Machine learning-based surface temperature estimation method

Abstract

The invention discloses a machine learning-based surface temperature estimation method, which comprises the following steps of: collecting near-surface meteorological observation data, and establishing a meteorological data database; integrating the meteorological data of the meteorological data database to obtain a sample required by modeling; carrying out layer-by-layer progressive three-level situation division on the sample, and respectively establishing a training database of each situation; in combination with the training database of each situation, optimal independent variable combinations and algorithms of the model are optimized for different situations respectively; the optimal independent variable combination of each situation and an algorithm are adopted to establish a model of each situation, the models are tested one by one through cross validation, the optimal parameters of the models under each situation are determined, and model training is completed; and inputting the to-be-estimated independent variable into the estimation model of the corresponding situation, and carrying out land surface temperature estimation, the land surface temperature estimation problem is researched, situation division is carried out, the land surface temperature estimation model of each situation is established, and the land surface temperature estimation precision and the model application universality are effectively improved.

Description

Technical Field

[0001] The present application relates to the field of surface temperature estimation and machine learning, and in particular to a method for estimating surface temperature based on machine learning. Background Art

[0002] As a physical property of soil, surface temperature has an important impact on plant growth, development and other physiological processes, and is one of the important bases for agricultural disaster loss assessment and production decision-making. Due to the limited number of sites for surface temperature observation and a short observation history, surface temperature observation data is far from meeting application requirements. In agricultural production, near-surface air temperature is often used instead of surface temperature for research, but surface temperature is affected by factors such as vegetation cover, radiation, and soil moisture. Using near-surface air temperature instead of surface temperature has a large error. Therefore, estimating surface temperature has an important application demand.

[0003] Most current surface temperature estimation methods are based on near-surface meteorological observation data using linear regression. However, the relationship between near-surface meteorological elements and surface temperature is very complex and not linear. Such methods can only fit a single site or a certain period of time, and have time and geographical limitations, and the estimation errors are generally large. Currently, there is a lack of surface temperature estimation technology solutions with a wide range of applications and good results. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a surface temperature estimation method based on machine learning, which solves the problem that the existing surface temperature estimation method only fits a single site or a certain time period, has time and geographical limitations, and the estimation error is generally large.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a method for estimating ground surface temperature based on machine learning, comprising: S1. Collect near-surface meteorological observation data and establish a meteorological data database; S2, integrating meteorological data in the meteorological data database to obtain samples required for modeling; S3, divide the samples required for modeling into three levels of situations step by step, and establish training databases for each situation; S4. Combine the training database of each scenario to select the optimal independent variable combination and algorithm of the model for different scenarios; S5. Use the optimal independent variable combination and algorithm for each scenario to establish the model for each scenario, test the models one by one through cross-validation, determine the optimal parameters of the model under each scenario, and complete the model training; S6. Input the independent variables to be estimated into the estimation model of the corresponding scenario to estimate the surface temperature.

[0006] Furthermore: the meteorological data database established in S1 includes the daily observation data of long-term near-surface meteorological elements of the meteorological stations in the target area, including: surface temperature, daily maximum temperature, daily minimum temperature, daily average temperature, sunshine hours and daily precipitation.

[0007] Further: In S2, the variables describing the sample characteristics include daily maximum temperature, daily minimum temperature, daily average temperature, sunshine hours, daily precipitation, weather station code, and day sequence; Weather station codes are used to reflect differences in the underlying surface; The daily sequence is used to reflect the differences in the relationship between surface temperature and near-surface meteorological conditions at different times of the year.

[0008] Further: In S3, three levels of scenarios are divided into progressive levels according to temperature, sunshine and soil moisture; The steps for determining the thresholds for dividing the temperature scenario and the sunshine scenario include: S31, taking the daily maximum temperature, daily minimum temperature, daily average temperature, sunshine hours, daily precipitation, weather station code and day sequence as characteristic variables, taking the daily surface temperature observation data as the target variable, referring to the clustering results of the decision tree algorithm, and determining the reference values ​​of all thresholds for dividing the temperature and sunshine scenarios; S32, based on the reference value of each threshold value, each unit change upward and downward is a preset value, the temperature changes up and down by 5 units, and the sunshine hours change up and down by 2 units, to form a set of preset values ​​for each threshold value A i ={ x 1 , x 2 ,..., x n}; in, i represents the threshold value, n is the number of preset values ​​of the threshold; S33, calculating the Cartesian product of the preset value sets of all thresholds to obtain all threshold configuration combinations; S34, traverse all scenarios divided by all threshold configuration combinations, and substitute them into the machine learning model for modeling; S35. Select the one with the lowest root mean square error of the machine learning model among all threshold configuration combinations as the optimal threshold configuration combination; S36, completing the division of temperature scenarios and sunshine scenarios according to the optimal threshold configuration combination; For the division of soil moisture scenarios, the sliding window method is used to divide the soil into drought and wet scenarios based on precipitation data.

