CN114564767A

CN114564767A - Under-cloud surface temperature estimation method based on sun-cloud-satellite observation geometry

Info

Publication number: CN114564767A
Application number: CN202210204091.5A
Authority: CN
Inventors: 独文惠; 覃志豪; 范锦龙; 鲍宽乐; 赵春亮; 李仕峰
Original assignee: Institute of Agricultural Resources and Regional Planning of CAAS
Current assignee: Institute of Agricultural Resources and Regional Planning of CAAS
Priority date: 2022-03-03
Filing date: 2022-03-03
Publication date: 2022-05-31

Abstract

The invention belongs to the technical field of remote sensing application, and discloses an underground cloud surface temperature estimation method considering sun-cloud-satellite observation geometry, in particular to an underground cloud surface temperature estimation algorithm based on an integrated learning model, wherein the estimation of the underground cloud surface temperature mainly comprises the steps of dividing an underground cloud pixel into a cloud shielding pixel and a cloud covering pixel according to the geometrical observation structure of the sun-earth surface-satellite in consideration of an FY-3D MERSI-II image, and estimating the surface temperature by using the integrated learning model according to a temperature cause mechanism with different pixel numbers. In the estimation process, an MODTRAN radiation transmission model is combined, multiple earth surface temperature driving factors related to solar radiation, atmosphere, earth surface and precipitation meteorological conditions and the like are integrated, a set of flow method for achieving automatic operation of the earth surface temperature in the cloud is established, automatic all-weather earth surface temperature image estimation of the FY-3D MERSI-II image is achieved, and the applicability of the FY-3D MERSI-II high-resolution earth surface temperature data is improved.

Description

Under-cloud surface temperature estimation method based on sun-cloud-satellite observation geometry

Technical Field

The invention belongs to the technical field of remote sensing application, and particularly relates to a cloud-to-ground surface temperature estimation method based on sun-cloud-satellite observation geometry.

Background

Surface Temperature (LST) is a comprehensive characterization of energy exchange results of surface-atmosphere interaction on regional and global scales, and is widely applied to the fields of surface energy balance, drought monitoring, farmland soil moisture management, crop yield estimation, numerical weather forecast, climate change and the like. The remote sensing observation has the characteristics of continuous space, short period and low cost, so that the method is an effective technical means for acquiring regional or global-scale LST space-time distribution information. In the actual invention and application, due to the convenience of measurement and the mature LST inversion algorithm in clear sky, the surface temperature data is still mainly obtained through Thermal Infrared (TIR) remote sensing observation data. At present, surface temperature products such as MODIS, GLASS, FY3C/VIRR (visible light infrared scanning radiometer) and the like based on thermal infrared remote sensing are widely applied to change detection, drought monitoring and climate change. However, TIR LSTs are not usable in the presence of clouds, the TIR signal cannot penetrate the cloud cover to the surface, and the thermal radiation of the cloud pixels observed by the thermal infrared sensor is primarily from the top of the cloud cover, not the thermal radiation of the surface below the cloud cover. The temperature of the cloud pixel in the thermal infrared remote sensing image is actually the cloud top temperature, but not the surface temperature under the cloud. The lack of the cloud coverage pixel earth surface temperature information damages the integrity and continuity of the space-time distribution of the regional earth surface temperature to a great extent, and seriously hinders and limits the application and development of earth surface temperature products. In many remote sensing applications, such as climate change investigation, evapotranspiration estimation, urban heat islands and the like, the problem that the invention area cannot be comprehensively and effectively monitored because the surface temperature of the cloud pixel cannot be estimated is often encountered. Data loss due to cloud cover is an important obstacle for current surface temperature data applications. The method comprehensively masters the distribution of all-weather LST spatial patterns and the time evolution rule, and has very important practical significance for the invention in the related scientific fields of surface resource environment monitoring, energy balance of a ground gas system, an ecosystem and the like.

