Construction Method of GPP Estimation Model

Abstract

The present invention discloses a method for constructing a GPP estimation model, comprising: obtaining sunlight-induced chlorophyll fluorescence (SIF) product data and a local corresponding GPP database, combining the data with vegetation data and meteorological data of the study area to form a data set, preprocessing the data set, and dividing the processed data into a training set, a validation set, and a test set; selecting an existing GPP model, adding SIF characteristic factors, vegetation characteristic factors, and meteorological characteristic factors to the existing GPP model using an XGBOOST model, and training the parameters of the XGBOOST model using a training set; validating and further adjusting the parameters of the XGBOOST model using a validation set; resampling the test set and performing random bootstrapping selection, importing the test set into the XGBOOST model obtained in step 3 to perform traversal iteration to obtain an optimal GPP estimation model. The present invention takes into account characteristic factors such as vegetation and meteorology in the existing SIF-GPP relationship, breaks the regional limitations of the application of the linear SIF-GPP model, and improves the efficiency and accuracy of GPP measurement.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a method for constructing a GPP estimation model. Background Art

[0002] Currently, the main method for estimating terrestrial ecosystem carbon sinks relies on quadrat measurements, which are extremely time-consuming and labor-intensive, subject to significant representational errors, and fail to reflect the dynamics of ecosystem carbon sinks. Therefore, developing new, large-scale, high-precision, and low-cost methods for estimating terrestrial ecosystem carbon sinks has important academic and socioeconomic value. Gross primary productivity (GPP) of terrestrial ecosystems refers to the total amount of organic carbon fixed by green plants through photosynthesis per unit time and area, also known as gross primary productivity. Net carbon sinks are calculated by deducting factors such as autotrophic and heterotrophic respiration and disturbance from GPP. While established models exist for calculating respiration, the primary uncertainty in net carbon sink estimates stems from GPP estimates. Sunlight-induced chlorophyll fluorescence (SIF), a byproduct of photosynthesis, can more quickly and accurately reflect changes in the physiological state of vegetation. In recent years, SIF has been shown to accurately estimate GPP at a single point scale. However, the SIF-GPP relationship is influenced by multiple factors, such as temperature, humidity, vegetation type, drought, and even observation time, leading to significant uncertainty in surface GPP estimation. Therefore, it is necessary to address the dynamics of the SIF-GPP relationship under different stress factors in order to more accurately estimate GPP. Summary of the Invention

[0003] The purpose of the present invention is to address the shortcomings of the existing technology and provide a method for constructing a GPP estimation model. This method takes into account characteristic factors such as vegetation and meteorology in the existing SIF-GPP relationship, breaks the regional limitations of the application of the linear SIF-GPP model, and improves the efficiency and accuracy of GPP measurement.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0005] A method for constructing a GPP estimation model comprises the following steps:

[0006] Step 1: Obtain sunlight-induced chlorophyll fluorescence (SIF) product data and the corresponding local GPP database, then combine them with vegetation data and meteorological data of the study area to form a dataset. Preprocess the data in the dataset to remove abnormal data, and divide the processed data into training set, validation set, and test set.

[0007] Step 2: Select an existing GPP model and use the XGBOOST model to add SIF characteristic factors, vegetation characteristic factors, and meteorological characteristic factors to the existing GPP model to improve the existing GPP model and construct a GPP model suitable for large-scale use. Use the training set to train the parameters of the XGBOOST model.

[0008] Step 3: Use the validation set to verify and further adjust the parameters of the XGBOOST model to obtain an XGBOOST model with the required accuracy;

[0009] Step 4: Resample the test set and perform random bootstrapping, import it into the XGBOOST model obtained in step 3, and iterate to obtain the optimal GPP estimation model.

[0010] Furthermore, the vegetation data in step 1 include vegetation type and soil moisture, and the meteorological data include precipitation, downward radiation, average temperature, maximum temperature, and minimum temperature.

[0011] Furthermore, the existing GPP model selected is:

[0012]

[0013] Where G PP is the total primary productivity; S IF is chlorophyll fluorescence; L UE is the light energy utilization rate; L UEf is the fluorescence quantum yield; f esc is the canopy escape rate of chlorophyll fluorescence;

[0014] The above GPP model is transformed to obtain the transformed SIF-GPP model:

[0015] G PP =S IF T MINscalar V PDscalar a+b

[0016] Where, a, b are constants obtained by model fitting; T MINscalar is the minimum temperature regulation coefficient, V PDscalar is the saturated vapor pressure difference adjustment coefficient;

[0017] The XGBOOST model was used to add vegetation characteristic factors and meteorological characteristic factors to the above-mentioned transformed SIF-GPP model.

