Spatial distribution prediction method for soil respiration temperature sensitivity

Abstract

The invention discloses a spatial distribution prediction method for soil respiration temperature sensitivity. The method comprises the following steps: step 1, data acquisition and processing; 2, calculating parameters of a Van't Hoff index chemical reaction-temperature equation; 3, calculating the soil respiration temperature sensitivity Q10; step 4, acquiring and processing influence factor data of the soil respiration temperature sensitivity Q10; step 5, building a deep neural network model and training the deep neural network model; and step 6, determining the spatial distribution of the soil respiration temperature sensitivity Q10 in the research area. According to the method disclosed by the invention, the availability of the existing data is greatly enriched, and the capability of extracting effective information from the soil respiration data is remarkably improved; meanwhile, on the basis of enriching data availability, a deep neural network model is adopted, a nonlinear relation between the soil respiration temperature sensitivity Q10 and influence factors of the Q10 is constructed, the Q10 of a point scale is successfully expanded to a regional scale, and spatial distribution prediction of the soil respiration temperature sensitivity is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of environmental monitoring, and in particular relates to a spatial distribution prediction method for soil respiration temperature sensitivity. Background Art

[0002] The temperature sensitivity of soil respiration indicates the response of soil respiration to temperature changes, usually expressed as the multiple of soil respiration change when the temperature rises by 10°C (Q 10 ) value. 10 It is a key parameter for determining global terrestrial carbon dynamics under climate change. The higher its value, the more sensitive the soil respiration rate is to temperature changes. 10 The accuracy of the value of Q will determine the accuracy of Earth system models in predicting the strength of the carbon-climate feedback. 10 and estimating their spatial distribution is an important prerequisite for the ecosystem modeling process.

[0003] Currently, Q 10 The value is estimated by fitting the function using the monitored soil respiration flux data and temperature data. 10 The determination of Q values ​​relies heavily on continuously monitored soil respiration flux and temperature data. However, the reality is that the available continuous monitoring data are very limited, which will result in the Q values ​​available for analysis. 10 The data are seriously insufficient. Soil respiration carbon emissions represent the amount of carbon dioxide released by soil respiration over a period of time. Carbon emission data also contain key information about soil respiration, which can supplement the limited continuous monitoring data and be used to calculate Q 10 However, there is no method to calculate Q from soil respiration carbon emission data. 10 In addition, improving Q 10 The availability of data will also facilitate Q 10 There are methods that can predict the spatial distribution of Q in a region. 10 The spatial distribution of carbon is predicted by terrestrial carbon models, climate driven models, and random forests, but there is a lack of effective Q 10 The validation of these values ​​leads to large uncertainties, especially in areas where data are scarce. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting the spatial distribution of soil respiration temperature sensitivity to solve the above technical problems.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The present invention discloses a method for predicting the spatial distribution of soil respiration temperature sensitivity, which comprises the following steps:

[0007] Step 1, data acquisition and processing: obtain soil respiration monitoring data and daily temperature data set in the study area. The soil respiration monitoring data includes two types of data. The first type of data is the daily average flux data of soil respiration, and the second type of data is the carbon emission data of soil respiration. According to the location and time information of the acquired soil respiration monitoring data, extract the daily average temperature data from the daily temperature data set;

[0008] Step 2: Calculate the parameters of the Van't Hoff exponential chemical reaction-temperature equation: The Van't Hoff exponential chemical reaction-temperature equation is used to describe the relationship between the daily average flux of soil respiration and the daily average temperature:

[0009] Rs=ae bT (1)

[0010] Where Rs is the daily average flux of soil respiration, in μmolCm -2 s -1 ; T is the daily average temperature, in °C; a is the first empirical parameter, b is the second empirical parameter;

[0011] For the first type of data described in step 1, the Van't Hoff exponential chemical reaction-temperature equation, i.e., formula (1), is fitted using the least square method to determine the first empirical parameter a and the second empirical parameter b of the equation;

[0012] Among them, the least squares method finds the best matching function for fitting data by minimizing the sum of squares of errors, and its loss function is as follows:

[0013]

[0014] Where L is the loss function of the least squares method, that is, the sum of squared errors; (x i ,y i ) is the i-th sample; f() is the theoretical function of the unknown parameters;

[0015] For the second type of data described in step 1, the first empirical parameter a and the second empirical parameter b are solved by the following method;

[0016] According to the Van't Hoff exponential chemical reaction-temperature equation, the daily average flux of soil respiration during the monitoring period is expressed as:

[0017]

[0018] In the formula, Rs i represents the daily average flux of soil respiration on the i-th day; T i represents the daily average temperature on the i-th day;

