Large-area forest vegetation carbon reserve inversion method based on GEE platform multi-source data fusion

Abstract

The invention discloses a large-area forest vegetation carbon reserve inversion method based on GEE platform multi-source data fusion. The method comprises the following steps: data acquisition and analysis: obtaining forest carbon storage literature analysis data and an index factor data set; performing correlation analysis, constructing a forest carbon reserve index system, and establishing a corresponding characteristic variable relationship between the independent variable index factor and the forest carbon reserve sample; a seq2seq deep learning model based on LSTM (Long Short Term Memory) is imported; precision analysis is carried out; and inverting a large-area forest vegetation carbon reserve data set by using the deep learning model, and analyzing main determinants of forest carbon reserves. According to the method, a Google Earth engine is introduced, factors such as waveband information, vegetation indexes and long-term change trend characteristics are extracted from the Google Earth engine through a huge remote sensing data file of the Google Earth engine and a strong parallel computing capability of the Google Earth engine, and the forest carbon reserve in the loess plateau region is inverted and main determinants of the forest carbon reserve are analyzed based on a seq2seq deep learning model of LSTM.

Description

Technical Field

[0001] The present invention relates to the field of terrestrial ecosystem carbon storage monitoring and inversion technology, and more specifically, to an inversion method for large-area forest vegetation carbon storage based on multi-source data fusion on a GEE platform. Background Art

[0002] Global climate change, characterized by warming and frequent extreme weather, is a common challenge facing humanity today. Compared with 1850-1900, the global average temperature rose by 1.09°C from 2011 to 2020. Among them, anthropogenic carbon emissions are the main cause of the imbalance of atmospheric carbon dioxide.

[0003] Forests act as major carbon sinks through photosynthesis and become sources of carbon emissions due to deforestation or land cover change. Carbon storage is determined by biomass and its corresponding carbon content, so forest biomass is a key factor in studying forest carbon storage. The carbon pool of forest ecosystems is mainly composed of soil carbon, dead organic carbon, and aboveground and underground vegetation carbon. Commonly used carbon storage and biomass estimation parameters include biomass expansion factor (BEF), wood density (WD), root-to-shoot ratio (RSR), carbon content coefficient, etc. The biomass expansion factor can first convert the accumulation into biomass, and then expand it into aboveground biomass and total biomass. The underground biomass value, which is difficult to obtain, is obtained by calculating the root-to-shoot ratio and aboveground biomass, and the forest carbon stock is finally obtained by multiplying the obtained biomass sum with the carbon content coefficient. These methods for estimating forest ecosystem carbon storage based on forest resource inventory and biomass model have laid the foundation for accurately grasping the carbon storage of forest ecosystems. So far, the main methods for estimating forest biomass are: sample plot inventory method, micrometeorological method, model simulation method and remote sensing estimation method. The sample plot inventory method has high accuracy, but it is costly, time-consuming and easy to damage the sample trees, so it is more suitable for small-scale research. Accurate and rapid estimation of forest carbon storage in large areas and large scales has always been a hot topic and difficulty in forestry research.

[0004] The observation capability of remote sensing technology in continuous time and space enables researchers to estimate forest biomass over large areas at a lower cost and with higher accuracy. Vegetation accumulates biomass through photosynthesis. By analyzing the spectral characteristics of different vegetation in remote sensing images and the correlation between vegetation spectral information and measured biomass, a mathematical model for biomass estimation can be established. Remote sensing inversion methods can provide multi-scale and large-scale remote sensing data, realize real-time monitoring of surface data, and quickly and quantitatively obtain forest biomass at a regional scale, as well as analyze and evaluate the spatiotemporal pattern and dynamic changes of forest biomass. Due to the complexity, species diversity and regional heterogeneity of forest ecosystems, it is impossible to accurately estimate and monitor forest carbon storage using a unified model or calculation method. At present, remote sensing data with multi-scale and multi-spectral characteristics are widely used to extract forest vegetation information. Various machine learning algorithms can be used to distinguish forest types and estimate biomass. With the exponential growth of the number of remote sensing satellites and the amount of data, cloud storage and cloud computing technologies have risen rapidly. Summary of the invention

