High-precision land carbon flux reconstruction method based on four-dimensional space-time deep learning

Abstract

The invention relates to a high-precision land carbon flux reconstruction method based on four-dimensional space-time deep learning. The method comprises the steps of obtaining a multi-source heterogeneous data set and performing preprocessing; constructing a four-dimensional space-time deep learning model architecture which is sequentially connected with a multi-scale space-time feature extraction module, a space-time feature fusion module, a prediction layer and an uncertainty quantization layer from input to output, and taking the preprocessed multi-source heterogeneous data set as input and a carbon flux reconstruction result as output; constructing a loss function, adopting a hierarchical learning rate strategy and introducing a cosine annealing scheduler, and training a four-dimensional space-time deep learning model architecture based on the preprocessed multi-source heterogeneous data set to obtain a four-dimensional space-time deep learning model; and obtaining a carbon flux data reconstruction set, and inputting the four-dimensional space-time deep learning model to generate a carbon flux reconstruction result. By adopting the method, multi-source heterogeneous data can be effectively integrated, the complex spatial-temporal dynamic characteristics of the carbon flux are captured, and accurate reconstruction of the land carbon flux with high spatial-temporal resolution is realized.

Description

Technical Field

[0001] This application relates to the field of carbon cycle and climate change monitoring technology, and in particular to a high-precision method for reconstructing terrestrial carbon flux based on four-dimensional spatiotemporal deep learning. Background Technology

[0002] Terrestrial ecosystem carbon flux is a key component of the global carbon cycle. Accurate reconstruction and estimation of terrestrial carbon flux are crucial for understanding global climate change mechanisms, formulating emission reduction policies, and achieving "dual carbon" goals. In recent years, with the development of remote sensing technology, a new generation of satellites has provided higher-resolution observational data. This high-resolution remote sensing data offers new possibilities for overcoming the resolution limitations of current carbon flux estimation. Remote sensing data-driven machine learning methods have been widely used in recent years, estimating regional-scale carbon flux by establishing statistical relationships between remote sensing indices and carbon flux. However, traditional machine learning methods (such as random forests and support vector machines) mainly build models based on a single spatial or temporal dimension, neglecting the spatiotemporal coupling characteristics of carbon flux and making it difficult to capture the nonlinear variation patterns of carbon flux. Furthermore, traditional terrestrial carbon flux data products exhibit a clear scale trade-off in terms of spatiotemporal resolution. Typical products such as MODIS GPP / NPP (MOD17) provide 8-day, 500-meter resolution data, FLUXCOM provides daily-scale 0.5° grid carbon flux data, and GLASS GPP provides 8-day, 1-kilometer resolution total vegetation primary productivity datasets. While some products achieve daily temporal scales, their spatial resolution often decreases to several kilometers or even coarser. This inherent trade-off in spatiotemporal resolution limits the ability to characterize the fine spatiotemporal heterogeneity of ecosystem carbon fluxes. In particular, under the influence of rapidly changing landscape patterns and short-term climate events, medium- to high-resolution dynamic simulations of carbon fluxes still face significant challenges.

[0003] However, effectively utilizing this high-resolution data remains a challenge. Traditional methods have limitations in processing multi-source, multi-scale remote sensing data, making it difficult to fully extract the spatiotemporal information contained within the data. Simultaneously, the spatiotemporal dynamics of carbon flux are complexly influenced by various environmental factors, requiring more advanced modeling methods to capture these intricate relationships. Therefore, there is an urgent need to develop carbon flux reconstruction methods that can fully utilize four-dimensional spatiotemporal information and fuse multi-source high-resolution remote sensing data to improve the accuracy and reliability of carbon flux estimation at regional and global scales, providing data support and scientific basis for climate change research and the achievement of dual-carbon goals. Summary of the Invention

[0004] Therefore, it is necessary to provide a high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning that can improve the accuracy and reliability of carbon flux estimation, in order to address the above-mentioned technical problems.

[0005] This application provides a high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning, which executes steps S1-S3 to generate a four-dimensional spatiotemporal deep learning model, captures the complex spatiotemporal dynamic features of carbon flux, and achieves accurate land carbon flux reconstruction with high spatiotemporal resolution, including: Step S1: Obtain multi-source heterogeneous datasets, including flux observation network data, satellite remote sensing data, and meteorological reanalysis data, and preprocess the multi-source heterogeneous datasets. The preprocessing includes spatiotemporal matching and feature construction of the multi-source heterogeneous datasets and updating the multi-source heterogeneous datasets. Step S2: Construct a four-dimensional spatiotemporal deep learning model architecture that connects the multi-scale spatiotemporal feature extraction module, the spatiotemporal feature fusion module, the prediction layer, and the uncertainty quantification layer sequentially from input to output. The four-dimensional spatiotemporal deep learning model architecture takes the preprocessed multi-source heterogeneous dataset as input and the carbon flux reconstruction result as output. Step S3: Construct temporal scale loss function, spatial scale loss function, and total loss function. Adopt a hierarchical learning rate strategy and introduce a cosine annealing scheduler. Train the four-dimensional spatiotemporal deep learning model architecture based on the preprocessed multi-source heterogeneous dataset to obtain the four-dimensional spatiotemporal deep learning model. Obtain the carbon flux data reconstruction set, input it into the four-dimensional spatiotemporal deep learning model to generate carbon flux reconstruction results, and use a multi-scale verification strategy to verify the carbon flux reconstruction results and perform uncertainty assessment.

