Training method and reconstruction method of normalized difference vegetation index reconstruction model

Abstract

The present disclosure provides a training method for a normalized vegetation index reconstruction model and a normalized vegetation index reconstruction method, the training method comprising: obtaining normalized vegetation index data and corresponding surface parameter reanalysis data, the surface parameter reanalysis data comprising a plurality of time series parameters, wherein the plurality of time series parameters comprise a steady-state factor and a dynamic variable, the spatial distribution of the steady-state factor varying with time less than the spatial distribution of the dynamic variable varying with time; generating a training sample data set according to the surface parameter reanalysis data and the normalized vegetation index; and training a normalized vegetation index reconstruction model to be trained using the training sample data set to obtain the normalized vegetation index reconstruction model.

Description

Technical Field

[0001] The present disclosure relates to the technical field of surface vegetation growth, and in particular to a training method for a historical normalized vegetation index reconstruction model based on machine learning and a historical normalized vegetation index reconstruction method. Background Art

[0002] Vegetation index is a quantitative expression of the growth status of surface vegetation, mainly reflecting the difference in reflectivity between vegetation in the visible light, near-infrared bands and soil background. Vegetation index sequences derived from satellite remote sensing data can reflect information such as vegetation growth and surface cover changes at multiple scales, and have become the data basis for global change research. The reconstruction of historical vegetation indices is a scientific practice to explore the impact of global change and connect with the United Nations Sustainable Development Goals. It has important data support significance for the analysis of historical climate change and vegetation evolution mechanisms. However, there are relatively few studies on the mining and reconstruction of vegetation index data before the maturity of remote sensing technology. The time interval span of historical reconstruction data supported by proxy data is large, and it is impossible to form a complete and continuous high-precision vegetation index product. In order to explore and understand the laws and characteristics of vegetation growth changes over a long period of time under the background of global change, support global climate change research and regional sustainable development, it is urgent to use multi-source parameter data and new model technology to construct long-term series vegetation index data products.

[0003] Time series data reconstruction aims to use a variety of numerical analysis methods to simulate the law of data change over time, so as to optimize it and provide a more complete data basis for related research. Vegetation index reconstruction mainly includes two aspects. One is to reconstruct the data outside the time range of the current vegetation index data product by combining relevant factor data, and to expand and extend the data time series range; the other is to improve the spatiotemporal resolution of the existing vegetation index data products through algorithms, reduce atmospheric interference and thus improve the quality of vegetation index data.

[0004] In terms of filling and reconstructing vegetation indices in historical periods, existing scientific research has mainly reconstructed them based on a large number of historical records, proxy data and existing data products, thereby exploring the characteristics and changing laws of vegetation and climate in historical periods. For example, Pound et al. (2011) reconstructed the global vegetation and animal distribution during the Tortonian period based on 223 vegetation sample points provided by the Tertiary Environment Vegetation Information System, combined with historical documents and the climate and vegetation model of the late Miocene. Gao Weidong et al. (2012) reconstructed the maximum normalized difference vegetation index (NDVI) of grassland in the Hutubi River Basin in Xinjiang from 1662 to September 2001 based on tree ring indices. Liu et al. (2013) constructed a quantitative reconstruction model of vegetation cover in North China based on sediment profiles of 15 lakes in northern China and their modern remote sensing vegetation index. Bunting et al. (2016) constructed a historical relative pollen productivity model through linear regression based on the ground pollen deposition data of broad-leaved forests and shrub forests in three areas of East Anglia, UK, and reconstructed the historical vegetation index of the three study areas. Shang Huaming et al. (2016) reconstructed the NDVI of central Tibet, China from July to August 1565 to 2006 using tree ring width. Shao et al. (2018) reconstructed the global vegetation density during the Last Glacial Maximum using statistical methods based on the relationship between climate, vegetation and CO2 and historical pollen record data. Zhang Xu et al. (2018) reconstructed the NDVI of the southern forest area of ​​Xiaoxing'anling from July 1900 to 2015 based on the tree ring index of four tree species. Li et al. (2019) quantitatively reconstructed the vegetation information of China since the last glacial period based on 286 fossil pollen records analyzed by the biome regionalization method, revealing the spatial distribution changes of different species such as grasslands, deserts, and forests during the historical period. In terms of improving and reconstructing existing vegetation index products, related research mainly uses threshold method, filtering method, function fitting method and other methods to improve the spatiotemporal resolution of existing vegetation index data and reduce noise interference factors. However, due to different definitions, design ideas and calculation methods, various reconstruction methods have great differences in noise removal, smoothing effect and fidelity of time series data.