[0005] To this end, the technical problem to be solved by the present invention is to provide an inversion method for large-area forest vegetation carbon storage based on multi-source data fusion of the GEE platform, so as to solve the difficulty that traditional plot inventory methods, micrometeorological methods, model simulation methods and other models are difficult to accurately and quickly estimate forest carbon storage in large areas and large scales.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] The inversion method of large-area forest vegetation carbon storage based on multi-source data fusion on the GEE platform includes the following steps:

[0008] (1) Data acquisition and analysis: Literature review, management and analysis based on the ArcGIS platform, and obtaining forest carbon storage literature analysis data and indicator factor data sets;

[0009] (2) Conduct correlation analysis on forest carbon storage literature analysis data and index factor data, construct a forest carbon storage index system, and establish the corresponding characteristic variable relationship between the independent variable index factor and the forest carbon storage sample;

[0010] (3) Import the corresponding feature variable relationship into the LSTM-based seq2seq deep learning model; and analyze the accuracy of the forest carbon storage inversion results of the deep learning model;

[0011] (4) Use the deep learning model to invert a large-area forest vegetation carbon storage dataset and analyze the main determinants of forest carbon storage.

[0012] In the above-mentioned inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform, in step (1), the index factors that can characterize the forest carbon content are determined as follows:

[0013] There are 12 index factors preliminarily determined, including: vegetation cover FVC, leaf area index LAI, photosynthetically active radiation absorption fraction FPAR, gross primary productivity GPP, net primary productivity NPP, normalized difference vegetation index NDVI, enhanced vegetation index EVI, land use index LCI, radar tree height index RHI, elevation data DEM, precipitation data PRE, and temperature data TEM.

[0014] The above-mentioned inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform is

[0015] By setting up the cloud removal function, the accumulated FVC, LAI, FPAR, GPP, NPP, NDVI, EVI, LCI, RHI, DEM, PRE, and TEM data can be calculated and downloaded; the cloud function is a data screening step.

[0016] The data of the 12 preliminarily determined index factors were subjected to VIF multicollinearity analysis (VIF test, i.e., variance inflation factor (VIF) test, is a multicollinearity diagnostic method commonly used in regression analysis), and the NDVI, LCI, and RHI indices with strong multicollinearity were eliminated. Finally, 9 index factors representing forest carbon content were selected for calculation. The 9 final index factors included vegetation cover FVC, leaf area index LAI, photosynthetically active radiation absorption fraction FPAR, gross primary productivity GPP, net primary productivity NPP, enhanced vegetation index EVI, elevation data DEM, precipitation data PRE, and temperature data TEM.

[0017] In the above-mentioned inversion method of forest vegetation carbon storage in a large area based on multi-source data fusion of GEE platform, in step (2), forest carbon storage samples are obtained after outlier screening based on forest carbon storage survey data extracted from relevant literature; according to the correlation analysis results between independent variable index factors and sample plot biomass, the corresponding relationship between independent variable index factors and forest carbon storage samples is established (used to reflect the physical parameters of vegetation): vegetation coverage FVC and enhanced vegetation index EVI are used to characterize vegetation growth conditions, leaf area index LAI is used to distinguish the carbon fixation capacity of different vegetation types, photosynthetically active radiation absorption fraction FPAR, gross primary productivity GPP, net primary productivity NPP are used to characterize the ability of vegetation ecosystems in the study area to receive solar energy and convert growth, elevation data DEM is used to reflect vegetation types at different altitudes, and precipitation data PRE and temperature data TEM mainly affect the growth status of vegetation ecosystems within a certain period of time. Through many indicator factors that can characterize forest carbon content, forest carbon content is directly inverted. The above parameters are usually used for linear or nonlinear carbon storage inversion. This paper uses a deep learning inversion model.

[0018] In the above-mentioned inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform, in step (3), according to the correlation analysis results of the independent variable index factor and the biomass of the forest carbon storage sample site, the corresponding feature variables are imported into the LSTM-based seq2seq deep learning model, and the determination coefficient R is used to calculate the carbon storage of forest vegetation in the large-area forest. 2 And the prediction accuracy was evaluated.

[0019] The above-mentioned inversion method for large-area forest vegetation carbon storage based on multi-source data fusion of the GEE platform has the following corresponding characteristic variables: the independent variable is the 9 index factors finally selected to characterize the forest carbon content, and the target data is the forest carbon storage point data, which is the dependent variable. The independent variables and dependent variables form a data set and are imported into the seq2seq deep learning model for learning and prediction.