[0006] In one embodiment, step S1, spatiotemporal matching of multi-source heterogeneous datasets, includes: The bilinear interpolation method is used to uniformly resample the spatial resolution of each data source in the multi-source heterogeneous dataset to the target grid resolution. The bilinear interpolation method is as follows: ; in, For target grid points The interpolated value of the position. and The target grid points are respectively The original data value and weight coefficient corresponding to the m-th sampling point selected as the center, where n is the total number of sampling points participating in the interpolation calculation; The cubic spline interpolation method is used to unify the temporal resolution of each data point in the multi-source heterogeneous dataset to a daily level. The cubic spline interpolation method is as follows: ; Where S(t) is the interpolation function value at time point t. Let be the starting time point of the interval, and let a, b, c, and d be the interpolation coefficients determined by the spline conditions, representing the coefficients of the constant term, linear term, quadratic term, and cubic term, respectively.

[0007] In one embodiment, step S1, feature construction of the multi-source heterogeneous dataset, includes constructing vegetation physiological features, meteorological features, and terrain features from the spatiotemporally matched multi-source heterogeneous dataset, wherein: Vegetation physiological characteristics were constructed using satellite remote sensing data from a multi-source heterogeneous dataset. The NDVI vegetation index was calculated using the following formula: ; in, R nir The reflectance is 842 nm in the near-infrared band. R red The reflectance at 665 nm in the red band; The EVI vegetation index is calculated using the following formula: ; in, R blue The reflectivity at 490nm in the blue light band; Meteorological feature construction includes constructing daily time series features of temperature, precipitation, solar radiation, air pressure, and wind speed; Terrain feature construction includes calculating slope and aspect based on multi-source heterogeneous datasets, using the following formula: ; Where Z represents the elevation value in the multi-source heterogeneous dataset. and are the partial derivatives of elevation in the x and y directions, respectively, representing the gradient change of the terrain.

[0008] In one embodiment, the multi-scale spatiotemporal feature extraction module of the four-dimensional spatiotemporal deep learning model architecture in step S2 includes a spatial feature extraction branch and a temporal feature extraction branch. The inputs of the spatial feature extraction branch and the temporal feature extraction branch serve as the two inputs of the multi-scale spatiotemporal feature extraction module, respectively. The outputs of the spatial feature extraction branch and the temporal feature extraction branch are jointly connected to the input of the spatiotemporal feature fusion module as the output of the multi-scale spatiotemporal feature extraction module. The spatial feature extraction branch uses a multi-scale 3D convolutional neural network to extract spatial features, as shown in the following equation: ; in, F s Let X be the spatial feature, W be the preprocessed multi-source heterogeneous dataset used as input features, and W be the convolution kernel weights. b Here, represents the hidden layer bias vector; BN stands for batch normalization. It is a non-linear activation function; It is a three-dimensional convolution function; The temporal feature extraction branch combines a long short-term memory network and a temporal convolutional network to simultaneously construct short-term fluctuation models and long-term trend models to extract temporal features. The long short-term memory network is used to capture long-term dependencies as follows: ; in, h t Let be the hidden state vector at time t. h t-1 Let be the hidden state vector from the previous time step. X t Let W be the input feature vector at time t. h This is the weight matrix from hidden state to hidden state. W x Let σ be the weight matrix input to the hidden state, and σ be the activation function. Using a temporal convolutional network to capture local temporal patterns is as follows: ; in, F t The time characteristic is represented by d, where d is the inflation rate. Let X be a one-dimensional convolution function, and let X be a preprocessed multi-source heterogeneous dataset used as input features.

[0009] In one embodiment, the spatiotemporal feature fusion module of the four-dimensional spatiotemporal deep learning model architecture in step S2 adopts a dual-stream spatiotemporal feature fusion network. The input of the spatiotemporal feature fusion module is connected to the output of the multi-scale spatiotemporal feature extraction module, and the output of the spatiotemporal feature fusion module is connected to the input of the prediction layer. The features extracted by the spatial feature extraction branch and the temporal feature extraction branch are adaptively integrated through a gating mechanism, as shown in the following equation: ; in, As weight; The spatiotemporal feature fusion module uses an adaptive feature fusion layer to dynamically adjust the weights of spatial and temporal features.

[0010] In one embodiment, the prediction layer of the four-dimensional spatiotemporal deep learning model architecture in step S2 adopts a multi-task learning framework, simultaneously predicting net ecosystem exchange (NEE), gross primary productivity (GPP), and ecosystem respiration (Re). A physical constraint layer is introduced to ensure the carbon balance relationship NEE = Re - GPP. The input of the prediction layer is connected to the output of the spatiotemporal feature fusion module, and the output of the prediction layer is connected to the input of the uncertainty quantification layer, as shown in the following equation: ; in, This is the physical constraint loss function.

[0011] In one embodiment, the uncertainty quantization layer of the four-dimensional spatiotemporal deep learning model architecture in step S2 uses Monte Carlo random deactivation technique to quantify the uncertainty of the model prediction. The input of the uncertainty quantization layer is connected to the output of the prediction layer, as shown in the following equation: ; in, To predict the mean, This represents the output of the four-dimensional spatiotemporal deep learning model architecture after the i-th random deactivation, where N is the total number of Monte Carlo samplings.

[0012] In one embodiment, the time-scale loss function is constructed in step S3 as follows: ; The spatial scale loss function is constructed as follows: ; in, This is the output of a four-dimensional spatiotemporal deep learning model architecture under real-world conditions. This is the result of predicting the output of a four-dimensional spatiotemporal deep learning model architecture under ideal conditions; T is the total number of steps within the cosine variation period, t is the current step; N is the total number of Monte Carlo samplings; The total loss function is constructed as follows: ; in, The weights of the time-scale loss function, The weights of the spatial scale loss function are... The weights of the physical constraint loss function, This is the physical constraint loss function.