[0009] Further: In S4, the method of selecting the optimal independent variable combination of the model for each scenario includes: S401, combining independent variables by exhaustive method to obtain different independent variable combinations; S402, taking the surface temperature as the target variable, traversing all independent variable combinations and substituting them into the machine learning model in combination with the training database of the scenario to perform a fitting experiment; S403. Select the independent variable combination with the lowest root mean square error in the fitting experiment results as the optimal independent variable combination.

[0010] Further: In S4, the method of selecting the optimal algorithm of the model for each scenario includes: S411, substitute the optimal independent variable combination into different machine learning algorithms to conduct fitting experiments; S412. Select the machine learning algorithm with the lowest root mean square error in the fitting experiment results as the optimal algorithm.

[0011] The beneficial effects of the present invention are: 1. The complex response mechanism between surface temperature and near-surface meteorological conditions is a key difficulty in estimating surface temperature. Compared with the widely used time series linear regression method, machine learning is better at processing complex nonlinear relationships and high-dimensional data, and can better reflect the response mechanism and improve the accuracy of surface temperature estimation; 2. The present invention incorporates the weather station code and date sequence into the independent variables input into the estimation model, so the constructed model can be better applied to different regions and different times, improving the universality of the model. In addition, in the actual application process, as the number of observation sites included in the database increases and the database expands, the estimation effect of the model will be further improved; 3. The invention adopts a scenario-based estimation method and sets up a surface temperature estimation model with 10 scenarios at three levels of "temperature-sunshine-soil moisture", which also improves the overall estimation accuracy and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 Flowchart of the surface temperature estimation method based on machine learning.

[0013] Figure 2 This is a distribution map of 20 meteorological observation sites in an embodiment of the present invention.

[0014] Figure 3 Schematic diagram of the scenario division results and sample size of each scenario database according to an embodiment of the present invention.

[0015] Figure 4 This is a standardized residual histogram of the surface temperature estimation result according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0017] like Figure 1 As shown, in one embodiment of the present invention, a method for estimating ground surface temperature based on machine learning is provided, comprising: S1. Collect near-surface meteorological observation data and establish a meteorological data database; S2, integrating meteorological data in the meteorological data database to obtain samples required for modeling; S3, divide the samples required for modeling into three levels of situations step by step, and establish training databases for each situation; S4. Combine the training database of each scenario to select the optimal independent variable combination and algorithm of the model for different scenarios; S5. Use the optimal independent variable combination and algorithm for each scenario to establish the model for each scenario, test the models one by one through cross-validation, determine the optimal parameters of the model under each scenario, and complete the model training; S6. Input the independent variables to be estimated into the estimation model of the corresponding scenario to estimate the surface temperature.

[0018] The meteorological data database established in S1 includes the daily observation data of long-term near-surface meteorological elements of the meteorological stations in the target area, including: surface temperature, daily maximum temperature, daily minimum temperature, daily average temperature, sunshine hours and daily precipitation.

[0019] In S2, the variables describing the sample characteristics include daily maximum temperature, daily minimum temperature, daily average temperature, sunshine hours, daily precipitation, weather station code, and day sequence; Weather station codes are used to reflect differences in the underlying surface; The daily sequence is used to reflect the differences in the relationship between surface temperature and near-surface meteorological conditions at different times of the year.