[0020] In the above-mentioned inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of the GEE platform, in step (4), the GEE platform is called to obtain remote sensing data of the location to be inverted, and the data spatial resolution is set to 500 meters:

[0021] Using MODIS image data, the following data were obtained: vegetation cover FVC, leaf area index LAI, photosynthetically active radiation absorption fraction FPAR, gross primary productivity GPP, net primary productivity NPP, enhanced vegetation index EVI; elevation data DEM was obtained using NASA; precipitation data PRE was obtained using UCSB-CHG; and temperature data TEM was obtained using ECMWF.

[0022] The above-mentioned inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of the GEE platform will import the data obtained from the GEE platform into the LSTM seq2seq deep learning model to perform large-area forest carbon storage inversion, obtain a forest carbon storage data set in a certain area over a certain period of time, and analyze the main determinants of forest carbon storage.

[0023] The technical solution of the present invention achieves the following beneficial technical effects:

[0024] 1. Based on literature analysis, a forest carbon storage sample data set was constructed; in the large-area inversion test phase, the forest carbon storage was first inverted using GEE embedded machine learning models such as support vector machine, random forest, and random gradient boosting, and the inversion accuracy of forest vegetation carbon storage was compared with the LSTM-based seq2seq deep learning model proposed in this study. Finally, the deep learning model with the highest accuracy was applied to the inversion of forest vegetation carbon storage in the Loess Plateau.

[0025] Google Earth Engine (GEE) provided by Google Cloud Infrastructure provides great technical convenience for vegetation remote sensing analysis. The huge remote sensing data archive in GEE, coupled with its powerful parallel computing capabilities, overcomes the limitations of research scope and data volume, and effectively improves recognition efficiency by extracting various vegetation information from long time series data sets.

[0026] 2. The present invention aims to introduce Google Earth Engine, through its huge remote sensing data archive and its powerful parallel computing capabilities, to construct multi-index cumulative annual mean synthetic data based on MODIS images of GEE platform, and extract band information, vegetation index, long-term trend characteristics and other factors. On this basis. Add NASA digital elevation data, UCSB-CHG precipitation data and ECMWF temperature data to form a characteristic factor set to comprehensively characterize the forest carbon storage capacity of the study area. The LSTM-based seq2seq deep learning model is used to invert the forest vegetation carbon storage to improve the inversion accuracy.

[0027] 3. This application is based on a seq2seq deep learning model based on LSTM, which is a deep learning architecture designed for tasks involving sequences as input and output. The key idea behind the seq2seq model is to process an input sequence and generate an output sequence of possibly different lengths, which is achieved through the main components of the encoder and decoder. The encoder takes the input sequence and processes it into a fixed-length representation. It captures basic information from the input sequence in the context vector; the decoder takes the context vector from the encoder and generates the output sequence step by step. At each step, the internal state of the decoder and the previously generated output are used to predict the next element of the output sequence. Attention mechanisms are often used in seq2seq models to help the decoder focus on different parts of the input sequence when generating each element of the output sequence. This helps the model process long sequences and improves the performance of various tasks.

[0028] 4. This study is based on forest vegetation carbon storage inventory data, MODIS remote sensing images, NASA surface elevation data, UCSB-CHG precipitation data, and ECMWF temperature data to conduct relevant research on the forest ecosystem in the Loess Plateau. The LSTM-based seq2seq deep learning model is used to invert the forest carbon storage in the Loess Plateau from 2019 to 2021 and analyze the main determinants of forest carbon storage. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 , Flowchart of the inversion method of forest vegetation carbon storage in a large area based on multi-source data fusion on the GEE platform;

[0030] Figure 2 , forest carbon stock inversion index system: (a) FVC, (b) LAI, (c) FPAR, (d) GPP, (e) NPP, (f) EVI, (g) DEM, (h) PRE, (i) TEM;

[0031] Figure 3 , Forest carbon storage inversion results in the Loess Plateau region based on LSTM-based seq2seq deep learning model: (a) 2019, (b) 2020, (c) 2021;

[0032] Figure 4 , importance analysis of indicator factors. DETAILED DESCRIPTION

[0033] GEE cloud platform is a free spatial analysis platform that provides global-scale geospatial information data supported by Google cloud infrastructure. In the era of big data, GEE takes advantage of the cloud for data query and acquisition, and has stable performance. The seq2seq deep learning model based on LSTM (Long Short-Term Memory Network) has been proven to be effective for various sequential data tasks, especially when the input and output have complex relationships and variable lengths. This model is a popular and powerful architecture for sequence-related tasks because they are able to capture complex patterns in sequence data and use attention mechanisms to improve performance.