[0019] The daily soil respiration emission was calculated based on the daily average flux of soil respiration, and the formula is:

[0020]

[0021] In the formula, is the daily soil respiration emission on the ith day, in gCm -2 ; m represents the mass of carbon per micromole, which is 12×10 -6 g μmol -1 ; t represents the number of seconds in a day, which is 86400s; η is the conversion factor, which is 1.0368, and the unit is gμmol -1 s;

[0022] The cumulative value of soil respiration carbon emissions during the monitoring period is calculated based on the daily emissions of soil respiration. The formula is:

[0023]

[0024] In the formula, P represents the number of days of monitoring time, the unit is d; is the cumulative value of soil respiration carbon emission within P days, in gCm -2 ;

[0025] When there are two soil respiration carbon emission data at different monitoring times at the same location, the values ​​of parameters a and b are obtained by solving the following nonlinear equations:

[0026]

[0027] In the formula, m and n represent different monitoring days, and the unit is d;

[0028] Step 3: Calculate soil respiration temperature sensitivity Q 10 : Based on the parameters of the Van't Hoff exponential chemical reaction-temperature equation obtained in step 2, calculate the soil respiration temperature sensitivity Q 10 , the calculation formula is:

[0029] Q 10 =e 10b (7)

[0030] In the formula, Q 10 is the multiple of soil respiration change when the temperature rises by 10℃, which is used to indicate the temperature sensitivity of soil respiration; b is the second empirical parameter;

[0031] Step 4: Obtain soil respiration temperature sensitivity Q 10 The impact factor data of the study area is obtained and processed: 10 The impact factor data of the obtained Q 10 The impact factor data of Q10 The impact factor data are resampled to the required resolution, then outliers are removed and missing values ​​are interpolated;

[0032] Step 5: Build and train a deep neural network model: This includes the following steps:

[0033] Step 51: Q calculated in step 3 10 is the label, with the calculated Q 10 Q of location 10 The influencing factors are used as features to form a training data set;

[0034] Step 52: Q in the training data set 10 and Q 10 Normalize the impact factor data to generate a normalized training data set;

[0035] Step 53: Build a deep neural network model and add L1 or L2 regularization to prevent the deep neural network model from overfitting;

[0036] Step 54: Use K-fold cross validation combined with grid search to determine the optimal hyperparameter combination;

[0037] Step 55: Use the normalized training data set to train a deep neural network model to establish Q 10 With Q 10 The mapping relationship between the influencing factors is used to obtain the predicted Q 10 The regression model is the trained deep neural network model;

[0038] Step 6: Determine the temperature sensitivity of soil respiration in the study area Q 10 Spatial distribution of: Q of all pixels at the required resolution in the study area 10 The impact factor data of the study area is used as the prediction data set, and the prediction data set is normalized to generate a normalized prediction data set; the generated normalized prediction data set is input into the trained deep neural network model to obtain the normalized Q 10 value; the normalized Q 10 The value was inversely normalized to obtain the soil respiration temperature sensitivity Q in the entire study area. 10 The value is used to determine the temperature sensitivity of soil respiration in the study area, Q 10 spatial distribution.

[0039] Furthermore, the soil respiration monitoring data in step 1 is obtained through literature, public databases or site monitoring; the daily temperature dataset comes from a public raster dataset, and the raster dataset is remote sensing data or reanalysis data.

[0040] Furthermore, in step 4, Q 10 The influencing factor data include annual average temperature, annual precipitation, normalized difference vegetation index NDVI, pH, soil organic carbon density SOCD, elevation DEM, and soil property data.

[0041] Furthermore, the soil property data includes soil bulk density, clay content, sand content, and silt content.

[0042] Furthermore, the required resolution in step 4 is 1 km×1 km.

[0043] Furthermore, the specific process of building the deep neural network in step 53 is: setting the number of nodes in the input layer, the number of hidden layers, the number of neurons in the hidden layer, the output layer, the activation function, the loss function, the optimizer, the learning rate, the number of training times, and the L1 or L2 regularization parameter of the deep neural network; the number of nodes in the input layer and Q 10 The number of impact factors is related to

[0044] Among them, L1 and L2 regularization limits the complexity of the model to prevent the model from overfitting on the training data, thereby improving the generalization ability of the model;

[0045] L1 Regularization:

[0046]

[0047] L2 Regularization:

[0048]

[0049] In the formula, and They are the loss functions of L1 and L2 regularization respectively; is the data loss of the model; λ is the regularization parameter; W is the weight matrix; || ||1 and || ||2 are the L1 and L2 norms respectively.

