SIF changes. For each sub-region, the statistical characteristics of key covariates in the time domain and their geometric characteristics in the spatial domain of the corresponding representative spatial points are extracted. The construction of a four-dimensional spatiotemporal cube, comprising time and spatial domains, based on variable feature information from representative spatial points includes: Based on the statistical characteristics of representative spatial points in the temporal domain and their geometric characteristics in the spatial domain, a four-dimensional spatiotemporal cube containing both the temporal and spatial domains is constructed.

3. The method according to claim 1, characterized in that, The LSTM-Transformer model includes an input module, an LSTM module, a Transformer module, and an output module. The step of inputting the temporal feature vector into the trained LSTM-Transformer model for spatiotemporal prediction, obtaining the trend-seasonal point value prediction result for each representative spatial point, includes: The input module receives the time-series feature vector and preprocesses it to obtain a unified dimension vector. The LSTM module is used to extract features from the unified dimension vector to analyze the long-term dependencies in the unified dimension vector. The transmission and forgetting of information are dynamically controlled through a gating mechanism to obtain multivariate time-series features. The Transformer module extracts features from the multivariate temporal features through a multi-head attention mechanism to analyze the correlation between different spatial locations and transform the multivariate temporal features into a spatiotemporal fusion code with high-order spatiotemporal features. The output module performs nonlinear transformation and dimensional adjustment on the spatiotemporal fusion code to obtain the trend-seasonal point value prediction results for each representative spatial point.

4. The method according to claim 3, characterized in that, The LSTM module includes cascaded multi-layer bidirectional LSTM layers, each of which includes multiple LSTM units; each LSTM unit includes a cell state, a forget gate, an input gate, and an output gate. 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 cell state with the input information at time t. The output gate controls the output of the cell state.

5. The method according to claim 3, characterized in that, The Transformer module includes: a location encoding layer, a multi-head self-attention layer, and a feedforward network layer; the Transformer module extracts features from the multivariate temporal features through a multi-head attention mechanism to analyze the correlation between different spatial locations, and transforms the multivariate temporal features into a spatiotemporal fusion encoding with high-order spatiotemporal features, including: Through the location encoding layer, temporal and spatial location information are encoded into vectors and embedded into the multivariate temporal features to obtain sequential encoding features; Through the multi-head self-attention layer, the sequential encoding features are dynamically weighted and fused in multiple modes. The residual connection preserves the original information and the layer normalization optimizes the gradient propagation to generate global association features that represent global spatiotemporal dependence. The global correlation features are subjected to two fully connected projections and nonlinear activations through the feedforward network layer. The original input and the layer-normalized stable feature distribution are then fused through residual connections to output the spatiotemporal fusion code.

6. The method according to claim 1, characterized in that, The method further includes: Framework for building the initial LSTM-Transformer model; Adaptive gradient adjustment is performed using the Adam optimizer, which dynamically adjusts the update step size of the model parameters based on a preset learning rate to optimize the learning process of the initial model. Historical datasets were obtained from multiple data sources, including SIF variables, second meteorological variables, second vegetation physiological state variables, second soil variables, and second human activity variables. The historical dataset is cyclically split into three subsets using triple cross-validation, and three rounds of training-validation are performed. In each round, the initial model is trained with two subsets and the generalization of the initial model is validated with one subset. The initial model's hyperparameter combinations are traversed using GridSearchCV to optimize these combinations, which include the number of hidden units, the number of layers, the learning rate, and the Dropout ratio.

7. The method according to claim 1, characterized in that, The method further includes: The actual SIF change sequence of the target prediction region is obtained, and the predicted SIF value in the large-scale spatiotemporal prediction map and the real SIF value in the actual SIF change sequence are compared point by point to obtain the evaluation index of the trained LSTM-Transformer model. The evaluation index includes: coefficient of determination, adjusted coefficient of determination, root mean square error, mean absolute percentage error, explained variance and mean directional accuracy. Based on the evaluation metrics, the performance of the trained LSTM-Transformer model is evaluated, and the evaluation results are obtained. Based on the evaluation results, the trained LSTM-Transformer model is updated.

8. The method according to claim 1, characterized in that, The preprocessing of the original dataset includes: geometric correction, projection stitching, resampling, maximum value synthesis, and missing value imputation.

9. A solar-induced chlorophyll fluorescence spatiotemporal prediction device, characterized in that, include: The module comprises a data acquisition module, a partitioning module, a feature construction module, a prediction module, and a spatial interpolation module; among which: The data acquisition module is used to acquire raw datasets from multiple data sources. The raw datasets include a first meteorological variable, a first vegetation physiological state variable, a first soil variable, and a first human activity variable in a large-scale region. The raw datasets are preprocessed to generate a spatiotemporal dataset in a unified format. The partitioning module is used to perform surface aggregation based on the spatiotemporal dataset, taking the county-level administrative units in the large-scale region as the reference unit, and combining empirical wavelet decomposition and attention mechanism to divide the large-scale region into multiple homogeneous sub-regions. Each of the homogeneous sub-regions has similar environmental features, and representative spatial points are selected from each of the sub-regions. The feature construction module is used to extract variable feature information of key covariates related to the change of solar-induced chlorophyll fluorescence (SIF) at representative spatial points for each sub-region, and to construct a four-dimensional spatiotemporal cube containing time and space domains based on the variable feature information of each representative spatial point. Each representative spatial point in the four-dimensional spatiotemporal cube corresponds to variable feature information at a specific spatial location and at a specific time point. The prediction module is used to perform STL decomposition on the temporal data of each spatiotemporal unit in the four-dimensional spatiotemporal cube to obtain a trend term, a seasonal term, and a residual term. The trend term, the seasonal term, and the residual term are flattened to obtain a corresponding temporal feature vector. The temporal feature vector is then input into a trained LSTM-Transformer model for spatiotemporal prediction to obtain the trend-seasonal point value prediction result for each representative spatial point. The spatial interpolation module is used to interpolate the point value prediction results and preset covariates to a preset target prediction region using the ANUSPLIN thin-plate smooth spline interpolation method, thereby generating a large-scale SIF spatiotemporal prediction map with a smooth and continuous surface.

Images