[0034] The optical image data, land use data, elevation data, and meteorological data used in this study are all from the GEE cloud platform. By integrating high-resolution active microwave remote sensing and optical remote sensing information and obtaining a large amount of measured point data from literature, a vegetation carbon storage dataset for the Loess Plateau was developed with a spatial resolution of 500 meters.

[0035] Table 1 Data types and sources

[0036]

[0037]

[0038] The specific steps include the following: Figure 1 Flowchart of the inversion method for large-area forest vegetation carbon storage based on multi-source data fusion on the GEE platform.

[0039] 1. Data acquisition and analysis: Literature review, management and analysis based on ArcGIS platform were conducted to obtain forest carbon storage literature analysis data and its corresponding indicator factor data set.

[0040] Through ArcGIS platform management, the analytical data and related index factors of forest carbon storage literature were analyzed, and finally the index factor that can reflect the forest carbon storage was determined.

[0041] (1-1) A total of 12 index factors that can characterize forest carbon storage were preliminarily determined, including vegetation cover (FVC), leaf area index (LAI), photosynthetically active radiation absorption fraction (FPAR), gross primary productivity (GPP), net primary productivity (NPP), normalized differential vegetation index (NDVI), enhanced vegetation index (EVI), land use index (LCI), relative height index (RHI), digital elevation model (DEM), precipitation data (PRE), and temperature data (TEM).

[0042] (1-2) By setting functions such as cloud removal, calculate and download the accumulated FVC, LAI, FPAR, GPP, NPP, NDVI, EVI, LCI, RHI, DEM, PRE, and TEM data.

[0043] (1-3) The 12 index factors initially determined were subjected to VIF multicollinearity analysis, and the NDVI, LCI, and RHI indices with strong multicollinearity were eliminated. Finally, 9 index factors representing forest carbon storage were selected for calculation. The 9 final index factors included vegetation coverage FVC, leaf area index LAI, photosynthetically active radiation absorption fraction FPAR, gross primary productivity GPP, net primary productivity NPP, enhanced vegetation index EVI, elevation data DEM, precipitation data PRE, and temperature data TEM, such as Figure 1 shown.

[0044] 2. Conduct correlation analysis on forest carbon storage literature analysis data and indicator factor data, construct a forest carbon storage indicator system, and establish the corresponding characteristic variable relationship between the independent variable index factor and the forest carbon storage sample.

[0045] This application utilizes the GEE cloud platform to obtain global-scale geospatial information data. By integrating high-resolution active microwave remote sensing and optical remote sensing information, a large amount of measured point data from literature is obtained, thereby constructing a corresponding relationship between the index factor and the forest carbon stock sample, laying the foundation for inversion of the seq2seq deep learning model input into the LSTM (long short-term memory network).

[0046] All data were downloaded from the GEE platform, and the spatial resolution of all data was set to 500 meters.

[0047] Among the above 9 indicator factors, the corresponding relationship between the independent variable index factor and the forest carbon storage sample is as follows:

[0048] Vegetation coverage fraction (FVC) and enhanced vegetation index (EVI) are used to characterize vegetation growth conditions;

[0049] The leaf area index LAI is used to distinguish the carbon sequestration capacity of different vegetation types;

[0050] The fraction of photosynthetically active radiation absorbed (FPAR), gross primary productivity (GPP), and net primary productivity (NPP) are used to characterize the ability of vegetation ecosystems in the study area to receive solar energy and transform it into growth.

[0051] Elevation data DEM is used to reflect vegetation types at different altitudes;

[0052] Precipitation data PRE and temperature data TEM mainly affect the growth of vegetation ecosystems within a certain period of time. Figure 2 shown.

[0053] 3. Import the corresponding feature variable relationship into the LSTM-based seq2seq deep learning model; and analyze the accuracy of the forest carbon storage inversion results of the deep learning model.