[0050] Furthermore, the specific process of using K-fold cross validation combined with grid search to determine the optimal hyperparameter combination in step 54 is:

[0051] (1) Identify the hyperparameters that need to be adjusted and their possible values, and create a hyperparameter grid, that is, possible combinations of hyperparameter values;

[0052] (2) Perform grid search and K-fold cross validation on all possible hyperparameter combinations;

[0053] The specific steps of the K-fold cross validation are:

[0054] 1) Split the entire data set into K subsets;

[0055] 2) Select one subset for validation without duplication, and use the remaining K-1 subsets for training the model;

[0056] 3) Summarize the K verification results and calculate the average value as the final verification result;

[0057] (3) Compare the verification results of all hyperparameter combinations and determine the optimal hyperparameter combination.

[0058] Furthermore, the training of the deep neural network in step 55 continuously adjusts and optimizes the weights and biases {W, b} of the deep neural network through the back propagation algorithm to minimize the loss function:

[0059]

[0060] In the formula, is the loss function, usually the root mean square error between the predicted and measured data.

[0061] The beneficial effects of the present invention are as follows: the present invention provides a spatial distribution prediction method for soil respiration temperature sensitivity, which constructs a method for calculating soil respiration temperature sensitivity Q from different types of soil respiration data. 10 In particular, the method of calculating the soil respiration temperature sensitivity Q from soil respiration carbon emissions 10 , greatly enriching the availability of existing data and significantly improving the ability to extract effective information from soil respiration data. 10 Based on the data availability, a deep neural network model was used to construct the soil respiration temperature sensitivity Q 10 With Q 10 The nonlinear relationship between the influencing factors and the successful point-scale Q 10 Expanded to the regional scale, the spatial distribution prediction of soil respiration temperature sensitivity was achieved.

[0062] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a flow chart of the method of the present invention;

[0064] Figure 2 A schematic diagram of deep neural network model training and prediction in Example 1;

[0065] Figure 3 The soil respiration temperature sensitivity Q of the Qinghai-Tibet Plateau predicted in Example 1 10 Spatial distribution map. DETAILED DESCRIPTION

[0066] The present invention discloses a method for predicting the spatial distribution of soil respiration temperature sensitivity. Figure 1 As shown, the method comprises the following steps:

[0067] Step 1: Data acquisition and processing:

[0068] The soil respiration monitoring data and daily temperature data set of the study area were obtained. The soil respiration monitoring data include two types of data: the first type of data is the daily average flux data of soil respiration, and the second type of data is the carbon emission data of soil respiration, which are generally obtained through literature, public databases or site monitoring. The unit of the daily average flux data of soil respiration is μmolCm -2 s -1 ; Soil respiration carbon emission data refers to the cumulative value of soil respiration carbon emission during the monitoring period, in gCm -2 The daily temperature dataset comes from a public raster dataset, which can be remote sensing data or reanalysis data, such as the China Regional Meteorological Element Driven Dataset (CMFD) dataset.

[0069] Then, according to the location and time information of the acquired soil respiration monitoring data, the daily average temperature data is extracted from the daily temperature data set. For the first type of data, the daily average temperature corresponding to the daily average flux is extracted, and for the second type of data, the daily average temperature of each day during the monitoring period is extracted. This is because the acquired soil respiration monitoring data lacks temperature information, so it is necessary to extract the corresponding temperature information from the daily temperature data set according to the location and time information of the soil respiration monitoring data.

[0070] Step 2: Calculate the parameters of the Van't Hoff exponential chemical reaction-temperature equation:

[0071] The Van't Hoff exponential chemical reaction-temperature equation is used to describe the relationship between the daily average flux of soil respiration and the daily average temperature:

[0072] Rs=ae bT (1)

[0073] Where Rs is the daily average flux of soil respiration, in μmolCm -2 s -1 ; T is the daily average temperature, unit is ℃; a is the first empirical parameter, b is the second empirical parameter.

[0074] For the first type of data described in step 1, the Van't Hoff exponential chemical reaction-temperature equation, that is, formula (1), is fitted using the least squares method to determine the first empirical parameter a and the second empirical parameter b of the equation.

[0075] Among them, the least squares method is a commonly used method to solve regression problems. It finds the best matching function for fitting data by minimizing the sum of squares of errors. Its loss function is as follows:

[0076]

[0077] Where L is the loss function of the least squares method, that is, the sum of squared errors; (x i ,y i ) is the i-th sample; f() is the theoretical function of the unknown parameters.

[0078] For the second type of data described in step 1, the first empirical parameter a and the second empirical parameter b are solved by the following method.

[0079] According to the Van't Hoff exponential chemical reaction-temperature equation, the daily average flux of soil respiration during the monitoring period is expressed as:

[0080]

[0081] In the formula, Rs i represents the daily average flux of soil respiration on the i-th day; T i represents the average daily temperature on the i-th day.