[0018] Furthermore, in step 2, the XGBOOST model is:

[0019]

[0020] Among them, FS is the decision tree set, x i is the vector composed of the eigenvalues of the i-th data, including vegetation data and meteorological data, f n (x i ) is the nth independent decision tree, which contains the structure and weight information of the tree, N is the total number of decision trees, is the predicted value of the i-th data.

[0021] Furthermore, in step 2, the XGBOOST model is trained using a ten-fold cross-validation method, and the loss function during training is:

[0022]

[0023] Where M represents the number of training sets. is the predicted value and the true value y i Here, the mean square error is used as the loss function, Ω(f n ) is the regularization term of the decision tree.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention combines the advantages of high GPP estimation accuracy and simple algorithm of SIF with strong robustness, wide applicability and increased application area after training of XGBOOST, and obtains a GPP estimation model that only needs to input the local SIF value and other selected characteristic values (vegetation type, soil moisture, precipitation, downward radiation, average temperature, maximum temperature, minimum temperature) to directly obtain a GPP estimation value with high precision and high robustness. This improves the Chinese regional research on the GPP dataset based on sunlight-induced chlorophyll fluorescence (SIF), breaks the regional limitations of the use of the linear SIF model, and obtains a GPP estimation model with high precision and high robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of a method for constructing a GPP estimation model according to an embodiment of the present invention;

[0026] Figure 2 This is a line graph showing the changing trends of SIF and GPP at various sites from 2008 to 2013, according to an embodiment of the present invention. (a) is the KFS site, (b) is the Kon site, (c) is the Ne-1 site, and (d) is the Ne-3 site.

[0027] Figure 3 Schematic diagram of the influence of each characteristic factor on the GPP results in an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram of establishing a GPP model and GPP estimation results using the XGBOOST model in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0030] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0031] The present invention will be further described below with reference to specific examples, but they are not intended to limit the present invention.

[0032] like Figure 1 As shown, an embodiment of the present invention proposes a method for constructing a GPP estimation model. First, the existing SIF and corresponding GPP, vegetation data, and meteorological data are used to train an XGBOOST model to obtain the importance of eigenvalues and corresponding weights and adjust the parameters. The SIF and eigenvalues of the study area are input into the trained XGBOOST model to obtain a GPP prediction model for the study area. This improves the Chinese regional research based on the GPP dataset of sunlight-induced chlorophyll fluorescence (SIF), breaks the regional limitations of the application of the linear SIF model, and obtains a highly accurate and robust GPP estimation model. The method specifically includes the following steps:

[0033] Step 1: Obtain sunlight-induced chlorophyll fluorescence (SIF) product data and the corresponding local GPP database, then combine them with vegetation data and meteorological data of the study area to form a dataset, preprocess the data in the dataset, and divide the processed data into training set, validation set, and test set;

[0034] In this embodiment, the SIF740 product measured by the GOME-2 sensor on the MetOp-A satellite, the existing 2000-2017 global SIF product data GOSIF obtained by processing MODIS and MERRA-2 data, and the local corresponding GPP database (MODIS data may be considered), are combined with local vegetation data and meteorological data to form a data set. The prepared data includes vegetation type and soil moisture, and the meteorological data includes precipitation, downward radiation, average temperature, maximum temperature, and minimum temperature.

[0035] Next, preprocess the data. Consider using Python. During data preprocessing, first preview the data to review the data file and understand the characteristics of each column. This will facilitate specifying the appropriate type when reading and speed up the data. Create a code for the category and digitize the code using the create_num function. However, the original df_raw file must be retained. When interpreting the model, the digitized code must be converted back to the original code created by the category. After data preprocessing, remove outliers, convert the data to a format that is more convenient for subsequent operations, and store it as the training set.

[0036] Here you can consider using autonomous sampling, holdout, and cross-validation methods to construct training sets, validation sets, and test sets to provide data sets for subsequent training, validation, and testing.

[0037] Step 2: Select an existing GPP model and use the XGBOOST model to add SIF characteristic factors, vegetation characteristic factors, and meteorological characteristic factors to the existing GPP model to improve the existing GPP model and construct a GPP model suitable for large-scale use. Use the training set to train the parameters of the XGBOOST model.