[0013] In one embodiment, the hierarchical learning rate strategy used in step S3 is as follows: ; in, The initial learning rate, As the attenuation factor, For the first l The depth function of the layer, l is the index of each trainable layer in the four-dimensional spatiotemporal deep learning model architecture, with a value range from 1 to L, where L is the total number of all trainable layers in the four-dimensional spatiotemporal deep learning model architecture. The cosine annealing scheduler is introduced as follows: ; in, Let be the learning rate at the t-th iteration. To minimize the learning rate, This represents the maximum learning rate.

[0014] In one embodiment, step S4 employs a multi-scale verification strategy to verify the carbon flux reconstruction results, including site-scale verification and regional-scale verification, wherein the site-scale verification is as follows: ; Wherein, RMSE is the root mean square error. This is the output of a four-dimensional spatiotemporal deep learning model architecture under real-world conditions. To predict the output of a four-dimensional spatiotemporal deep learning model architecture under ideal conditions, N is the total number of Monte Carlo samplings; The regional scale verification is as follows: ; CI stands for Consistency Index. and This is the output of a four-dimensional spatiotemporal deep learning model architecture under real-world conditions. It is the average of N sample results.

[0015] The aforementioned high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning constructs a four-dimensional spatiotemporal deep learning model architecture. It combines a multi-scale three-dimensional convolutional neural network and a long short-term memory network with a temporal convolutional network to simultaneously capture the spatial structural features and temporal dynamic changes of carbon flux. A dual-stream spatiotemporal feature fusion network and an adaptive feature fusion layer are employed to achieve dynamic weighted integration of spatial and temporal features, improving the model's adaptability to carbon flux characteristics in different regions and time periods. A multi-scale loss function design method is developed to simultaneously optimize model performance at multiple temporal and spatial scales, and a physical constraint loss term ensures the carbon balance relationship. A deep learning framework incorporating spatial and temporal self-attention mechanisms is constructed, enabling adaptive capture of long-distance dependencies between different regions and identification of key time nodes. An uncertainty quantification layer is designed, using Monte Carlo random inactivation techniques and an error propagation model to assess the uncertainty of the model's prediction results, thereby improving the reliability of the results. Through the above methods, this application can effectively integrate multi-source heterogeneous data, capture the complex spatiotemporal dynamic characteristics of carbon flux, and achieve accurate reconstruction of land carbon flux with high spatiotemporal resolution. It overcomes the technical problems of insufficient spatial representativeness, large simulation bias, and insufficient capture of spatiotemporal coupling characteristics in traditional technologies, and provides more accurate and reliable carbon flux data support for global climate change research, refined weather forecasting, and environmental monitoring. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning in one embodiment. Detailed Implementation

[0017] Terrestrial ecosystem carbon flux is a key component of the global carbon cycle. Accurate reconstruction and estimation of terrestrial carbon flux are crucial for understanding global climate change mechanisms, formulating emission reduction policies, and achieving the "dual carbon" goal. Currently, methods for estimating terrestrial carbon flux mainly include direct measurement via flux observation networks, process model simulation, and remote sensing inversion. Biogeochemical process models can simulate the spatiotemporal variations of terrestrial ecosystem carbon flux, but due to limitations in parameterization schemes and model structure uncertainties, simulation results often deviate significantly from actual observations. Particularly under the influence of extreme climate events, the models have limited ability to simulate the response of ecosystem carbon flux disturbances, making it difficult to accurately characterize the temporal dynamics of carbon flux. Remote sensing data-driven machine learning methods have been widely used in recent years, estimating regional-scale carbon flux by establishing statistical relationships between remote sensing indices and carbon flux. However, traditional machine learning methods (such as random forests and support vector machines) primarily build models based on a single spatial or temporal dimension, neglecting the spatiotemporal coupling characteristics of carbon flux and failing to capture the nonlinear variation patterns of carbon flux. Furthermore, traditional terrestrial carbon flux data products exhibit a clear scale trade-off in terms of spatiotemporal resolution. Typical products such as MODIS GPP / NPP (MOD17) provide 8-day, 500-meter resolution data, FLUXCOM provides daily 0.5° grid carbon flux data, and GLASS GPP provides 8-day, 1-kilometer resolution total vegetation primary productivity datasets. While some products achieve daily temporal resolution, their spatial resolution is often correspondingly reduced to several kilometers or even coarser. This inherent trade-off in spatiotemporal resolution limits the ability to characterize the fine spatiotemporal heterogeneity of ecosystem carbon fluxes. In particular, under the influence of rapidly changing landscape patterns and short-term climate events, medium- to high-resolution dynamic simulations of carbon fluxes still face significant challenges.

[0018] In recent years, with the development of remote sensing technology, a new generation of satellites has provided higher-resolution observational data. Among them, the Sentinel-2 satellite continuously monitors global land surface changes with a high spatial resolution of 10–60 meters, significantly improving the ability to observe the dynamic processes of surface vegetation and ecosystems. This high-resolution remote sensing data offers new possibilities for overcoming the resolution limitations of current carbon flux estimation. However, effectively utilizing this high-resolution data still faces challenges. Existing methods have limitations in processing multi-source, multi-scale remote sensing data, making it difficult to fully extract the spatiotemporal information contained within the data. Simultaneously, the spatiotemporal dynamics of carbon flux are complexly influenced by various environmental factors, requiring more advanced modeling methods to capture these complex relationships. Therefore, there is an urgent need to develop carbon flux reconstruction methods that can fully utilize four-dimensional spatiotemporal information and fuse multi-source high-resolution remote sensing data to improve the accuracy and reliability of carbon flux estimation at regional and global scales, providing data support and scientific basis for climate change research and the achievement of "dual carbon" goals.