[0005] At present, the reconstruction of vegetation index is mainly to improve the spatiotemporal resolution and remove noise of existing data, and the results are limited to the time range covered by vegetation index data. The research on data reconstruction methods and models outside the time range of existing vegetation index products is relatively simple, mostly based on proxy data to study paleoclimate and paleovegetation characteristics, lacking the output of modern and contemporary continuous data products, resulting in insufficient understanding of the continuous changes and succession patterns of vegetation in historical periods.

[0006] Existing vegetation data reconstruction studies are mostly based on simple regression simulations due to the small number of input variables, which makes it difficult to characterize the subtle differences in time and space of long-sequence vegetation indices under complex climatic conditions. Machine learning has excellent simulation capabilities for multivariable, nonlinear complex systems. Since its inception, it has been widely used in research such as earth system change simulation, remote sensing image information extraction, image classification and target recognition, and climate and environmental element inversion. Currently, there is a lack of research on the reconstruction of historical vegetation index spatiotemporal series using machine learning algorithms. By leveraging the advantages and characteristics of machine learning algorithms, the development of a spatiotemporal series reconstruction model for vegetation index data based on machine learning algorithms will help promote the development of spatiotemporal series reconstruction algorithms for vegetation index data and expand the time range and wide application of vegetation index data. Summary of the invention

[0007] In order to reconstruct historical NDVI data more accurately, the present disclosure provides a training method for a normalized vegetation index reconstruction model, including:

[0008] Acquire normalized vegetation index data and corresponding surface parameter reanalysis data, wherein the surface parameter reanalysis data includes a plurality of time series parameters, wherein the plurality of time series parameters include a steady-state factor and a dynamic variable, and the spatial distribution of the steady-state factor varies less than the spatial distribution of the dynamic variable over time;

[0009] Generating a training sample data set according to the surface parameter reanalysis data and the normalized difference vegetation index; and

[0010] The training sample data set is used to train the normalized vegetation index reconstruction model to be trained, so as to obtain the normalized vegetation index reconstruction model.

[0011] According to some embodiments of the present disclosure, the steady-state factor includes any one or a combination of the following: digital elevation model, dominant soil component, eco-geographical zoning, topsoil clay content, topsoil gravel content, and topsoil USDA texture classification; and

[0012] The dynamic variables include any one or a combination of the following: aerodynamic conductivity, air temperature, surface air pressure, snowfall flux, rainfall flux, underground runoff, specific humidity, wind speed, soil profile moisture, canopy water content, vegetation transpiration flux, surface downward shortwave radiation net flux, ground heat flow, surface upward latent heat flux and surface upward latent heat flux.

[0013] According to some embodiments of the present disclosure, using the training sample data set to train the normalized vegetation index reconstruction model to be trained includes:

[0014] The iteration does the following:

[0015] Randomly selecting samples from the training sample data set to establish multiple independent subsets, establishing multiple regression trees based on each subset, importing random forest regression functions as leaf nodes of the regression trees, wherein a sub-model is constructed in each node of each regression tree to simulate the relationship between the time series parameter and the normalized vegetation index;

[0016] Using the multiple regression trees to test samples of the training sample data set, and outputting contribution rates of the multiple time series parameters to the normalized difference vegetation index respectively; and

[0017] Determine whether the contribution rates of the multiple time series parameters are all greater than or equal to a preset threshold; if not, remove the time series parameters whose contribution rates are less than the preset threshold from the training sample data set, and re-establish multiple regression trees for training; if so, stop the iteration and use the model when the iteration stops as the normalized vegetation index reconstruction model.

[0018] According to some embodiments of the present disclosure, obtaining surface parameter reanalysis data corresponding to normalized difference vegetation index data includes:

[0019] The data of multiple time series parameters are synthesized at a time window scale respectively to obtain the surface parameter reanalysis data at a window scale, wherein the time window scale synthesis includes synthesis by obtaining the cumulative value within the time window, or synthesis by obtaining the average value within the time window.

[0020] According to some embodiments of the present disclosure, the time window scale includes a monthly scale, and the normalized vegetation index reconstruction model to be obtained by training with the training sample data set is a monthly scale reconstruction model.

[0021] According to some embodiments of the present disclosure, the training sample data set includes training sets and corresponding validation sets of multiple monthly scales, and the ratio of the number of samples in the training set to that in the validation set is 7:3.