[0054] LSTM is a special type of recurrent neural network with a memory structure for learning long-term information. Sequence-to-sequence (seq2seq) model is a deep learning architecture designed for tasks involving sequences as input and output. The key idea behind seq2seq model is to process input sequence and generate output sequence of possibly different lengths, which is achieved through encoder and decoder main components. Encoder takes input sequence and processes it into fixed-length representation. It captures essential information from input sequence in context vector; decoder takes context vector from encoder and generates output sequence step by step. At each step, internal state of decoder and previously generated output are used to predict next element of output sequence. Attention mechanism is often used in seq2seq model to help decoder focus on different parts of input sequence while generating each element of output sequence. This helps the model to process long sequences and improves performance on various tasks.

[0055] Based on the forest carbon storage survey (inversion) data extracted from relevant literature, the forest carbon storage samples obtained after outlier screening, cleaning and statistics were used as dependent variables. 219 sample points were selected for coniferous forests, broad-leaved forests and mixed coniferous and broad-leaved forests, of which 150 were used for modeling and 69 for verification; 372 sample points were selected for forests, of which 215 were used for modeling and 157 for verification. According to the correlation analysis results between the exponential factor (independent variable) and the biomass of the sample plot (dependent variable), the independent variable and the dependent variable constitute a data set, and the corresponding characteristic variables (the corresponding relationship between the independent variable and the dependent variable) are respectively imported into GEE embedded machine learning models such as support vector machine, random forest, and stochastic gradient boosting, as well as LSTM-based seq2seq deep learning model, and the accuracy of the experimental results of the above models is evaluated.

[0056] Statistical analysis. According to the verification results of the biomass samples in the study area (Table 2), from the perspective of the four regression models, in the inversion of multi-feature forest biomass, the deep learning regression model in this paper has the best estimation effect, and the forest carbon storage prediction model R 2 All of them reached above 0.8921, which was significantly higher than the multi-SVM model and RF model. 2 The minimum is 0.7861. The Seq2seq model regression can make predictions based on the existing data model, has good nonlinear fitting ability, and is strong in processing high-dimensional data. The accuracy of the multivariate regression models established for other forest types is lower than that of the Seq2seq model regression model (87.02%), and the RF model has the lowest accuracy, which is 78.58%. 2 As well as the prediction accuracy, the Seq2seq model has the best regression fit and is relatively robust, which can be applied to the inversion of vegetation carbon storage in large areas.

[0057] Table 2 Accuracy of four models in forest carbon stock inversion

[0058] SVM RF GBDT This article's deep learning model <![CDATA[R 2 ]]> 0.8135 0.8294 0.7861 0.8921 Prediction accuracy 81.28％ 78.58％ 80.21％ 87.02％

[0059] 4. Use the deep learning model to invert a large-area forest vegetation carbon storage dataset and analyze the main determinants of forest carbon storage.

[0060] The LSTM-based seq2seq deep learning model has been widely used in forest biomass estimation in recent years. It can quickly construct a large number of regression trees for prediction. It is robust, easy to understand, and insensitive to noisy data.

[0061] The model has good adaptability to complex forest parameters and does not require a priori models. It can also effectively analyze the nonlinear relationship between forest parameters and overcome the collinearity problem between data.

[0062] The inversion results of forest vegetation carbon storage in the Loess Plateau from 2019 to 2021 based on the LSTM-based seq2seq deep learning model are as follows: Figure 3 As shown, the gray part is grassland, farmland, desert and other land cover types. Analysis shows that the vegetation in the study area is mainly distributed in the southeastern river valleys, plains and other areas. Over the past three years, the spatial distribution pattern of forest carbon storage has changed little, with high carbon storage in the middle of the southern forest area of ​​the Loess Plateau, and the areas on both sides gradually transitioning from high to low. In order to explore the causes of the spatial distribution pattern of forest carbon storage in the Loess Plateau, we conducted an importance analysis of the above-mentioned index factors and forest carbon storage ( Figure 4 ). It can be obtained that the net primary productivity NPP in the study area has the greatest correlation with the gross primary productivity GPP, followed by the enhanced vegetation index EVI and vegetation coverage FVC, while the correlation with meteorological factors is the smallest. This shows that the forest carbon storage capacity in the study area is closely related to the strength of vegetation photosynthesis and vegetation type. At the same time, meteorological factors are also an important condition for determining carbon storage. Under the conditions of global climate change, different forest vegetation systems are threatened by extreme weather, resulting in a decrease in carbon storage capacity.

[0063] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the claims of this patent application.