[0020] In S3, three levels of scenarios are divided into progressive levels according to temperature, sunshine and soil moisture; The steps for determining the thresholds for dividing the temperature scenario and the sunshine scenario include: S31, taking the daily maximum temperature, daily minimum temperature, daily average temperature, sunshine hours, daily precipitation, weather station code and day sequence as characteristic variables, taking the daily surface temperature observation data as the target variable, referring to the clustering results of the decision tree algorithm, and determining the reference values ​​of all thresholds for dividing the temperature and sunshine scenarios; The specific process of creating a decision tree model is as follows: input the daily maximum temperature, daily minimum temperature, daily average temperature, sunshine hours, daily precipitation, weather station code and day sequence as the feature variables for decision tree modeling, and input the daily surface temperature observation data as the target variable for decision tree modeling. The decision tree algorithm creates a decision tree by recursively splitting the data set. Each internal node of the output decision tree represents a feature, each branch represents a decision rule, and each leaf node represents a category. The preset value of the threshold is determined by referring to the clustering result of the output decision tree.

[0021] S32, based on the reference value of each threshold value, each unit change upward and downward is a preset value, the temperature changes up and down by 5 units, and the sunshine hours change up and down by 2 units, to form a set of preset values ​​for each threshold value A i ={ x 1 , x 2 ,..., x n}; in, i represents the threshold value, n is the number of preset values ​​of the threshold; S33, calculating the Cartesian product of the preset value sets of all thresholds to obtain all threshold configuration combinations; S34, traverse all scenarios divided by all threshold configuration combinations, and substitute them into the machine learning model for modeling; S35. Select the machine learning model with the lowest root mean square error among all threshold configuration combinations to determine the optimal threshold configuration combination; S36, completing the division of temperature scenarios and sunshine scenarios according to the optimal threshold configuration combination; Among them, for the division of soil moisture scenarios, the sliding window method is adopted to divide the soil drought and wet scenarios according to the precipitation data. In this embodiment, specifically: within the 15-day sliding window, if the daily precipitation of a specific day and the previous 14 days is less than 1 mm, then the specific day is judged as a drought scenario; within the 5-day sliding window, if the daily precipitation of a specific day and any one of the previous 4 days is ≥40 mm, then the specific day is judged as a wet scenario.

[0022] In S4, the method of selecting the optimal independent variable combination of the model for different scenarios includes: S401, combining independent variables by exhaustive method to obtain different independent variable combinations; S402, taking the surface temperature as the target variable, traversing all independent variable combinations and substituting them into the machine learning model in combination with the training database of the scenario to perform a fitting experiment; S403. Select the independent variable combination with the lowest root mean square error in the fitting experiment results as the optimal independent variable combination.

[0023] In S4, the method of selecting the optimal algorithm of the model for different scenarios includes: S411, substitute the optimal independent variable combination into different machine learning algorithms to conduct fitting experiments; S412. Select the machine learning algorithm with the lowest root mean square error in the fitting experiment results as the optimal algorithm.

[0024] In one embodiment of the present invention, Jiangsu Province is taken as the target area, 20 meteorological stations are collected as cases, and the specific implementation steps are as follows: S1. Obtain the daily surface temperature (GST_Max), daily maximum temperature (TEM_Max), daily minimum temperature (TEM_Min), daily average temperature (TEM_Avg), sunshine hours (dh), and daily precipitation (Pre) of 20 meteorological stations in the target area from 1990 to 2022 from the meteorological department to form a meteorological data database; S2. Obtain the samples required for modeling. The variables describing the sample characteristics include daily maximum temperature (TEM_Max), daily minimum temperature (TEM_Min), daily average temperature (TEM_Avg), sunshine hours (dh), daily precipitation (Pre), weather station code (Station) and day sequence (DOY); S3. Divide the scenarios. First, the first level is divided according to the daily average temperature. Through the traversal method, the daily average temperature is divided into three first-level temperature scenarios: low temperature (TEM_Avg≤10℃), medium temperature (10℃<TEM_Avg<34℃) and high temperature (TEM_Avg≥34℃). The sample sizes of the three temperature scenario databases are 9742, 24380 and 27 respectively. Based on the temperature scenario, the secondary scenarios are divided according to the sunshine hours, which are divided into two secondary scenarios: low sunshine (dh≤3) and normal sunshine (dh>3); Among them, for the high temperature (TEM_Avg ≥ 34℃) scenario, there is rarely a lack of sunshine, so this temperature scenario is not divided into the secondary sunshine scenario; According to the classification standard of the dry and wet conditions of soil moisture in this scheme, the scenarios are further divided into the third level. Among them, the estimation models of the drought scenarios under the two sunshine scenarios under the low temperature (TEM_Avg≤10℃) scenario are very different. After merging the two drought scenarios, it does not affect the model fitting effect, so the scenarios are merged. However, there are many samples of normal sunshine scenarios in the medium temperature (10℃<TEM_Avg<34℃), and the model fitting effect is poor due to the large temperature range. Therefore, on the basis of the three-level scenario division, the traversal method is used again according to the temperature to divide the first-level scenario into low temperature (10℃<TEM_Avg≤18℃) and high temperature (18℃<TEM_Avg<34℃). Finally, there are very few samples of soil wet conditions in the high temperature (TEM_Avg≥34℃) scenario, so the second-level scenario division of soil dry and wet conditions is not carried out.