[0022] According to some embodiments of the present disclosure, obtaining normalized vegetation index data includes:

[0023] Get the original normalized difference vegetation index data;

[0024] Processing the original NDVI data by the maximum value synthesis method and generating monthly NDVI data; and

[0025] The bilinear interpolation method is used to resample the monthly NDVI data to obtain the NDVI data to be used.

[0026] According to some embodiments of the present disclosure, testing the samples of the training sample data set using the multiple regression trees includes:

[0027] The votes of the multiple regression trees are counted, and the leaf node with the most votes is output as the normalized vegetation index.

[0028] According to some embodiments of the present disclosure, the training sample data set also includes a test set;

[0029] After outputting the normalized difference vegetation index, the method further includes: using the test set and statistical indicators to evaluate the accuracy of the output normalized difference vegetation index, and outputting the evaluation result;

[0030] The statistical indicators include any one or a combination of the following: Pearson correlation coefficient, error sum of squares, determination coefficient, root mean square error, mean deviation error, standard deviation and mean value.

[0031] The present disclosure also provides a normalized vegetation index reconstruction method, comprising:

[0032] Obtaining reanalysis data of surface parameters corresponding to the normalized difference vegetation index to be reconstructed;

[0033] The surface parameter reanalysis data is input into a normalized vegetation index reconstruction model, and a reconstruction result is output, wherein the normalized vegetation index reconstruction model is trained by the above-mentioned normalized vegetation index reconstruction model training method.

[0034] Through the above-mentioned embodiments of the present disclosure, the factors affecting NDVI changes are determined by studying the relationship between hydrothermal environmental conditions and spatial distribution characteristics of vegetation, and then a random forest-based NDVI reconstruction model is designed and implemented based on reanalysis data and multi-source remote sensing data. The optimal reconstruction parameters are determined by test set accuracy evaluation and control of input parameters, and a training method for a normalized vegetation index reconstruction model based on machine learning is obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flowchart schematically shows a training method for a normalized vegetation index reconstruction model according to an embodiment of the present disclosure;

[0036] Figure 2 A block diagram schematically illustrates a training method for a normalized vegetation index reconstruction model according to an embodiment of the present disclosure; and

[0037] Figure 3 The schematic diagram schematically shows the contribution rate of time series parameters to the normalized vegetation index in the training method of the normalized vegetation index reconstruction model in the embodiment of the present disclosure. DETAILED DESCRIPTION

[0038] In order to make the objectives, technical solutions and advantages of the present disclosure more clearly understood, the present disclosure is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0039] However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, the description of the known technology is omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0040] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The term "comprising" used herein indicates the existence of features, steps, operations, but does not exclude the existence or addition of one or more other features.

[0041] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification, and should not be interpreted in an idealized or overly rigid manner.

[0042] Figure 1 The flowchart of the training method of the normalized vegetation index reconstruction model according to the embodiment of the present disclosure is schematically shown.

[0043] In order to solve the defects of the solutions in the related art and to reconstruct the historical NDVI data more accurately, the present disclosure provides a training method for the normalized vegetation index reconstruction model, such as Figure 1 As shown, it includes operations S1 to S3.

[0044] According to some embodiments of the present disclosure, operation S101 includes: acquiring normalized vegetation index data and corresponding surface parameter reanalysis data, where the surface parameter reanalysis data includes multiple time series parameters.

[0045] In this embodiment, the multiple time series parameters include steady-state factors and dynamic variables, and the spatial distribution of the steady-state factors varies less than the spatial distribution of the dynamic variables over time.

[0046] According to some embodiments of the present disclosure, operation S102 includes: generating a training sample data set according to surface parameter reanalysis data and a normalized difference vegetation index.

[0047] According to some embodiments of the present disclosure, operation S103 includes: training a normalized vegetation index reconstruction model to be trained using a training sample data set to obtain a normalized vegetation index reconstruction model.

[0048] According to some embodiments of the present disclosure, the steady-state factor includes any one or a combination of the following: digital elevation model, dominant soil component, eco-geographical zoning, topsoil clay content, topsoil gravel content, and topsoil USDA texture classification.

[0049] According to some embodiments of the present disclosure, the dynamic variables include any one or a combination of the following: aerodynamic conductivity, air temperature, surface air pressure, snowfall flux, rainfall flux, underground runoff, specific humidity, wind speed, soil profile moisture, canopy water content, vegetation transpiration flux, surface downward shortwave radiation net flux, ground heat flow, surface upward latent heat flux, and surface upward latent heat flux.