[0038] In this embodiment, the existing reference chlorophyll fluorescence GPP estimation model is selected as:

[0039]

[0040] Where G PP ——Gross primary productivity, g·c / (m 2 d)

[0041] S IF ——Chlorophyll fluorescence, m·W / (m 2 ·nm·sr)

[0042] L UE ——Light energy utilization rate, %

[0043] L UEf ——fluorescence quantum yield, %

[0044] f esc -Canopy escape rate of chlorophyll fluorescence

[0045] Since the far infrared band is less affected by the canopy, f esc The default value is usually 1. According to the calculation method used by MODl7 products, there are:

[0046] L UE =L UEmax T MINscalar VPDscalar (2)

[0047] The transformed SIF-GPP model is constructed as:

[0048] G PP =S IF T MINscalar V PDscalar a+b (3)

[0049] In the formula, a, b are constants obtained by model fitting

[0050] Among them, T MINscalar Minimum temperature regulation coefficient and V PDscalar The calculation formula of the saturated water vapor pressure difference adjustment coefficient is as follows:

[0051] L UE =L UEmax T MINscalar V PDscalar (4)

[0052] in,

[0053]

[0054] Where, L UEmax ——Maximum light energy utilization rate, %

[0055] T MINscalar ——Minimum temperature adjustment coefficient

[0056] T MIN ——Lowest temperature of the day, ℃

[0057] T MINmin ——The lowest temperature when the vegetation light energy utilization rate is 0, ℃

[0058] T MINmax ——The lowest temperature when vegetation light energy utilization reaches its maximum value, ℃

[0059] V PDscalar ——Saturated vapor pressure difference adjustment coefficient

[0060] V PD ——Daytime average saturated water vapor pressure difference, Pa

[0061] V PDmin ——The average saturated water vapor pressure difference during the day when the vegetation light energy utilization rate is 0, Pa

[0062] V PDmax ——The average saturated water vapor pressure difference during the day when the vegetation light energy utilization rate is the highest, Pa

[0063] L UEmax 、TMINmin 、T MINmax 、V PDmin 、V PDmax This can be found in the MOD17 product description. Figure 2 As shown in Figure 3, it can be seen that under certain time and space constraints, the changes in SIF and GPP are highly consistent, providing a reference for formula (3).

[0064] The reference chlorophyll fluorescence GPP estimation model was transformed to obtain the linear model of SIF-GPP (Formula (3), small area). In order to make it applicable to a wider range, the XGBOOST model was used to add characteristic factors (vegetation, meteorological factors) to the corresponding relationship of SIF-GPP so that it presents a nonlinear relationship in a large area.

[0065] After determining the idea, start XGBOOST model training. XGBOOST is a machine learning algorithm developed from the GBDT (GradientBoosting Decision Tree) algorithm. It optimizes the objective function based on ensemble learning and combines gradient information to obtain the optimal solution. In this embodiment, vegetation characteristic factors and meteorological characteristic factors are added to the SIF-GPP model to optimize the SIF-GPP model. The XGBOOST algorithm is composed of a series of decision trees:

[0066]

[0067] Among them, FS (Forest Sets) is the decision tree set, x i is a vector composed of the eigenvalues of the i-th data, including local vegetation type, soil moisture, precipitation, downward radiation, average temperature, maximum temperature, and minimum temperature data. n (x i ) is the nth independent decision tree, which contains the structure and weight information of the tree, N is the total number of decision trees, is the predicted value of the i-th data.

[0068] The loss function during training is defined as:

[0069]

[0070] Where M represents the number of training sets, is the predicted value and the true value y i Here, the mean square error is used as the loss function, Ω(f n ) is the regularization term of the decision tree.

[0071] The parameters of XGBOOST can be continuously obtained through iteration, where the loss function of the t-th round of iteration can be expressed as formula (9):

[0072]

[0073] In order to simplify the calculation process, Substituting formula (9) into formula (10) we can obtain:

[0074]

[0075] Performing a second-order Taylor expansion on formula (10) yields formula (11);

[0076]

[0077] where Δf t (x i )=f t (x i )-f t-1 (x i ), due to l(y i ,f t-1 (x i )) is the error value in the previous iteration, which can be eliminated and formula (11) can be expanded to formula (12):

[0078]

[0079] in, Will Substituting into formula (12), we can obtain formula (13);

[0080]

[0081] Since the samples are divided into T decision trees, formula (13) can be rewritten as formula (14):

[0082]

[0083] where w j is the leaf weight of the j-th decision tree, i∈I jRepresents the samples belonging to the decision tree node j, λ is the parameter of the regularization term, which can control the score of the leaf node from being too large to prevent overfitting; λ (lambda) generally has a default value of 1. The larger the parameter, the less likely the model is to overfit; γ also represents the parameter of the regularization term to control the number of leaf nodes. γ (gamma) generally has a default value of 0. The minimum loss required for further partitioning on the leaf nodes of the tree is reduced. The larger its value, the more conservative the algorithm. Applying formula (14) to w can obtain the optimal estimated value of the leaf weight of each decision tree.