[0082] The daily soil respiration emission was calculated based on the daily average flux of soil respiration, and the formula is:

[0083]

[0084] In the formula, is the daily soil respiration emission on the ith day, in gCm -2 ; m represents the mass of carbon per micromole, which is 12×10 -6 g μmol -1 ; t represents the number of seconds in a day, which is 86400s; η is the conversion factor, which is 1.0368, and the unit is gμmol -1 s.

[0085] The cumulative value of soil respiration carbon emissions during the monitoring period is calculated based on the daily emissions of soil respiration. The formula is:

[0086]

[0087] In the formula, P represents the number of days of monitoring time, the unit is d; is the cumulative value of soil respiration carbon emission within P days, in gCm -2 .

[0088] When there are two soil respiration carbon emission data at different monitoring times at the same location, the values ​​of parameters a and b are obtained by solving the following nonlinear equations:

[0089]

[0090] Where m and n represent different monitoring days, and the unit is d.

[0091] Step 3: Calculate soil respiration temperature sensitivity Q 10 :

[0092] According to the parameters of the Van't Hoff exponential chemical reaction-temperature equation obtained in step 2, the soil respiration temperature sensitivity Q 10 , the calculation formula is:

[0093] Q 10 =e 10b (7)

[0094] In the formula, Q 10 It is the multiple of soil respiration change when the temperature rises by 10℃, which is used to indicate the temperature sensitivity of soil respiration; b is the second empirical parameter.

[0095] Step 4: Obtain soil respiration temperature sensitivity Q 10 Impact factor data and processing:

[0096] Get Q in the study area 10 The influencing factor data include annual average temperature, annual precipitation, Normalized Vegetation Index (NDVI), pH (potential of hydrogen), soil organic carbon density (SOCD), Digital Elevation Model (DEM), soil properties, etc. 10 The impact factor data also comes from public raster datasets of various types.

[0097] To obtain the Q 10 The impact factor data of Q 10 The impact factor data is resampled to the required resolution, and then outliers are removed and missing values ​​are interpolated.

[0098] Step 5: Build and train a deep neural network model:

[0099] The specific steps include:

[0100] Step 51: Q calculated in step 3 10is the label, with the calculated Q 10 Q of location 10 The influencing factors are used as features to form the training data set.

[0101] Step 52: Q in the training data set 10 and Q 10 The impact factor data is normalized to generate a normalized training data set.

[0102] Step 53: Build a deep neural network model and add L1 or L2 regularization to prevent the deep neural network model from overfitting.

[0103] The specific process of building a deep neural network is as follows: Set the number of nodes in the input layer of the deep neural network (with Q 10 The number of influencing factors is related to the number of hidden layers, the number of neurons in the hidden layers, the output layer, the activation function, the loss function, the optimizer, the learning rate, the number of training times, the L1 or L2 regularization parameters, etc.

[0104] Among them, L1 and L2 regularization limit the complexity of the model to prevent the model from overfitting on the training data, thereby improving the generalization ability of the model.

[0105] L1 Regularization:

[0106]

[0107] L2 Regularization:

[0108]

[0109] In the formula, and They are the loss functions of L1 and L2 regularization respectively; is the data loss of the model; λ is the regularization parameter; W is the weight matrix; || ||1 and || ||2 are the L1 and L2 norms respectively.

[0110] Step 54: Use K-fold cross validation combined with grid search to determine the optimal hyperparameter combination. The specific process is as follows:

[0111] (1) Identify the hyperparameters that need to be adjusted and their possible values, and create a hyperparameter grid (i.e., possible combinations of hyperparameter values);

[0112] (2) Perform grid search and K-fold cross validation on all possible hyperparameter combinations;

[0113] The specific steps of the K-fold cross validation are:

[0114] 1) Split the entire data set into K subsets;

[0115] 2) Select one subset for validation without duplication, and use the remaining K-1 subsets for training the model;

[0116] 3) Summarize the K verification results and calculate the average value as the final verification result;

[0117] (3) Compare the verification results of all hyperparameter combinations and determine the optimal hyperparameter combination.

[0118] Step 55: Use the normalized training data set to train a deep neural network model to establish Q 10 With Q 10 The mapping relationship between the influencing factors is used to obtain the predicted Q 10 The regression model is the trained deep neural network model.

[0119] Among them, the training of the deep neural network continuously adjusts and optimizes the weights and biases {W, b} of the deep neural network through the back propagation algorithm to achieve the purpose of minimizing the loss function:

[0120]

[0121] In the formula, is the loss function, usually the root mean square error between the predicted and measured data.