[0050] According to some embodiments of the present disclosure, training a normalized vegetation index reconstruction model to be trained using a training sample data set includes iteratively performing operations S201 to S203.

[0051] According to some embodiments of the present disclosure, operation S201 includes: randomly selecting samples from a training sample data set to establish multiple independent subsets, establishing multiple regression trees based on each subset, importing a random forest regression function as a leaf node of the regression tree, wherein a sub-model is constructed in each node of each regression tree to simulate the relationship between the time series parameter and the normalized vegetation index.

[0052] According to some embodiments of the present disclosure, operation S202 includes: testing samples of the training sample data set using multiple regression trees, and outputting contribution rates of multiple time series parameters to the normalized vegetation index.

[0053] According to some embodiments of the present disclosure, operation S203 includes: determining whether the contribution rates of multiple time series parameters are all greater than or equal to a preset threshold; if not, removing the time series parameters whose contribution rates are less than the preset threshold from the training sample data set, and re-establishing multiple regression trees for training; if so, stopping the iteration and using the model at the time when the iteration stops as the normalized vegetation index reconstruction model.

[0054] According to some embodiments of the present disclosure, obtaining surface parameter reanalysis data corresponding to normalized vegetation index data includes: performing time window scale synthesis on data of multiple time series parameters respectively to obtain window scale surface parameter reanalysis data, wherein the time window scale synthesis includes synthesis by obtaining the cumulative value within the time window, or synthesis by obtaining the average value within the time window.

[0055] According to some embodiments of the present disclosure, the time window scale includes a monthly scale, and the normalized vegetation index reconstruction model to be obtained by training with the training sample data set is a monthly scale reconstruction model.

[0056] According to some embodiments of the present disclosure, the training sample data set includes training sets and corresponding validation sets of multiple monthly scales, and the ratio of the number of samples in the training set to that in the validation set includes 7:3.

[0057] According to some embodiments of the present disclosure, obtaining normalized vegetation index data includes operations S301 to S303.

[0058] According to some embodiments of the present disclosure, operation S301 includes: acquiring original normalized vegetation index data.

[0059] According to some embodiments of the present disclosure, operation S302 includes: processing original NDVI data by a maximum synthesis method and generating monthly NDVI data.

[0060] According to some embodiments of the present disclosure, operation S303 includes: resampling the monthly normalized vegetation index data using a bilinear interpolation method to obtain normalized vegetation index data to be used.

[0061] According to some embodiments of the present disclosure, testing samples of a training sample data set using multiple regression trees includes: counting votes of the multiple regression trees, and outputting a leaf node with the most votes as a normalized vegetation index.

[0062] According to some embodiments of the present disclosure, the training sample data set also includes a test set.

[0063] In this embodiment, after outputting the normalized vegetation index, the method further includes: using a test set and statistical indicators to evaluate the accuracy of the output normalized vegetation index, and outputting the evaluation result.

[0064] According to some embodiments of the present disclosure, the statistical indicators include any one or a combination of the following: Pearson correlation coefficient, sum of squared errors, coefficient of determination, root mean square error, mean deviation error, standard deviation and mean.

[0065] The present disclosure also provides a normalized vegetation index reconstruction method, including operations S401 and S402.

[0066] According to some embodiments of the present disclosure, operation S401 includes: acquiring surface parameter reanalysis data corresponding to the normalized difference vegetation index to be reconstructed.

[0067] According to some embodiments of the present disclosure, operation S402 includes: inputting the surface parameter reanalysis data into a normalized vegetation index reconstruction model, and outputting a reconstruction result.

[0068] In this embodiment, the normalized difference vegetation index reconstruction model is trained by the above-mentioned normalized difference vegetation index reconstruction model training method.

[0069] The present disclosure is further explained below in conjunction with specific embodiments. It should be understood that the specific embodiments are only for those skilled in the art to better understand the technical solutions of the present disclosure, and should not be construed as limiting the scope of protection of the present disclosure.

[0070] Figure 2 A block diagram schematically illustrates a training method for a normalized vegetation index reconstruction model according to an embodiment of the present disclosure.