[0084]

[0085] Input the training set and train the XGBOOST model according to the ten-fold cross-validation method, that is, divide the training set into 10 parts on average, use nine of the ten samples as the training set for each training, and use the remaining sample as the test set for verification and output the training accuracy. Repeat the training ten times and output the final training accuracy. After training, the influence of each feature factor on the GPP result is obtained as follows Figure 3 As shown, from Figure 3 It can be seen that the weight example diagram of each feature factor obtained after training the XGBOOST model, the weight factor reference formula is shown in formula (15).

[0086] Step 3: Use the validation set to verify and further adjust the parameters of the XGBOOST model to obtain an XGBOOST model with the required accuracy;

[0087] The num_boost_round parameter in the XGBOOST model uses its built-in xgb.cv function, initially giving it a sufficiently large initial value for training, and then waiting for it to return the optimal result value.

[0088] Other parameters can be tuned using the grid training tool GridSearchCV in sklearn.

[0089] The max_depth and min_weight parameters have a great impact on the results, so they should be tuned first, with a large-scale initial adjustment followed by a small-scale adjustment.

[0090] The subsample and colsample_bytree parameters can also be tuned. Starting at 0.6, gradually adjust towards 1. Smaller values can lead to overfitting.

[0091] The validation set is used to verify the parameter adjustment results and update the parameters until the accuracy of the XGBOOST model reaches the target value. In this embodiment, if the average accuracy of the XGBOOST model is higher than 0.9, the output accuracy meets the XGBOOST model.

[0092] Step 4: The test group data is resampled using the ADASYN method, and random selection is performed by bootstrapping. The data is then imported into the XGBOOST model output in step 3 and iterated to obtain the optimal GPP model.

[0093] See Figure 4 In this embodiment, the SIF and eigenvalues of the test set (local vegetation type, soil moisture, precipitation, downward radiation, average temperature, maximum temperature, and minimum temperature data) are resampled using the ADASYN method, and random bootstrap selection is performed. The XGBOOST model obtained in step 3 is imported and iterated to obtain the optimal GPP estimation model. The optimal GPP model is then used to predict the GPP of the study area, thereby obtaining the GPP value.

[0094] This step completed the establishment of the GPP estimation model based on sunlight-induced chlorophyll fluorescence and the XGBOOST algorithm, improved the Chinese regional research on the GPP dataset based on sunlight-induced chlorophyll fluorescence (SIF), broke the regional limitations of the application of the linear SIF model, improved the efficiency and accuracy of GPP measurement, and obtained a highly accurate and robust GPP estimation model.

[0095] The above are only preferred embodiments of the present invention and do not limit the implementation mode and protection scope of the present invention. For those skilled in the art, it should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the contents of the present invention specification should be included in the protection scope of the present invention.

Claims

1. A method for constructing a GPP estimation model, characterized in that: The steps include: Step 1: Obtain sunlight-induced chlorophyll fluorescence (SIF) product data and the corresponding local GPP database, then combine them with vegetation data and meteorological data of the study area to form a dataset. Preprocess the data in the dataset to remove abnormal data, and divide the processed data into training set, validation set, and test set. Step 2: Select an existing GPP model and use the XGBOOST model to add SIF characteristic factors, vegetation characteristic factors, and meteorological characteristic factors to the existing GPP model to improve the existing GPP model and construct a GPP model suitable for large-scale use. Use the training set to train the parameters of the XGBOOST model. Step 3: Use the validation set to verify and further adjust the parameters of the XGBOOST model to obtain an XGBOOST model with the required accuracy; Step 4: resample the test set and perform random bootstrapping, import it into the XGBOOST model obtained in step 3, and iterate to obtain the optimal GPP estimation model; The vegetation data in step 1 include vegetation type and soil moisture, and the meteorological data include precipitation, downwelling radiation, average temperature, maximum temperature, and minimum temperature; The existing GPP models selected are: ； Where, is the gross primary productivity; is chlorophyll fluorescence; is the light energy utilization rate; is the fluorescence quantum yield; is the canopy escape rate of chlorophyll fluorescence; The above GPP model is transformed to obtain the transformed SIF-GPP model: Where, ——Constants obtained from model fitting; is the minimum temperature regulation coefficient, is the saturated vapor pressure difference adjustment coefficient; The XGBOOST model was used to add vegetation characteristic factors and meteorological characteristic factors to the above-mentioned transformed SIF-GPP model; In step 2, the XGBOOST model is: ； in, is a decision tree ensemble, is the vector composed of the eigenvalues of the i-th data, including vegetation data and meteorological data. is the nth independent decision tree, which contains the tree structure and weight information, is the total number of decision trees, is the predicted value of the i-th data; In step 2, the XGBOOST model is trained using a 10-fold cross-validation method. The loss function during training is: ； Where M represents the number of training sets. is the predicted value and the true value Here we use the mean square error as the loss function. is the regularization term of the decision tree.

Images