[0025] Finally, a total of 10 scenarios were obtained. The division steps and results are detailed in Figure 3 ,The database sample size of each scenario is also shown in the figure.

[0026] S4. Combine the training database of each scenario to select the optimal independent variable combination and algorithm of the model for different scenarios; S4-1. When selecting the independent variable combination for each scenario, the independent variables are combined by exhaustive method to obtain different independent variable combinations. All independent variable combinations are traversed with surface temperature as the target variable, and the same machine learning algorithm is substituted for fitting experiments. The fitting experimental results of each variable combination are sorted according to the root mean square error (RMSE), and the independent variable combination with the lowest root mean square error (RMSE) is selected to enter the training of the estimation model. Finally, the independent variables of the 1st to 9th categories of input are selected as follows: daily maximum temperature (TEM_Max), daily average temperature (TEM_Avg), sunshine hours (dh), daily precipitation (Pre), day sequence (DOY), and weather station code (Station), a total of 6 input variables. The characteristic variables of the 10th category of input are daily maximum temperature (TEM_Max), daily minimum temperature (TEM_Min), daily average temperature (TEM_Avg), sunshine hours (dh), daily precipitation (Pre), day sequence (DOY), and weather station code (Station), a total of 7 input variables.

[0027] S4-2. When optimizing the algorithm for each scenario, take the surface temperature as the target variable, use the three machine learning algorithms, XGBoosting, random forest, and support vector machine, to bring the optimal independent variable combination of the scenario generated in S4-1 into the fitting experiment, and sort the fitting experimental results of each machine learning algorithm according to the root mean square error (RMSE), and select the machine learning algorithm with the lowest root mean square error (RMSE) as the training algorithm for the estimation model.

[0028] For this example, the random forest algorithm performs best in the 3rd, 5th, and 10th categories, and the XGBoosting algorithm performs best in the remaining categories.

[0029] S5. Use the optimal independent variable combination and algorithm for each scenario to establish the model for each scenario, test the models one by one through cross-validation, determine the optimal parameters of the model under each scenario, and complete the model training; S5-1. When establishing each scenario model, a training set and a test set are established for each scenario database. Among them, 70% of the samples are selected as the training set and 30% of the samples are selected as the test set, which are used for training and testing the model respectively; S5-2. The specific process of cross-validation is as follows: call the Caret package of R software (R Version: 4.3.1) for cross-validation; use k-fold cross-validation, and set the k value to 5; for each category, use the train function to train the model, and use the specified parameter grid for tuning. Among them, the random forest model specifies the mtry value parameter for tuning, and the XGBoosting model specifies the nrounds, max_depth, eta, gamma, colsample_bytree, min_child_weight, and subsample parameter grids for tuning.

[0030] Here, the search range of mtry is specified as [1,2,3,4,5,6,7,8,9]; the search range of nrounds is [100,200,300]; the search range of max_depth is [3,5,6,10]; the search range of eta is [0.01,0.1,0.4]; the search range of gamma is [0.2,0.3,0.4,0.5]; the search range of colsample_bytree is [0.5,0.6,0.7,1.0]; the search range of min_child_weight is [1,5,7,10]; the search range of subsample is [0.5,0.7,0.8,1.0]. The tuning results of each model parameter are shown in Table 1 and Table 2: Table 1 Random forest algorithm model parameter tuning results category mtry 3 3 5 2 10 2 Table 2 XGBooting algorithm model parameter tuning results

[0031] The estimated accuracy of the model after training (root mean square error (RMSE), mean absolute error (MAE) and R-squared (R 2 )) See Table 3 for details, and the overall estimation effect can be seen Figure 4 .