In recent years, Ensemble Learning (Ensemble Learning) has become a new trend of development of machine Learning. By combining a series of weak learners (weak learners) into a new model with higher precision and stronger generalization capability, the defects of single function and easy occurrence of overfitting of the original algorithm can be improved. Random Forest (RF), gradient enhanced regression tree (GBRT) and autoregressive extreme gradient ascent algorithm (XGboost) are emerging integrated learning models developed in recent years, have good applicability to different scenes and have outstanding accuracy. Among them, GBRT and XGBoost use boosting's integrated decision tree algorithm, and Random Forest (RF) is using bagging's integrated decision tree algorithm. The three integrated algorithms are widely applied in various fields of biology, medicine, geography, traffic, energy, agricultural engineering, industry and the like, and achieve relatively ideal classification or regression effects. The problem of estimating the cloud-to-ground surface temperature belongs to a multiple regression problem, so that the simulation operation of the cloud-to-ground surface temperature of FY-3D MERSI-II is invented by applying the three integrated regression algorithms.

A Fengyun three-number (FY-3) series satellite is a second generation polar orbit meteorological satellite in China, FY-3D (Fengyun-3D) is a fourth satellite, the satellite is transmitted in a Taiyuan satellite transmission center in 2017 in 11 and 15 months, the orbit type of the satellite is a near polar region sun synchronous orbit, and 10 observation instruments such as a medium resolution ratio spectral imager (MERSI-II), a microwave imager (MWRI), a microwave thermometer (MWHS-2) and the like are carried together. Medium Resolution Spectral Imager II (MERSI-II) is one of the main sensors of the satellite, and the observation capability thereof is greatly improved, so that the satellite can be compared with the united polar orbit satellite newly launched in the United states. In order to fully exert the advantages of basic data resources and technologies of national meteorological satellite data at home and abroad and improve the comprehensive application value and service capability of the domestic meteorological satellite, the basic algorithm invention of the domestic meteorological satellite data needs to be further strengthened.

The cloud spatial distribution on the remote sensing image is related to the solar altitude and the satellite observation angle (wanfe, 2017; WANG et al, 2017). When the sun's incident angle does not coincide with the satellite observation angle, the cloud observed by the satellite and the cloud of the actual position produce a positional deviation (SHI et al, 2017). The greater the relative angle between the direction of incidence of the sun and the direction of observation of the sensor, the greater this offset. Specifically, in practical application, the remote sensing data obtains the cloud position through cloud detection to judge the distribution condition of clear sky and cloud, and determine whether the pixel is suitable for a clear sky or cloudy weather algorithm. However, the cloud pixel detected by the remote sensing image is only the projection of the real cloud cluster in the observation direction of the sensor, and does not correspond to the orthographic projection position of the cloud cluster on the ground surface, nor completely correspond to the influence area of the cloud on the ground surface radiation. The real ground surface area affected by the cloud corresponds to a ground projection area formed by the whole cloud cluster in the direction of the incident angle of the sun, wherein the visible part is a shadow area on the remote sensing image. This effect of the sun-cloud-satellite geometry on the surface parameters is known as the scsg (solar cluster satellite geometry) geometric effect (WANG et al, 2017; WANG et al, 2019). These effects are often ignored in the present invention of estimating the surface temperature under the cloud, and the problem of the sun-cloud-satellite geometric effect (SCSG) of the surface temperature and the radiant flux has not been sufficiently taken into consideration.

Based on the background of the invention, the invention aims at FY-3D MERSI-II data to carry out the estimation invention of the surface temperature under the cloud. The method comprises the steps of firstly, carrying out clear sky surface temperature inversion based on a Two-Factor split window algorithm (TFSWA), considering the sun-cloud-satellite geometric effect of surface temperature and radiant flux, and developing an under-cloud surface temperature estimation algorithm suitable for FY-3D MERSI-II data based on an integrated learning model. And finally, all-weather surface temperature data of the FY-3DMERSI-II are constructed, the further development of thermal infrared remote sensing is promoted, and the comprehensive benefit and the service capability of the domestic satellite are improved.