[0019] Therefore, in view of the technical problems of traditional land carbon flux estimation methods, such as insufficient spatial representativeness due to sparse site distribution, simulation bias caused by parameterization simplification of process models, neglect of spatiotemporal coupling characteristics by traditional machine learning methods, and limited simulation capability for response to extreme climate events, this application aims to propose a high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning.

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] In one exemplary embodiment, such as Figure 1 As shown, a high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning is provided. This method fuses multi-source data through deep learning technology to achieve high-precision spatiotemporal reconstruction of land carbon flux. The high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning provided by this invention includes the following steps: Step S1: Obtain multi-source heterogeneous datasets, including flux observation network data, satellite remote sensing data, and meteorological reanalysis data, and preprocess the multi-source heterogeneous datasets. The preprocessing includes spatiotemporal matching and feature construction of the multi-source heterogeneous datasets and updating the multi-source heterogeneous datasets. For example, this embodiment includes multi-source heterogeneous data collection and preprocessing. First, multi-source heterogeneous data such as flux observation network data, satellite remote sensing data, and meteorological reanalysis data are acquired. Missing value processing, outlier detection, and filtering are performed on the flux observation data, and quantile normalization is used to eliminate systematic errors caused by instrument and environmental factors.

[0022] Furthermore, bilinear interpolation was used to resample satellite remote sensing data and meteorological reanalysis data at different spatial resolutions to a target resolution grid (10 m × 10 m), and temporal interpolation was used to unify data at different temporal resolutions to the same time step (daily). Multi-scale spatiotemporal features were constructed, including vegetation physiological characteristics (NDVI, EVI, etc.), meteorological characteristics (temperature, precipitation, solar radiation, wind speed, air pressure, humidity), land use types, temporal characteristics (seasonal indicators, year), and topographic features (elevation, slope, aspect). NDVI is the Normalized Difference Vegetation Index, and EVI is the Enhanced Vegetation Index. Step S2: Construct a four-dimensional spatiotemporal deep learning model architecture that connects the multi-scale spatiotemporal feature extraction module, the spatiotemporal feature fusion module, the prediction layer, and the uncertainty quantification layer sequentially from input to output. The four-dimensional spatiotemporal deep learning model architecture takes the preprocessed multi-source heterogeneous dataset as input and the carbon flux reconstruction result as output. For example, this embodiment includes the construction of a four-dimensional spatiotemporal deep learning model. A four-dimensional spatiotemporal deep learning model architecture is designed and trained to obtain a model capable of simultaneously capturing the spatial structural features and temporal dynamic changes of carbon flux. The model includes a multi-scale spatiotemporal feature extraction module, a spatiotemporal feature fusion module, a prediction layer, and an uncertainty quantization layer. In the multi-scale spatiotemporal feature extraction module, the spatial feature extraction branch employs a multi-scale three-dimensional convolutional neural network (3D-CNN) to capture spatial dependence patterns of carbon flux at different spatial scales using convolutional kernels of different sizes. The temporal feature extraction branch combines a long short-term memory network (LSTM) and a temporal convolutional network (TCN) to balance modeling long-term dependencies with short-term dynamic changes, improving the ability to model multi-scale change patterns in time series. A spatial self-attention mechanism is introduced to adaptively capture long-distance dependencies between different regions. A temporal self-attention mechanism is also introduced to identify key time nodes at different time scales. The spatiotemporal feature fusion module employs a dual-stream spatiotemporal feature fusion network, adaptively integrating features extracted from spatial and temporal branches through a gating mechanism. An adaptive feature fusion layer dynamically adjusts the weights of spatial and temporal features based on the characteristics of different regions and time periods. The prediction layer utilizes a multi-task learning framework to simultaneously predict Net Ecosystem Exchange (NEE), Gross Primary Productivity (GPP), and Ecosystem Respiration (Re). A physical constraint layer is introduced to ensure the carbon balance relationship of NEE = Re - GPP. The uncertainty quantification layer employs Monte Carlo random inactivation technology, generating a prediction distribution through multiple forward propagations to quantify the uncertainty of the model predictions. An error propagation model is constructed to assess the impact of input data uncertainty on the prediction results.

[0023] Step S3: Construct temporal scale loss function, spatial scale loss function, and total loss function. Adopt a hierarchical learning rate strategy and introduce a cosine annealing scheduler. Train the four-dimensional spatiotemporal deep learning model architecture based on the preprocessed multi-source heterogeneous dataset to obtain the four-dimensional spatiotemporal deep learning model. Obtain the carbon flux data reconstruction set, input it into the four-dimensional spatiotemporal deep learning model to generate carbon flux reconstruction results, and use a multi-scale verification strategy to verify the carbon flux reconstruction results and perform uncertainty assessment.

[0024] For example, this embodiment includes model training and optimization strategies. Regarding the design of the multi-scale loss function, the time-scale loss calculates the mean square error between predicted and observed values ​​at four time scales: hourly, daily, monthly, and yearly. The spatial-scale loss calculates the consistency between predicted and reference data at the station and regional scales. The physical constraint loss incorporates a carbon balance constraint term to ensure the physical relationship between NEE, GPP, and Re. The total loss function can adaptively and weightedly combine the above losses. Regarding the optimization strategy, a hierarchical learning rate strategy is adopted, setting different learning rates for different network layers; a learning rate warm-up and cosine annealing scheduler are introduced to improve model convergence; gradient accumulation technology is used to process large amounts of data; and an early stopping strategy is implemented to avoid overfitting. Regarding transfer learning and pre-training, large-scale meteorological reanalysis data is used for model pre-training; and domain adaptation technology is used to improve the model's generalization ability in sparse data regions.