[0071] Step 1: Data acquisition and processing:

[0072] like Figure 2 As shown, the catchment land surface model (CLSM) GLDAS-CLSM data is downloaded from the Global Land Data Assimilation System (GLDAS), wherein GLDAS utilizes advanced land surface process model simulation and data assimilation technology, integrates a large amount of surface observation and satellite remote sensing observation data, and simulates and generates global high-resolution land surface state variables and flux data. GLDAS is one of the main models used in the present disclosure, and the GLDAS-CLSM data product produced by it provides daily 0.25 degree surface parameter reanalysis data for the world since January 1948, including 33 time series parameters.

[0073] According to some optional embodiments of the present disclosure, the downloaded GLDAS-CLSM data is stored in .nc4 format. First, a program is written to batch extract TIF data files from the .nc4 file, and then the 33 variables of the daily GLDAS-CLSM data are synthesized on a monthly scale in combination with the metadata information of the data set.

[0074] In this embodiment, the monthly value data are synthesized by taking the cumulative values ​​within the time window (monthly) for rate data such as precipitation, evapotranspiration, runoff, and snowmelt; and the monthly value data are synthesized by taking the average values ​​within the time window (monthly) for state data such as flux, temperature, soil moisture, snow depth, water storage, radiation, air pressure, and wind speed.

[0075] Acquisition and processing of true NDVI data: Advanced Very High Resolution Radiometers (AVHRR) are the main sensors carried on the NOAA series of polar-orbiting environmental satellites. The Global Inventory Modeling and Mapping Studies (GIMMS) NDVI data is a long-term series of global vegetation index change data based on NOAA meteorological satellites (NOAA7, 9, 11, 14 and 16) launched in November 2003. The AVHRR GIMMS3g NDVI dataset provides global 1 / 12 degree semi-monthly NDVI from July 1981 to December 2015. This paper generates monthly NDVI through the maximum synthesis method and uses the bilinear interpolation method to resample the spatial resolution of NDVI data from 1 / 12 degree to 0.25 degree as the target dataset for machine learning training and reconstruction.

[0076] Soil and terrain environmental parameters: In addition to being closely related to climatic factors such as meteorology and hydrology, vegetation growth is also closely related to the underlying soil environment and terrain conditions. In addition to the GLDAS-2CLSM time series dynamic variables, the present disclosure also selects six terrain and soil parameter factors such as Digital Elevation Model (DEM), eco-geographical zoning, soil type map, soil texture and soil components that are closely related to the vegetation index. Considering that the spatial distribution of the above-mentioned related data varies little over time, and the update frequency of these data is low, in the present disclosure, they are used as steady-state factors together with the GLDAS-2CLSM time series dynamic variables as independent variables to input the model for training.

[0077] Step 2: Construction of NDVI reconstruction algorithm based on machine learning.

[0078] (1) Design of conceptual model for historical NDVI reconstruction,

[0079] Historical NDVI is closely related to many factors such as vegetation type, vegetation distribution, climate environment, terrain characteristics, phenological changes, and human activities, and has obvious seasonal periodicity and spatial zonality. It is difficult to effectively simulate the NDVI data of the historical period with a simple linear or second-order polynomial regression model. Therefore, this paper designs a conceptual model for historical NDVI reconstruction:

[0080] y i =f(x i ,eco,dem,soil,…,i) (1)

[0081] Where i represents the month, with a value of 1-12; yi represents the target NDVI dataset for the i-th month; x i It represents the GLDAS-2CLSM reanalysis parameters of the ith month, including environmental factors such as precipitation, temperature, air pressure, wind speed, vegetation transpiration, vegetation moisture, soil moisture, solar radiation, and ground evapotranspiration; eco represents eco-geographical zoning data; dem represents terrain parameters; soil represents soil classification, texture and other parameters.

[0082] (2) Construction of historical NDVI reconstruction algorithm based on Random Forest (RF).

[0083] Random forest is an integrated algorithm derived from classification and regression tree algorithm. Multiple classification regression trees are established by randomly selecting different independent subsets from sample data, and a model is built in each node of all independent subtrees to simulate the relationship between independent variables and dependent variables. The basic unit of RF is a decision tree, and each decision tree is a regressor. The RF algorithm specifies the category with the most votes as the final output by integrating all regression voting results. The present disclosure is based on the Python environment provided by Anaconda. First, the RandomForestRegressor regression function is imported from the sklearn class library, and the data samples and test data sets for executing machine learning training are extracted as needed. The machine learning regressor function can return the score between the independent variable and the dependent variable in the training variable, which is used as a standard for testing the accuracy of the model and is also an important basis for modifying the input parameters of the model. By referring to and comparing the parameter adjustment results based on the classesklearn.model_selection.GridSearchCV() function and the result accuracy score of the oob_score_ output, the numerical settings of each main parameter are constantly adjusted to find the optimal fit result.