[0032] Table 3 Evaluation index table of each category model and overall estimation accuracy

[0033] In summary, the present invention conducts research on the problem of surface temperature estimation, divides the scenarios, and uses machine learning algorithms to build surface temperature estimation models for each scenario, which effectively improves the accuracy of surface temperature estimation and the universality of model application.

[0034] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for estimating land surface temperature based on machine learning, characterized in that: include: S1. Collect near-surface meteorological observation data and establish a meteorological data database; S2, integrating meteorological data in the meteorological data database to obtain samples required for modeling; S3, divide the samples required for modeling into three levels of situations step by step, and establish training databases for each situation; S4. Combine the training database of each scenario to select the optimal independent variable combination and algorithm of the model for different scenarios; S5. Use the optimal independent variable combination and algorithm for each scenario to establish the model for each scenario, test the models one by one through cross-validation, determine the optimal parameters of the model under each scenario, and complete the model training; S6. Input the independent variables to be estimated into the estimation model of the corresponding scenario to estimate the surface temperature.

2. The method for estimating ground surface temperature based on machine learning according to claim 1, characterized in that: The meteorological data database established in S1 includes the daily observation data of long-term near-surface meteorological elements of the meteorological stations in the target area, including: surface temperature, daily maximum temperature, daily minimum temperature, daily average temperature, sunshine hours and daily precipitation.

3. The method for estimating ground surface temperature based on machine learning according to claim 1, characterized in that: In S2, the variables describing the sample characteristics include daily maximum temperature, daily minimum temperature, daily average temperature, sunshine hours, daily precipitation, weather station code, and day sequence; Weather station codes are used to reflect differences in the underlying surface; The daily sequence is used to reflect the differences in the relationship between surface temperature and near-surface meteorological conditions at different times of the year.

4. The method for estimating ground surface temperature based on machine learning according to claim 1, characterized in that: In S3, three levels of scenarios are divided into progressive levels according to temperature, sunshine and soil moisture; The steps for determining the thresholds for dividing the temperature scenario and the sunshine scenario include: S31, taking the daily maximum temperature, daily minimum temperature, daily average temperature, sunshine hours, daily precipitation, weather station code and day sequence as characteristic variables, taking the daily surface temperature observation data as the target variable, referring to the clustering results of the decision tree algorithm, and determining the reference values ​​of all thresholds for dividing the temperature and sunshine scenarios; S32, based on the reference value of each threshold value, each unit change upward and downward is a preset value, the temperature changes up and down by 5 units, and the sunshine hours change up and down by 2 units, to form a set of preset values ​​for each threshold value A i ={ x 1, x 2,..., x n }; in, i represents the threshold value, n is the number of preset values ​​of the threshold; S33, calculating the Cartesian product of the preset value sets of all thresholds to obtain all threshold configuration combinations; S34, traverse all scenarios divided by all threshold configuration combinations, and substitute them into the machine learning model for modeling; S35. Select the one with the lowest root mean square error of the machine learning model among all threshold configuration combinations as the optimal threshold configuration combination; S36, completing the division of temperature scenarios and sunshine scenarios according to the optimal threshold configuration combination; For the division of soil moisture scenarios, the sliding window method is used to divide the soil into drought and wet scenarios based on precipitation data.

5. The method for estimating ground surface temperature based on machine learning according to claim 1, characterized in that: In S4, the method of selecting the optimal combination of independent variables of the model for each scenario includes: S401, combining independent variables by exhaustive method to obtain different independent variable combinations; S402, taking the surface temperature as the target variable, traversing all independent variable combinations and substituting them into the machine learning model in combination with the training database of the scenario to perform a fitting experiment; S403. Select the independent variable combination with the lowest root mean square error in the fitting experiment results as the optimal independent variable combination.

6. The method for estimating ground surface temperature based on machine learning according to claim 1, characterized in that: In S4, the method of selecting the optimal algorithm of the model for each scenario includes: S411, substitute the optimal independent variable combination into different machine learning algorithms to conduct fitting experiments; S412. Select the machine learning algorithm with the lowest root mean square error in the fitting experiment results as the optimal algorithm.

Images