[0025] Furthermore, the carbon flux data reconstruction set includes remote sensing data (Landsat 5 / 7 / 8 / 9) image series (1986-2024), 30-meter resolution; natural factors, including climate (precipitation, temperature, evapotranspiration), topography (elevation, slope, etc.); and human factors, including GDP, population, human footprint, and land use. All data were preprocessed to maintain spatiotemporal consistency.

[0026] Furthermore, this embodiment also includes high-precision spatiotemporal reconstruction and validation of carbon flux. Multi-scale carbon flux reconstruction, based on a trained four-dimensional spatiotemporal deep learning model, generates high-resolution (10 m × 10 m, daily) carbon flux products globally or regionally, i.e., the carbon flux reconstruction results; and constructs a long-term series (2000 to present) carbon flux dataset. The comprehensive validation strategy includes site-scale validation (using reserved flux site data for cross-validation), regional-scale validation, global-scale validation (comparing with estimated global carbon balance products), and extreme event response validation (specifically evaluating the model's predictive performance during extreme events such as droughts and heat waves). This embodiment employs site-scale and regional-scale validation. Regarding uncertainty assessment, the range of uncertainty in the model's predictions is quantified; the spatial distribution and temporal variation characteristics of uncertainty are analyzed; and the relationship between uncertainty and environmental factors is evaluated.

[0027] This application constructs a multi-scale spatiotemporal feature extraction and fusion framework to effectively integrate multi-source heterogeneous data, thereby capturing the complex spatiotemporal dynamic characteristics of carbon flux, achieving accurate reconstruction of land carbon flux with high spatiotemporal resolution, improving the accuracy and reliability of carbon flux estimation, and providing more precise data support for climate change research and environmental monitoring.

[0028] In an exemplary embodiment, step S1, multi-source heterogeneous data acquisition and preprocessing, includes multi-source heterogeneous data collection, data preprocessing, and spatiotemporal matching. Flux observation data, including daily net ecosystem exchange (NEE), gross primary productivity (GPP), and ecosystem respiration (Re), are acquired through a flux observation network; satellite remote sensing data, specifically 10 m × 10 m spatial resolution Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and 10 m × 10 m spatial resolution land use type data retrieved from the Sentinel-2 satellite, are also acquired. The data includes 30 m × 30 m spatial resolution digital elevation model (DEM) data provided by the SRTM (Shuttle Radar Topography Mission), and slope and aspect information are calculated based on the DEM; the meteorological reanalysis data includes 0.25° × 0.25° daily temperature, precipitation, solar radiation, air pressure, and wind speed provided by the ERA5 (ECMWF Reanalysis 5th Generation) dataset.

[0029] Furthermore, data preprocessing and spatiotemporal matching include missing value handling and outlier detection, spatial resampling (10m×10 m), temporal resampling (daily), and multi-scale spatiotemporal feature construction, specifically including spatiotemporal matching of multi-source data, feature construction, and feature standardization.

[0030] In an exemplary embodiment, the spatiotemporal matching of the multi-source heterogeneous dataset in step S1 includes using bilinear interpolation to unify the spatial resolution to a 10 m × 10 m grid: ; in, For target grid points The interpolated value of the position. and The target grid points are respectively The original data value and weight coefficient corresponding to the m-th sampling point selected as the center, where n is the total number of sampling points participating in the interpolation calculation; The time resolution is unified to daily using cubic spline interpolation. ; Where S(t) is the interpolation function value at time point t. Let be the starting time point of the interval, and let a, b, c, and d be the interpolation coefficients determined by the spline conditions, representing the coefficients of the constant term, linear term, quadratic term, and cubic term, respectively.

[0031] In an exemplary embodiment, the feature construction of the multi-source heterogeneous dataset in step S1 includes constructing vegetation physiological features, meteorological features, and terrain features from the spatiotemporally matched multi-source heterogeneous dataset, wherein: The vegetation physiological characteristics were constructed using NDVI and EVI vegetation indices calculated from Sentinel-2 (Level-2A) satellite remote sensing data within a multi-source heterogeneous dataset. This dataset has a 10-meter spatial resolution and a 5-day temporal resolution. The NDVI vegetation index calculation formula is as follows: ; Among them, R nir The reflectance is in the near-infrared band (842 nm), corresponding to the B8 band of Sentinel-2 data; R red The reflectance is for the red band (665 nm), corresponding to the B4 band of the Sentinel-2 data.

[0032] The formula for calculating the Strong Vegetation Index (EVI) is: ; Among them, R blue The reflectance is for the blue light band (490 nm), corresponding to the B2 band of Sentinel-2 data (center wavelength 490 nm).

[0033] The meteorological feature construction includes the construction of daily temperature, precipitation, solar radiation, air pressure, and wind speed time series features.

[0034] Terrain feature construction includes calculating slope and aspect based on DEM data, as shown in the following formula: ; Where Z represents the elevation value in the multi-source heterogeneous dataset. and are the partial derivatives of elevation in the x and y directions, respectively, representing the gradient change of the terrain.

[0035] In an exemplary embodiment, feature standardization involves normalizing all features, as shown in the following formula: ; Where X represents the original feature values, namely vegetation physiological features, meteorological features, and topographic features; μ is the mean of the normalized feature, and σ is the standard deviation of the normalized feature. X norm These are the standardized feature values.