[0084] In this embodiment, the random forest algorithm input parameters are set as follows: RandomForestRegressor(bootstrap=True,criterion='mse',max_depth=None,max_features='auto',max_leaf_nodes=None,min_impurity_decrease=0.0,min_impurity_split=None,min_samples_leaf=1,min_samples_split=2,min_weight_fraction_leaf=0.0,n_estimators=500,n_jobs=None,oob_score=False,random_state=None,verbose=0,warm_start=False).

[0085] Step 3: Reconstruct production of historical long-term series NDVI data.

[0086] (1) Random forest NDVI reconstruction with all variables added: The 33-year 396-period monthly 0.25-degree AVHRR GIMMS3g.v1 NDVI data from January 1982 to December 2014 were divided into training set and validation set. The random.sample() random sampling function was used for sampling in the program design process at a ratio of about 7:3. Any 24 years were selected as the training data set and input into the RF algorithm for training. The remaining 9 years of data were used as the validation data set to test the reliability of the training results. By controlling the month, the machine learning algorithm was trained on the 24 years of data on a monthly basis, and all the 20 parameters selected (Table 1) were input into the model calculation according to the corresponding month.

[0087] Table 1 List of 20 input variables for random forest model training

[0088]

[0089] (2) According to the idea of ​​going from complex to simple, first add all the prepared input variables to the training, such as Figure 3 As shown in the figure, combined with the importance of each variable given by feature_importances_ in the output result (i.e., the contribution rate to the reconstruction result), the variables with small contribution rate are gradually eliminated, and the accuracy of the training results is continuously evaluated to find the optimal input configuration and output result.

[0090] In this embodiment, if Figure 3As described above, the contribution rate of each time series parameter from January to December to the normalized vegetation index is shown in turn, among which, in each bar graph, from bottom to top, they are the contribution rate of aerodynamic conductivity, air temperature, surface air pressure, snowfall flux, rainfall flux, groundrunoff, specific humidity, wind speed, soil profile moisture, canopy water, vegetation transpiration flux, vegetation transpiration flux, ground heat flow, surface upward latent heat flux, surface upward latent heat flux, HWSD dominant soil components, WWF eco-geographical zoning, HWSD surface soil clay content, HWSD surface soil gravel content and HWSD surface soil USDA texture classification to the normalized vegetation index.

[0091] Taking the output results of January and June as an example, the contribution rate of TFV (vegetation transpiration flux) in January is the lowest, and the contribution rate of SSR (vegetation transpiration flux) in June is the lowest. They are removed and recalculated.

[0092] Step 4: Accuracy evaluation and analysis of NDVI reconstruction results.

[0093] (1) Accuracy evaluation method of NDVI reconstruction results based on validation dataset

[0094] According to the reconstruction process proposed in step 3, the NDVI reconstruction of each month is randomly assigned to the training set and the validation set. Therefore, the reconstruction results are first evaluated and compared using the 9-year validation set NDVI data for the time period with validation set data. This paper uses the Pearson correlation coefficient (r), the sum of squared errors (Sum of the Squared Errors, SSE), and the coefficient of determination (R 2 ), Root Mean Square Error (RMSE), Mean Bias Error (MBE), Standard Deviation (STD) and Average Value (AVG) and other statistical indicators are quantitatively evaluated, calculated and analyzed based on the validation data set. Among them, the Pearson correlation coefficient is used to measure the correlation between two variables. The present disclosure uses the correlation coefficient to measure the correlation between the reconstructed NDVI and the validation data set. The sum of squared errors is also called the residual sum of squares. The square of the residual between all samples and the corresponding fitted values ​​is used to measure the quality of the fitting results. The error between the reconstructed results and the validation NDVI data set can be quantitatively evaluated. The coefficient of determination (also called goodness of fit) reflects the proportion of the variation that can be explained by the independent variable to the total variation, generally expressed as R 2It is generally equivalent to the square of the correlation coefficient r in terms of value. The root mean square error is the square root of the ratio of the sum of the squares of the deviations between the observed value and the true value to the number of observations. The mean deviation error is the ratio of the sum of the deviations between all sample observations and the true value to the sample size. It is used in this disclosure to measure the deviation between the reconstructed NDVI and the GIMMS3g NDVI. At the same time, this disclosure uses the standard deviation and the mean value to measure the degree of dispersion between the reconstructed NDVI data.