Claims

1. The inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform is characterized by: The steps include: (1) Data acquisition and analysis: Literature review, management and analysis based on the ArcGIS platform, and obtaining forest carbon storage literature analysis data and indicator factor data sets; (2) Conduct correlation analysis on forest carbon storage literature analysis data and index factor data, construct a forest carbon storage index system, and establish the corresponding characteristic variable relationship between the independent variable index factor and the forest carbon storage sample; (3) Import the corresponding feature variable relationship into the LSTM-based seq2seq deep learning model; and analyze the accuracy of the forest carbon storage inversion results of the deep learning model; (4) Use the deep learning model to invert a large-area forest vegetation carbon storage dataset and analyze the main determinants of forest carbon storage.

2. The inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform according to claim 2 is characterized in that: In step (1), the index factors that can characterize the carbon storage content of the forest are determined as follows: There are 12 index factors preliminarily determined, including: vegetation cover FVC, leaf area index LAI, photosynthetically active radiation absorption fraction FPAR, gross primary productivity GPP, net primary productivity NPP, normalized difference vegetation index NDVI, enhanced vegetation index EVI, land use index LCI, radar tree height index RHI, elevation data DEM, precipitation data PRE, and temperature data TEM.

3. The inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform according to claim 2 is characterized in that: By setting up the cloud removal function, calculate and download the accumulated FVC, LAI, FPAR, GPP, NPP, NDVI, EVI, LCI, RHI, DEM, PRE, and TEM data; The data of the 12 preliminarily determined index factors were subjected to VIF multicollinearity analysis, and the NDVI, LCI, and RHI indices with strong multicollinearity were eliminated. Finally, 9 index factors representing forest carbon storage were selected for calculation. The 9 final index factors included vegetation cover FVC, leaf area index LAI, photosynthetically active radiation absorption fraction FPAR, gross primary productivity GPP, net primary productivity NPP, enhanced vegetation index EVI, elevation data DEM, precipitation data PRE, and temperature data TEM.

4. The inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform according to claim 3 is characterized in that: In step (2), forest carbon stock samples are obtained after outlier screening based on forest carbon stock survey data extracted from relevant literature; the corresponding relationship between independent variable index factors and forest carbon stock samples is established based on the correlation analysis results between independent variable index factors and sample plot biomass: vegetation coverage FVC and enhanced vegetation index EVI are used to characterize vegetation growth conditions, leaf area index LAI is used to distinguish the carbon fixation capacity of different vegetation types, photosynthetically active radiation absorption fraction FPAR, gross primary productivity GPP, and net primary productivity NPP are used to characterize the ability of vegetation ecosystems in the study area to receive solar energy and convert it into growth, elevation data DEM is used to reflect vegetation types at different altitudes, and precipitation data PRE and temperature data TEM mainly affect the growth status of vegetation ecosystems within a certain period of time.

5. The inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform according to claim 4 is characterized in that: In step (3), according to the correlation analysis results between the independent variable index factor and the biomass of the forest carbon storage sample site, the corresponding feature variables are imported into the LSTM-based seq2seq deep learning model, and the determination coefficient R 2 And the prediction accuracy was evaluated.

6. The inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform according to claim 4 is characterized in that: The corresponding characteristic variables are: the independent variable is the 9 index factors finally selected to characterize the forest carbon content, and the target data is the forest carbon storage point data, which is the dependent variable. The independent variable and the dependent variable constitute a data set and are imported into the seq2seq deep learning model for learning and prediction.

7. The inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform according to claim 1 is characterized in that: In step (4), the GEE platform is called to obtain the remote sensing data of the location to be inverted, and the data spatial resolution is set to 500 meters: Using MODIS image data, the following data were obtained: vegetation cover FVC, leaf area index LAI, photosynthetically active radiation absorption fraction FPAR, gross primary productivity GPP, net primary productivity NPP, enhanced vegetation index EVI; elevation data DEM was obtained using NASA; precipitation data PRE was obtained using UCSB-CHG; and temperature data TEM was obtained using ECMWF.

8. The inversion method of large-area forest vegetation carbon storage based on multi-source data fusion of GEE platform according to claim 5 is characterized in that: The data obtained from the GEE platform is imported into the LSTM seq2seq deep learning model to invert the forest carbon storage in a large area, obtain the forest carbon storage data set in a certain area over a certain period of time, and analyze the main determinants of forest carbon storage.

Images