[0036] In an exemplary embodiment, the construction of the four-dimensional spatiotemporal deep learning model in step S2 includes spatial feature extraction, temporal feature extraction, spatiotemporal feature fusion network and gating mechanism, multi-task prediction layer and physical constraint layer, and uncertainty quantification layer.

[0037] In one exemplary embodiment, spatial feature extraction includes extracting spatial features using a multi-scale 3D convolutional neural network (3D-CNN), as shown in the formula: ; in, F s X is the spatial feature map; W is the input feature, i.e., the preprocessed multi-source heterogeneous dataset; b is the convolution kernel weight; b is the bias term, i.e., the hidden layer bias vector; BN is batch normalization. It is a non-linear activation function; It is a three-dimensional convolution function.

[0038] Furthermore, temporal feature extraction includes using LSTM to capture long-term dependencies and using a Temporal Convolutional Network (TCN) to capture local temporal patterns, where the use of LSTM to capture long-term dependencies is as follows: ; in, h t Let be the hidden state vector at time t. h t-1 Let be the hidden state vector from the previous time step. X t Let W be the input feature vector at time t. h This is the weight matrix from hidden state to hidden state. W x The weight matrix is ​​input to the hidden state. b Let σ be the hidden layer bias vector, and σ be the activation function.

[0039] Using a Temporal Convolutional Network (TCN) to capture local temporal patterns is as follows: ; in, F td is the time feature, X is the dilation rate, W is the input feature (i.e., the preprocessed multi-source heterogeneous dataset), b is the convolution kernel weight, b is the bias term (i.e., the hidden layer bias vector), and BN is batch normalization. For nonlinear activation functions It is a one-dimensional convolution function.

[0040] In an exemplary embodiment, spatiotemporal feature fusion includes fusing spatial and temporal features through a gating mechanism, as follows: ; in, F s For spatial features, F t As a time feature, For weights.

[0041] In an exemplary embodiment, multi-task prediction includes simultaneously predicting NEE, GPP, and Re, combined with a physical constraint layer, as follows: ; in, That is Physical constraint loss function.

[0042] In one exemplary embodiment, uncertainty quantification includes using a Monte Carlo random deactivation technique, as follows: ; in, To predict the mean, This represents the output of the four-dimensional spatiotemporal deep learning model architecture after the i-th random deactivation, where N is the total number of Monte Carlo samplings. i The value ranges from 1 to N.

[0043] In an exemplary embodiment, the multi-scale spatiotemporal feature extraction module of the four-dimensional spatiotemporal deep learning model architecture includes a spatial feature extraction branch and a temporal feature extraction branch. The inputs of the spatial feature extraction branch and the temporal feature extraction branch serve as the two inputs of the multi-scale spatiotemporal feature extraction module, respectively. The outputs of the spatial feature extraction branch and the temporal feature extraction branch together serve as the output of the multi-scale spatiotemporal feature extraction module and are connected to the input of the spatiotemporal feature fusion module. The input of the spatiotemporal feature fusion module is connected to the output of the multi-scale spatiotemporal feature extraction module, and the output of the spatiotemporal feature fusion module is connected to the input of the prediction layer. The input of the prediction layer is connected to the output of the spatiotemporal feature fusion module, and the output of the prediction layer is connected to the input of the uncertainty quantization layer.

[0044] In an exemplary embodiment, model training and optimization in step S3 includes multi-scale loss function design, optimization strategy (hierarchical learning rate), transfer learning, and domain adaptation.

[0045] Specifically, the design of multi-scale loss functions includes temporal scale loss, spatial scale loss, and total loss function, where: The time-scale loss is as follows: ; The spatial scale loss function is constructed as follows: ; in, This is the output of a four-dimensional spatiotemporal deep learning model architecture under real-world conditions. This represents the output of a four-dimensional spatiotemporal deep learning model architecture under ideal conditions; T is the total number of steps within the cosine variation period, t is the current step; N is the total number of Monte Carlo samplings. i The value ranges from 1 to N; The total loss function is constructed as follows: ; in, The weights of the time-scale loss function, The weights of the spatial scale loss function are... The weights of the physical constraint loss function, This is the physical constraint loss function.

[0046] In one exemplary embodiment, the optimization strategy includes: Using a hierarchical learning strategy, as shown below: ; in, The initial learning rate, As the attenuation factor, For the first l The depth function of the layer, l This represents the index of each trainable layer in the four-dimensional spatiotemporal deep learning model architecture, with a value ranging from 1 to L, where L is the total number of trainable layers in the four-dimensional spatiotemporal deep learning model architecture.

[0047] Introducing a cosine annealing scheduler, as shown in the following equation: ; in, Let be the learning rate at the t-th iteration. To minimize the learning rate, The maximum learning rate is T, where T is the total number of steps within the cosine variation period, and t is the current step.

[0048] In an exemplary embodiment, step S3 is followed by carbon flux reconstruction and verification, namely, high-precision spatiotemporal reconstruction of carbon flux data. The carbon flux reconstruction result generated by the four-dimensional spatiotemporal deep learning model is a high-resolution carbon flux product with a spatial resolution of 10 m × 10 m and a temporal resolution of daily.

[0049] Furthermore, the multi-scale validation strategy includes site-scale validation and region-scale validation. Site-scale validation includes calculating the root mean square error (RMSE), as follows: ; Wherein, RMSE is the root mean square error. This is the output of a four-dimensional spatiotemporal deep learning model architecture under real-world conditions. To predict the output of a four-dimensional spatiotemporal deep learning model architecture under ideal conditions, N represents the total number of samples, which can also be understood as the total number of Monte Carlo samplings. i The value ranges from 1 to N; Regional-scale verification involves calculating the Consistency Index (CI), as shown in the following formula: ; CI stands for Consistency Index. and This is the output of a four-dimensional spatiotemporal deep learning model architecture under real-world conditions. It is the average of N sample results (sampling results).