[0095] Formulas (2)-(8) respectively give the calculation methods of various statistical indicators:

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102]

[0103] Among them, X i Refers to the GIMMS3g NDVI data sample used for verification, is the average value of all GIMMS3g NDVI samples involved in the calculation, Y i Refers to NDVI data samples reconstructed through machine learning, is the average value of all reconstructed NDVI samples involved in the calculation, and n is the total number of samples involved in the test calculation.

[0104] (2) Accuracy evaluation and analysis of NDVI reconstruction results based on medium-scale remote sensing images.

[0105] For the reconstructed NDVI data from 1972 to 1981, this paper selected Landsat 1-3 satellite images with overlapping time windows for accuracy evaluation and analysis. The multi-spectral scanner (MSS) sensor band range is 0.5-1.1μm, which includes the visible red band sensitive to vegetation chlorophyll content and the near-infrared band sensitive to vegetation structure and morphology. The NDVI data based on Landsat MSS images can be obtained by band calculation. By obtaining the relevant statistical indicators given by formulas (2)-(8) between Landsat MSS NDVI and the reconstructed NDVI of the corresponding month, the scatter point density between Landsat MSS NDVI data and the corresponding reconstructed NDVI is calculated month by month to evaluate the accuracy and consistency of the reconstructed historical NDVI data.

[0106] (3) Accuracy evaluation and analysis of NDVI reconstruction results based on drought index.

[0107] Existing studies have shown that natural factors such as precipitation and evapotranspiration are not only closely related to vegetation changes, but also important factors that determine the occurrence and change of meteorological drought. Therefore, this paper attempts to mine and analyze the consistency or correlation of the NDVI time series and the trend of drought index changes as a means of verifying the results of long-term NDVI reconstruction. The Global SPEI database (Global SPEI datebase, SPEIbase) (http: / / spei.csic.es / database.html) provides long-term inversion data of 0.5 degree monthly drought conditions on a global time scale of 1-48 months. Considering the persistence of drought, this paper uses the SPEI data calculated based on a 3-month time scale provided by the SPEIbase v2.5 data set to test the changing trend of the reconstructed NDVI data, and mines the long-term data change pattern, periodicity and outlier test between the reconstructed NDVI and the drought index through time series decomposition, so as to carry out the consistency and reliability analysis of the historical reconstructed NDVI data from a qualitative perspective.

[0108] Through the above-mentioned embodiments of the present disclosure, the factors affecting NDVI changes are determined by studying the relationship between hydrothermal environmental conditions and spatial distribution characteristics of vegetation, and then a random forest-based NDVI reconstruction model is designed and implemented based on reanalysis data and multi-source remote sensing data. The optimal reconstruction parameters are determined by test set accuracy evaluation and control of input parameters, and a training method for a normalized vegetation index reconstruction model based on machine learning is obtained.

[0109] Based on the reanalysis of historical record data and public remote sensing vegetation index data, this paper designs and implements a parameter-controllable NDVI data reconstruction model based on an optimized machine learning regression algorithm, which can reconstruct and produce monthly NDVI data products since 1948, filling the gap in vegetation index data before the maturity of remote sensing technology.

[0110] The present invention comprehensively considers multiple comprehensive factors such as climate, soil, ecology, and environment in the process of machine learning. It is currently the most comprehensive historical vegetation index reconstruction method, which solves the limitations of traditional algorithms in the application of large areas and long periods of time.

[0111] So far, the embodiments of the present disclosure have been described in detail in conjunction with the accompanying drawings. It should be noted that the implementation methods not shown or described in the drawings or the body of the specification are all forms known to ordinary technicians in the relevant technical field and are not described in detail. In addition, the above definitions of each component are not limited to the various specific structures, shapes or methods mentioned in the embodiments, and ordinary technicians in the field can simply change or replace them.

[0112] It should also be noted that, in the specific embodiments of the present disclosure, unless otherwise indicated, the numerical parameters in the present specification and the appended claims are approximate values ​​and can be changed according to the desired characteristics obtained through the content of the present disclosure. Specifically, all numbers used in the specification and claims to express the size, range conditions, etc. of the composition should be understood to be modified by the term "about" in all cases. In general, the meaning of the expression refers to the change of ±10% in some embodiments, ±5% in some embodiments, ±1% in some embodiments, and ±0.5% in some embodiments by a specific number.