[0122] Step 6: Determine the temperature sensitivity of soil respiration in the study area Q 10 Spatial distribution of:

[0123] The Q of all pixels at the required resolution in the study area 10 The impact factor data of the study area is used as the prediction data set, and the prediction data set is normalized to generate a normalized prediction data set; the generated normalized prediction data set is input into the trained deep neural network model to obtain the normalized Q 10 value; the normalized Q 10 The value was inversely normalized to obtain the soil respiration temperature sensitivity Q in the entire study area. 10 The value is used to determine the temperature sensitivity of soil respiration in the study area, Q 10 spatial distribution.

[0124] Embodiment 1

[0125] This embodiment is an application example of the above method.

[0126] This example uses the Qinghai-Tibet Plateau as the application area and 2000-2020 as the reference period. The above method is used to perform spatial mapping of the sensitivity of soil respiration in the Qinghai-Tibet Plateau with a spatial resolution of 1km×1km. The specific steps include:

[0127] Step 1: Data acquisition and processing:

[0128] The soil respiration monitoring data and daily temperature data set of the study area were obtained. The soil respiration monitoring data includes two types of data: the first type of data is the daily average flux data of soil respiration, and the second type of data is the carbon emission data of soil respiration. The unit of the daily average flux data of soil respiration is μmolCm -2 s -1 ; Soil respiration carbon emission data refers to the cumulative value of soil respiration carbon emission during the monitoring period, in gCm -2 The daily temperature dataset comes from a public raster dataset.

[0129] Then, according to the location and time information of the acquired soil respiration monitoring data, the daily average temperature data is extracted from the daily temperature data set. For the first type of data, the daily average temperature corresponding to the daily average flux is extracted, and for the second type of data, the daily average temperature of each day during the monitoring period is extracted.

[0130] Step 2: Calculate the parameters of the Van't Hoff exponential chemical reaction-temperature equation:

[0131] The Van't Hoff exponential chemical reaction-temperature equation is used to describe the relationship between the daily average flux of soil respiration and the daily average temperature:

[0132] Rs=ae bT (1)

[0133] Where Rs is the daily average flux of soil respiration, in μmolCm -2 s -1 ; T is the daily average temperature, unit is ℃; a is the first empirical parameter, b is the second empirical parameter.

[0134] For the first type of data described in step 1, the Van't Hoff exponential chemical reaction-temperature equation, that is, formula (1), is fitted using the least squares method to determine the first empirical parameter a and the second empirical parameter b of the equation.

[0135] Among them, the least squares method is a commonly used method to solve regression problems. It finds the best matching function for fitting data by minimizing the sum of squares of errors. Its loss function is as follows:

[0136]

[0137] Where L is the loss function of the least squares method, that is, the sum of squared errors; (x i ,y i ) is the i-th sample; f() is the theoretical function of the unknown parameters.

[0138] For the second type of data described in step 1, the first empirical parameter a and the second empirical parameter b are solved by the following method.

[0139] According to the Van't Hoff exponential chemical reaction-temperature equation, the daily average flux of soil respiration during the monitoring period is expressed as:

[0140]

[0141] In the formula, Rs i represents the daily average flux of soil respiration on the i-th day; T i represents the average daily temperature on the i-th day.

[0142] The daily soil respiration emission was calculated based on the daily average flux of soil respiration, and the formula is:

[0143]

[0144] In the formula, is the daily soil respiration emission on the ith day, in gCm -2 ; m represents the mass of carbon per micromole, which is 12×10 -6 g μmol -1 ; t represents the number of seconds in a day, which is 86400s; η is the conversion factor, which is 1.0368, and the unit is gμmol -1 s.

[0145] The cumulative value of soil respiration carbon emissions during the monitoring period is calculated based on the daily soil respiration emissions. The formula is:

[0146]

[0147] In the formula, P represents the number of days of monitoring, in d; is the cumulative value of soil respiration carbon emission within P days, in gCm -2 .

[0148] When there are two soil respiration carbon emission data at different monitoring times at the same location, the values ​​of parameters a and b are obtained by solving the following nonlinear equations:

[0149]

[0150] Where m and n represent different monitoring days, and the unit is d.

[0151] Step 3: Calculate soil respiration temperature sensitivity Q 10 :

[0152] According to the parameters of the Van't Hoff exponential chemical reaction-temperature equation obtained in step 2, the soil respiration temperature sensitivity Q10 , the calculation formula is:

[0153] Q 10 =e 10b (7)

[0154] In the formula, Q 10 It is the multiple of soil respiration change when the temperature rises by 10℃, which is used to indicate the temperature sensitivity of soil respiration; b is the second empirical parameter.