[0050] In terms of uncertainty assessment, the range of uncertainty predicted by the model is quantified; the spatial distribution and temporal variation characteristics of uncertainty are analyzed; and the relationship between uncertainty and environmental factors is assessed.

[0051] The high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning proposed in this application has the following significant advantages over traditional techniques. First, it improves the spatial resolution and temporal accuracy of carbon flux estimation. The method generates carbon flux products with a spatial resolution of 10 m × 10 m and a daily temporal resolution. Compared to point-scale observations in traditional flux observation networks and coarse-resolution simulations in traditional process models (typically 500 m × 500 m or coarser), the spatial resolution is improved by 50–100 times, and the temporal continuity is significantly enhanced. The spatial detail capture capability of the method in this application is significantly superior to traditional methods, reflecting the spatial heterogeneity of carbon flux caused by small-scale topographic and vegetation differences. Second, it achieves efficient fusion of multi-source heterogeneous data. The dual-stream spatiotemporal feature fusion network and adaptive feature fusion layer in this application can effectively integrate multi-source data with different sources and characteristics, reconstructing carbon flux based on data fusion. Experimental results show that compared to models with a single data source, the accuracy of the model after multi-source data fusion is significantly improved, especially in data-sparse regions, where the prediction performance improvement is more pronounced. Finally, this application provides a reliable uncertainty assessment mechanism. By employing Monte Carlo random inactivation techniques and error propagation models, the method presented in this application can quantify the uncertainty in carbon flux estimation results. The uncertainty range provided by this application covers the vast majority of actual observations, and the reliability of the uncertainty assessment is significantly higher than that of traditional methods. This application has broad applicability and scalability, applicable not only to carbon flux reconstruction but also extendable to the reconstruction of other ecosystem fluxes such as water and energy fluxes. In hydrothermal flux estimation, this method also achieves significant accuracy improvements, demonstrating good versatility and scalability.

[0052] At least some steps in the flowcharts involved in the embodiments described above may include multiple steps or multiple stages, which may be executed at different times, or may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0053] For those skilled in the art, various modifications and improvements can be made without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this application shall be determined by the appended claims.

Claims

1. A high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning, characterized in that, The method is used to execute steps S1-S3 to generate a four-dimensional spatiotemporal deep learning model, capture the complex spatiotemporal dynamic features of carbon flux, and achieve accurate reconstruction of land carbon flux with high spatiotemporal resolution, including: Step S1: Obtain multi-source heterogeneous datasets, including flux observation network data, satellite remote sensing data, and meteorological reanalysis data, and preprocess the multi-source heterogeneous datasets. The preprocessing includes spatiotemporal matching and feature construction of the multi-source heterogeneous datasets and updating the multi-source heterogeneous datasets. Step S2: Construct a four-dimensional spatiotemporal deep learning model architecture that connects the multi-scale spatiotemporal feature extraction module, the spatiotemporal feature fusion module, the prediction layer, and the uncertainty quantification layer sequentially from input to output. The four-dimensional spatiotemporal deep learning model architecture takes the preprocessed multi-source heterogeneous dataset as input and the carbon flux reconstruction result as output. Step S3: Construct temporal scale loss function, spatial scale loss function, and total loss function. Adopt a hierarchical learning rate strategy and introduce a cosine annealing scheduler. Train the four-dimensional spatiotemporal deep learning model architecture based on the preprocessed multi-source heterogeneous dataset to obtain the four-dimensional spatiotemporal deep learning model. Obtain the carbon flux data reconstruction set, input it into the four-dimensional spatiotemporal deep learning model to generate carbon flux reconstruction results, and use a multi-scale verification strategy to verify the carbon flux reconstruction results and perform uncertainty assessment.

2. The high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning according to claim 1, characterized in that, The spatiotemporal matching of the multi-source heterogeneous dataset in step S1 includes: The bilinear interpolation method is used to uniformly resample the spatial resolution of each data source in the multi-source heterogeneous dataset to the target grid resolution. The bilinear interpolation method is as follows: ； in, For target grid points The interpolated value of the position. and The target grid points are respectively The original data value and weight coefficient corresponding to the m-th sampling point selected as the center, where n is the total number of sampling points participating in the interpolation calculation; The cubic spline interpolation method is used to unify the temporal resolution of each data point in the multi-source heterogeneous dataset to a daily level. The cubic spline interpolation method is as follows: ； Where S(t) is the interpolation function value at time point t. Let be the starting time point of the interval, and let a, b, c, and d be the interpolation coefficients determined by the spline conditions, representing the coefficients of the constant term, linear term, quadratic term, and cubic term, respectively.

3. The high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning according to claim 2, characterized in that, The feature construction of the multi-source heterogeneous dataset in step S1 includes constructing vegetation physiological features, meteorological features, and terrain features from the spatiotemporally matched multi-source heterogeneous dataset, wherein: Vegetation physiological characteristics were constructed using satellite remote sensing data from a multi-source heterogeneous dataset. The NDVI vegetation index was calculated using the following formula: ； in, R nir The reflectance is 842 nm in the near-infrared band. R red The reflectance at 665 nm in the red band; The EVI vegetation index is calculated using the following formula: ； in, R blue The reflectance at 490 nm in the blue light band; Meteorological feature construction includes constructing daily time series features of temperature, precipitation, solar radiation, air pressure, and wind speed; Terrain feature construction includes calculating slope and aspect based on multi-source heterogeneous datasets, using the following formula: ； Where Z represents the elevation value in the multi-source heterogeneous dataset. and are the partial derivatives of elevation in the x and y directions, respectively, representing the gradient change of the terrain.