[0113] It will be appreciated by those skilled in the art that the features described in the various embodiments and / or claims of the present disclosure may be combined and / or combined in a variety of ways, even if such combinations and / or combinations are not explicitly described in the present disclosure. In particular, the features described in the various embodiments and / or claims of the present disclosure may be combined and / or combined in a variety of ways without departing from the spirit and teachings of the present disclosure. All of these combinations and / or combinations fall within the scope of the present disclosure.

[0114] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present disclosure. It should be understood that the above description is only a specific embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A training method for a normalized vegetation index reconstruction model, characterized in that: include: Acquire normalized vegetation index data and corresponding surface parameter reanalysis data, wherein the surface parameter reanalysis data includes a plurality of time series parameters, wherein the plurality of time series parameters include a steady-state factor and a dynamic variable, and the spatial distribution of the steady-state factor varies less than the spatial distribution of the dynamic variable over time; Generating a training sample data set according to the surface parameter reanalysis data and the normalized difference vegetation index; and Using the training sample data set to train the normalized vegetation index reconstruction model to be trained, to obtain the normalized vegetation index reconstruction model; The steady-state factors include any one or a combination of the following: digital elevation model, dominant soil components, eco-geographical zoning, topsoil clay content, topsoil gravel content, and topsoil USDA texture classification; and The dynamic variables include any one or a combination of the following: aerodynamic conductivity, air temperature, surface air pressure, snowfall flux, rainfall flux, underground runoff, specific humidity, wind speed, soil profile moisture, canopy water content, vegetation transpiration flux, surface downward shortwave radiation net flux, ground heat flux, surface upward latent heat flux, and surface upward latent heat flux; Wherein, using the training sample data set to train the normalized vegetation index reconstruction model to be trained includes: The iteration does the following: Randomly selecting samples from the training sample data set to establish multiple independent subsets, establishing multiple regression trees based on each subset, importing random forest regression functions as leaf nodes of the regression trees, wherein a sub-model is constructed in each node of each regression tree to simulate the relationship between the time series parameter and the normalized vegetation index; Using the multiple regression trees to test samples of the training sample data set, and outputting contribution rates of the multiple time series parameters to the normalized difference vegetation index respectively; and Determine whether the contribution rates of the multiple time series parameters are all greater than or equal to a preset threshold; if not, remove the time series parameters whose contribution rates are less than the preset threshold from the training sample data set, and re-establish multiple regression trees for training; if yes, stop the iteration and use the model at the time of the iteration stop as the normalized vegetation index reconstruction model; Among them, the reanalysis data of surface parameters corresponding to the normalized difference vegetation index data include: The data of multiple time series parameters are synthesized at a time window scale respectively to obtain the surface parameter reanalysis data at a window scale, wherein the time window scale synthesis includes synthesis by obtaining the cumulative value within the time window, or synthesis by obtaining the average value within the time window.

2. The method according to claim 1, characterized in that The time window scale includes a monthly scale, and the normalized vegetation index reconstruction model to be obtained by training with the training sample data set is a monthly scale reconstruction model.

3. The method according to claim 2, characterized in that The training sample data set includes training sets of multiple monthly scales and corresponding validation sets, and the ratio of the number of samples in the training set to that in the validation set is 7:

3.

4. The method according to claim 1, characterized in that: Obtaining normalized difference vegetation index data includes: Get the original normalized difference vegetation index data; Processing the original NDVI data by the maximum value synthesis method and generating monthly NDVI data; and The bilinear interpolation method is used to resample the monthly NDVI data to obtain the NDVI data to be used.

5. The method according to claim 1, characterized in that Testing the samples of the training sample data set using the multiple regression trees includes: The votes of the multiple regression trees are counted, and the leaf node with the most votes is output as the normalized vegetation index.

6. The method according to claim 5, characterized in that The training sample data set also includes a test set; After outputting the normalized difference vegetation index, the method further includes: using the test set and statistical indicators to evaluate the accuracy of the output normalized difference vegetation index, and outputting the evaluation result; The statistical indicators include any one or a combination of the following: Pearson correlation coefficient, error sum of squares, determination coefficient, root mean square error, mean deviation error, standard deviation and mean value.

7. A normalized vegetation index reconstruction method, characterized in that: include: Obtaining reanalysis data of surface parameters corresponding to the normalized difference vegetation index to be reconstructed; The surface parameter reanalysis data is input into a normalized vegetation index reconstruction model, and a reconstruction result is output, wherein the normalized vegetation index reconstruction model is trained by the training method for the normalized vegetation index reconstruction model according to any one of claims 1 to 6.

Images