[0155] Step 4: Obtain soil respiration temperature sensitivity Q 10 The impact factor data of

[0156] Get Q in the study area 10 The impact factor data of Q obtained in this embodiment 10 The influencing factors include annual average temperature, annual precipitation, NDVI, pH, SOCD, DEM, soil bulk density, clay content, sand content, and silt content.

[0157] The Q 10 The impact factor data were resampled to a spatial resolution of 1km×1km, and then outliers were removed and missing values ​​were interpolated.

[0158] Step 5: Build and train a deep neural network model:

[0159] The specific steps include:

[0160] Step 51: Q calculated in step 3 10 is the label, with the calculated Q 10 Q of location 10 The influencing factors are used as features to form the training data set.

[0161] Step 52: Q in the training data set 10 and Q 10 The impact factor data is normalized to generate a normalized training data set.

[0162] Step 53: Build a deep neural network model and add L1 or L2 regularization to prevent the deep neural network model from overfitting.

[0163] The specific process of building a deep neural network is as follows: Set the number of nodes in the input layer of the deep neural network (with Q 10 The number of influencing factors is related to the number of hidden layers, the number of neurons in the hidden layers, the output layer, the activation function, the loss function, the optimizer, the learning rate, the number of training times, the L1 or L2 regularization parameters, etc.

[0164] Among them, L1 and L2 regularization limit the complexity of the model to prevent the model from overfitting on the training data, thereby improving the generalization ability of the model.

[0165] L1 Regularization:

[0166]

[0167] L2 Regularization:

[0168]

[0169] In the formula, and They are the loss functions of L1 and L2 regularization respectively; is the data loss of the model; λ is the regularization parameter; W is the weight matrix; || ||1 and || ||2 are the L1 and L2 norms respectively.

[0170] In this embodiment, the number of neurons in the input layer is set to 10 (Q 10 The number of influencing factors is 10), the number of neurons in the output layer is 1, the activation function is relu, the optimizer is Adam, and the loss function is MSE. In order to obtain the deep neural network model with the optimal structure, the number of hidden layers, the number of neurons in the hidden layers, the learning rate, and the regularization parameter are determined by K-fold cross validation in step 54.

[0171] Step 54: Use K-fold cross validation combined with grid search to determine the optimal hyperparameter combination. The specific process is as follows:

[0172] (1) Identify the hyperparameters that need to be adjusted and their possible values, and create a hyperparameter grid (i.e., possible value combinations of hyperparameters). In this embodiment, the optimal hidden layer structure (number of layers and neurons) and the learning rate search grid are set as follows:

[0173] Hidden layer structure: [10, 10], [20, 20], [10, 10, 10], [20, 20, 20], [10, 20, 10]. [10, 10] means there are 2 hidden layers and 10 neurons in each layer.

[0174] Learning rate: 0.01, 0.001, 0.0001, 0.00001, 0.000001.

[0175] (2) Perform grid search and K-fold cross validation on all possible hyperparameter combinations;

[0176] The specific steps of the K-fold cross validation are:

[0177] 1) Split the entire data set into K subsets; in this embodiment, the number of folds of K-fold cross validation is set to 10 (ie, K=10).

[0178] 2) One subset is selected for verification without repetition, and the remaining K-1 subsets are used for training the model; in this embodiment, one subset is selected for verification without repetition, and the remaining 9 subsets are used for training the model, and a total of 10 verifications are performed.

[0179] 3) Summarize K verification results and calculate the average value as the final verification result; this embodiment summarizes the results of 10 verifications and calculates the tie value as the final verification result.

[0180] (3) Compare the verification results of all hyperparameter combinations to determine the optimal hyperparameter combination. The optimal parameters in this embodiment are: {hidden layer structure: [10, 10, 10]; learning rate: 0.0001}. This embodiment does not limit the type of optimal parameters selected by cross-validation.

[0181] Step 55: Use the normalized training data set to train a deep neural network model to establish a temperature sensitivity Q 10 With Q 10 The mapping relationship between the influencing factors is used to obtain the predicted Q 10 The regression model is the trained deep neural network model, such as Figure 2 shown.

[0182] Among them, the training of the deep neural network continuously adjusts and optimizes the weights and biases {W, b} of the deep neural network through the back propagation algorithm to achieve the purpose of minimizing the loss function:

[0183]

[0184] In the formula, is the loss function, usually the root mean square error between the predicted and measured data.