4. The high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning according to claim 1, characterized in that, In step S2, the multi-scale spatiotemporal feature extraction module of the four-dimensional spatiotemporal deep learning model architecture includes a spatial feature extraction branch and a temporal feature extraction branch. The inputs of the spatial feature extraction branch and the temporal feature extraction branch serve as the two inputs of the multi-scale spatiotemporal feature extraction module, respectively. The outputs of the spatial feature extraction branch and the temporal feature extraction branch together serve as the output of the multi-scale spatiotemporal feature extraction module and are connected to the input of the spatiotemporal feature fusion module. in: The spatial feature extraction branch uses a multi-scale 3D convolutional neural network to extract spatial features, as shown in the following equation: ； in, F s Let X be the spatial feature, W be the preprocessed multi-source heterogeneous dataset used as input features, and W be the convolution kernel weights. b Here, represents the hidden layer bias vector; BN stands for batch normalization. It is a non-linear activation function; It is a three-dimensional convolution function; The temporal feature extraction branch combines a long short-term memory network and a temporal convolutional network to simultaneously construct short-term fluctuation models and long-term trend models to extract temporal features. The long short-term memory network is used to capture long-term dependencies as follows: ； in, h t Let be the hidden state vector at time t. h t-1 Let be the hidden state vector from the previous time step. X t Let W be the input feature vector at time t. h This is the weight matrix from hidden state to hidden state. W x Let σ be the weight matrix input to the hidden state, and σ be the activation function. Using a temporal convolutional network to capture local temporal patterns is as follows: ； in, F t The time characteristic is represented by d, where d is the inflation rate. Let X be a one-dimensional convolution function, and let X be a preprocessed multi-source heterogeneous dataset used as input features.

5. A high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning according to claim 4, characterized in that, In step S2, the spatiotemporal feature fusion module of the four-dimensional spatiotemporal deep learning model architecture adopts a dual-stream spatiotemporal feature fusion network. The input of the spatiotemporal feature fusion module is connected to the output of the multi-scale spatiotemporal feature extraction module, and the output of the spatiotemporal feature fusion module is connected to the input of the prediction layer. The features extracted by the spatial feature extraction branch and the temporal feature extraction branch are adaptively integrated through a gating mechanism, as shown in the following equation: ； in, As weight; The spatiotemporal feature fusion module uses an adaptive feature fusion layer to dynamically adjust the weights of spatial and temporal features.

6. A high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning according to claim 5, characterized in that, In step S2, the prediction layer of the four-dimensional spatiotemporal deep learning model architecture adopts a multi-task learning framework, simultaneously predicting net ecosystem exchange (NEE), gross primary productivity (GPP), and ecosystem respiration (Re). A physical constraint layer is introduced to ensure the carbon balance relationship NEE = Re - GPP. The input of the prediction layer is connected to the output of the spatiotemporal feature fusion module, and the output of the prediction layer is connected to the input of the uncertainty quantification layer, as shown in the following equation: ； in, This is the physical constraint loss function.

7. A high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning according to claim 6, characterized in that, In step S2, the uncertainty quantization layer of the four-dimensional spatiotemporal deep learning model architecture uses Monte Carlo random deactivation technique to quantify the uncertainty of the model prediction. The input of the uncertainty quantization layer is connected to the output of the prediction layer, as shown in the following equation: ； in, To predict the mean, This represents the output of the four-dimensional spatiotemporal deep learning model architecture after the i-th random deactivation, where N is the total number of Monte Carlo samplings.

8. A high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning according to claim 1, characterized in that, In step S3, the time-scale loss function is constructed as follows: ； The spatial scale loss function is constructed as follows: ； in, This is the output of a four-dimensional spatiotemporal deep learning model architecture under real-world conditions. This is the result of predicting the output of a four-dimensional spatiotemporal deep learning model architecture under ideal conditions; T is the total number of steps within the cosine variation period, t is the current step; N is the total number of Monte Carlo samplings; The total loss function is constructed as follows: ； in, The weights of the time-scale loss function, The weights of the spatial scale loss function are... The weights of the physical constraint loss function, This is the physical constraint loss function.

9. A high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning according to claim 8, characterized in that, The hierarchical learning rate strategy adopted in step S3 is as follows: ； in, The initial learning rate, As the attenuation factor, For the first l The depth function of the layer, l is the index of each trainable layer in the four-dimensional spatiotemporal deep learning model architecture, with a value range from 1 to L, where L is the total number of all trainable layers in the four-dimensional spatiotemporal deep learning model architecture. The cosine annealing scheduler is introduced as follows: ； in, Let be the learning rate at the t-th iteration. To minimize the learning rate, This represents the maximum learning rate.

10. A high-precision land carbon flux reconstruction method based on four-dimensional spatiotemporal deep learning according to claim 1, characterized in that, In step S4, a multi-scale verification strategy is used to verify the carbon flux reconstruction results, including site-scale verification and regional-scale verification. The site-scale verification is as follows: ； Wherein, RMSE is the root mean square error. This is the output of a four-dimensional spatiotemporal deep learning model architecture under real-world conditions. To predict the output of a four-dimensional spatiotemporal deep learning model architecture under ideal conditions, N is the total number of Monte Carlo samplings; The regional scale verification is as follows: ； CI stands for Consistency Index. and This is the output of a four-dimensional spatiotemporal deep learning model architecture under real-world conditions. It is the average of N sample results.

Images