[0185] Step 6: Determine the temperature sensitivity of soil respiration in the study area Q 10 Spatial distribution of:

[0186] The Q of all pixels at the required resolution in the study area 10 The impact factor data of the Qinghai-Tibet Plateau is used as the prediction data set. 10 The impact factor data (2000-2020) predicts the temperature sensitivity of soil respiration Q in the Qinghai-Tibet Plateau during the period 2000-2020 10 The spatial distribution of the prediction data set is Q within all 1km×1km pixels. 10 Impact factor data.

[0187] The prediction data set is normalized to generate a normalized prediction data set; the normalized prediction data set of the study area is input into the trained deep neural network model, such as Figure 2 As shown, the normalized Q corresponding to the resolution of the study area is obtained 10 value; the normalized Q 10 The value was inversely normalized to obtain the soil respiration temperature sensitivity Q in the entire study area. 10 The value is used to determine the temperature sensitivity of soil respiration in the study area, Q 10 spatial distribution.

[0188] The temperature sensitivity of soil respiration Q in the Qinghai-Tibet Plateau predicted in this example during 2000-2020 10 Spatial distribution Figure 3 As shown. Figure 3 It can be seen that the temperature sensitivity of soil respiration in the Qinghai-Tibet Plateau is Q 10 The value is between 1 and 10, showing a distribution characteristic of smaller values ​​in the northwest and larger values ​​in the southeast. The deep neural network well presents the soil respiration sensitivity Q of the Qinghai-Tibet Plateau 10 spatial distribution.

[0189] The present invention makes full use of the limited soil respiration monitoring data (including flux data and carbon emission data) and mines effective information, significantly improving the availability of data in the research area. At the same time, the present invention combines deep learning technology to convert point-scale Q 10 The value is extended to the regional scale, improving Q 10 The accuracy of spatial distribution prediction. The present invention is applicable to the soil respiration temperature sensitivity Q in any region 10 The spatial distribution prediction of soil respiration is more advantageous in areas lacking continuous soil respiration monitoring (such as the Qinghai-Tibet Plateau).

[0190] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred arrangement scheme, a person skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting the spatial distribution of soil respiration temperature sensitivity, characterized in that: The method comprises the following steps: Step 1, data acquisition and processing: obtain soil respiration monitoring data and daily temperature data set in the study area. The soil respiration monitoring data includes two types of data. The first type of data is the daily average flux data of soil respiration, and the second type of data is the carbon emission data of soil respiration. According to the location and time information of the acquired soil respiration monitoring data, extract the daily average temperature data from the daily temperature data set; Step 2: Calculate the parameters of the Van't Hoff exponential chemical reaction-temperature equation: The Van't Hoff exponential chemical reaction-temperature equation is used to describe the relationship between the daily average flux of soil respiration and the daily average temperature: Rs=ae bT (1) Where Rs is the daily average flux of soil respiration, in μmolCm -2 s -1 ; T is the daily average temperature, in °C; a is the first empirical parameter, b is the second empirical parameter; For the first type of data described in step 1, the Van't Hoff exponential chemical reaction-temperature equation, i.e., formula (1), is fitted using the least square method to determine the first empirical parameter a and the second empirical parameter b of the equation; Among them, the least squares method finds the best matching function for fitting data by minimizing the sum of squares of errors, and its loss function is as follows: Where L is the loss function of the least squares method, i.e., the sum of squared errors; (x i ,y i ) is the i-th sample; f() is the theoretical function of the unknown parameters; For the second type of data described in step 1, the first empirical parameter a and the second empirical parameter b are solved by the following method; According to the Van't Hoff exponential chemical reaction-temperature equation, the daily average flux of soil respiration during the monitoring period is expressed as: In the formula, Rs i represents the daily average flux of soil respiration on the i-th day; T i represents the daily average temperature on the i-th day; The daily soil respiration emission was calculated based on the daily average flux of soil respiration, and the formula is: In the formula, is the daily soil respiration emission on the ith day, in gCm -2 ; m represents the mass of carbon per micromole, which is 12×10 -6 g μmol -1 ; t represents the number of seconds in a day, which is 86400s; η is the conversion factor, which is 1.0368, and the unit is gμmol -1 s; The cumulative value of soil respiration carbon emissions during the monitoring period is calculated based on the daily emissions of soil respiration. The formula is: In the formula, P represents the number of days of monitoring time, the unit is d; is the cumulative value of soil respiration carbon emission within P days, in gCm -2 ; When there are two soil respiration carbon emission data at different monitoring times at the same location, the values ​​of parameters a and b are obtained by solving the following nonlinear equations: In the formula, m and n represent different monitoring days, and the unit is d; Step 3: Calculate soil respiration temperature sensitivity Q 10 : Based on the parameters of the Van't Hoff exponential chemical reaction-temperature equation obtained in step 2, calculate the soil respiration temperature sensitivity Q 10 , the calculation formula is: Q 10 =e 10b (7) In the formula, Q 10 is the multiple of soil respiration change when the temperature rises by 10℃, which is used to indicate the temperature sensitivity of soil respiration; b is the second empirical parameter; Step 4: Obtain soil respiration temperature sensitivity Q 10 The impact factor data of the study area is obtained and processed: 10 The impact factor data of the obtained Q 10 The impact factor data of Q 10 The impact factor data are resampled to the required resolution, then outliers are removed and missing values ​​are interpolated; Step 5: Build and train a deep neural network model: This includes the following steps: Step 51: Q calculated in step 3 10 is the label, with the calculated Q 10 Q of location 10 The influencing factors are used as features to form a training data set; Step 52: Q in the training data set 10 and Q 10 Normalize the impact factor data to generate a normalized training data set; Step 53: Build a deep neural network model and add L1 or L2 regularization to prevent the deep neural network model from overfitting; Step 54: Use K-fold cross validation combined with grid search to determine the optimal hyperparameter combination; Step 55: Use the normalized training data set to train a deep neural network model to establish Q 10 With Q 10 The mapping relationship between the influencing factors is used to obtain the predicted Q 10 The regression model is the trained deep neural network model; Step 6: Determine the temperature sensitivity of soil respiration in the study area Q 10 Spatial distribution of: Q of all pixels at the required resolution in the study area 10 The impact factor data of the study area is used as the prediction data set, and the prediction data set is normalized to generate a normalized prediction data set; the generated normalized prediction data set is input into the trained deep neural network model to obtain the normalized Q 10 value; the normalized Q 10 The value was inversely normalized to obtain the soil respiration temperature sensitivity Q in the entire study area. 10 The value is used to determine the temperature sensitivity of soil respiration in the study area, Q 10 spatial distribution.

2. The method for predicting the spatial distribution of soil respiration temperature sensitivity according to claim 1, characterized in that: The soil respiration monitoring data in step 1 is obtained through literature, public databases or site monitoring; the daily temperature dataset comes from a public raster dataset, and the raster dataset is remote sensing data or reanalysis data.

3. The method for predicting the spatial distribution of soil respiration temperature sensitivity according to claim 1, characterized in that: Q as described in step 4 10 The influencing factor data include annual average temperature, annual precipitation, normalized difference vegetation index NDVI, pH, soil organic carbon density SOCD, elevation DEM, and soil property data.

4. The method for predicting the spatial distribution of soil respiration temperature sensitivity according to claim 3, characterized in that: The soil property data include soil bulk density, clay content, sand content, and silt content.

5. The method for predicting the spatial distribution of soil respiration temperature sensitivity according to claim 1, characterized in that: The required resolution described in step 4 is 1 km × 1 km.

6. The method for predicting the spatial distribution of soil respiration temperature sensitivity according to claim 1, characterized in that: The specific process of building the deep neural network in step 53 is: setting the number of nodes in the input layer, the number of hidden layers, the number of neurons in the hidden layer, the output layer, the activation function, the loss function, the optimizer, the learning rate, the number of training times, and the L1 or L2 regularization parameters of the deep neural network; the number of nodes in the input layer and Q 10 The number of impact factors is related to Among them, L1 and L2 regularization limits the complexity of the model to prevent the model from overfitting on the training data, thereby improving the generalization ability of the model; L1 Regularization: L2 Regularization: In the formula, and They are the loss functions of L1 and L2 regularization respectively; is the data loss of the model; λ is the regularization parameter; W is the weight matrix; 1 and 2 are the L1 and L2 norms respectively.

7. The method for predicting the spatial distribution of soil respiration temperature sensitivity according to claim 1, characterized in that: The specific process of using K-fold cross validation combined with grid search to determine the optimal hyperparameter combination in step 54 is: (1) Identify the hyperparameters that need to be adjusted and their possible values, and create a hyperparameter grid, that is, possible combinations of hyperparameter values; (2) Perform grid search and K-fold cross validation on all possible hyperparameter combinations; The specific steps of the K-fold cross validation are: 1) Split the entire data set into K subsets; 2) Select one subset for validation without duplication, and use the remaining K-1 subsets for training the model; 3) Summarize the K verification results and calculate the average value as the final verification result; (3) Compare the verification results of all hyperparameter combinations and determine the optimal hyperparameter combination.

8. The method for predicting the spatial distribution of soil respiration temperature sensitivity according to claim 1, characterized in that: The training of the deep neural network described in step 55 continuously adjusts and optimizes the weights and biases {W, b} of the deep neural network through the back propagation algorithm to minimize the loss function: In the formula, is the loss function, usually the root mean square error between the predicted and measured data.

Images