In general, the cloud-to-ground surface temperature estimation method developed by the invention has better applicability to different surface properties and climate zones, and can provide method reference for cloud-to-ground surface temperature estimation of polar orbit satellite data.

Disclosure of Invention

The invention belongs to the technical field of remote sensing application, and discloses a new method for estimating the surface temperature under the cloud based on a solar-cloud-satellite geometric mechanism. For all-weather remote sensing estimation of the surface temperature, the pixels are divided into a clear sky pixel and a cloud pixel, and the cloud pixel is further divided into a cloud shielding pixel and a cloud covering pixel. And (3) taking the solar-cloud-satellite geometric effect of the surface temperature into consideration, dividing the cloud pixels on the remote sensing image into two types of cloud coverage (the surface incident solar radiation is influenced by the cloud, zone B) and cloud shielding (the surface incident solar radiation is not influenced by the cloud, zone C), and respectively carrying out under-cloud LST estimation. The area C is a direct solar radiation area, and the theoretical clear sky surface temperature represents the cloud pixel surface temperature of the area; the cloud coverage area (area B) is the real projection of the cloud in the sun's viewing direction, and the solar radiation is affected by the cloud shading. Fig. 2 shows a schematic view of a solar-cloud-satellite observation geometry. The subsurface surface temperature under the cloud has multi-factor coupling complexity, and the integrated regression model has excellent performance on nonlinear and non-stationary process feature mining. The invention provides a new method for estimating the temperature of an underground surface under cloud based on an integrated regression learning model, which mainly comprises the following steps:

step one, radiometric calibration and geometric correction of remote sensing image

Radiometric calibration is a necessary step for quantitative remote sensing application, and converts a dimensionless DN value recorded by a sensor original observation into spectral reflectivity or radiance data with actual physical significance. And acquiring the image data of an L1-grade satellite, calculating the satellite reflectivity of a reflection channel and the brightness temperature of a heat emission channel, and performing geometric correction.

Step two, clear sky surface temperature inversion

The planck function was linearly expanded, a two-factor split window algorithm (TFSWA) was proposed through a series of derivations, and the LST was estimated using a series of linear combinations of the two luminance temperature channels, the corresponding band mean surface emissivity and the atmospheric transmittance, as follows:

Ts＝A0+A1T₂₄-A2T₂₅ (1-1)

A0＝a₂₄[D₂₅(1-C₂₄-D₂₄)/(C₂₄D₂₅-C₂₅D₂₄)]-a₂₅[D₂₄(1-C₂₅-D₂₅)/(C₂₄D₂₅-C₂₅D₂₄)] (1-2)

A1＝1+D₂₄/(C₂₄D₂₅-C₂₅D₂₄)+b₂₄[D₂₅(1-C₂₄-D₂₄)/(C₂₄D₂₅-C₂₅D₂₄)] (1-3)

A2＝D₂₄/(C₂₄D₂₅-C₂₅D₂₄)+b₂₅[D₂₄(1-C₂₅-D₂₅)/(C₂₄D₂₅-C₂₅D₂₄)] (1-4)

C₂₄＝ε₂₄τ_24n (1-5)

D₂₄＝[1-τ_24n][1+(1-ε₂₄)τ_24n] (1-6)

C₂₅＝ε₂₅τ_25n (1-7)

D₂₅＝[1-τ_25n][1+(1-ε₂₅)τ_25n] (1-8)

wherein Ts is the surface temperature (K); a0, A1 and A2 are intermediate parameters, T_iIn the case of bright temperature, tau in is the atmospheric transmission rate, i is the thermal infrared band, epsilon i is the surface emissivity corresponding to the thermal infrared band, and a₂₄、a₂₅、b₂₄、b₂₅Is a constant.

Step three, detecting clear sky, cloud shadow, cloud occlusion and cloud coverage area

The temperature estimation is expressed aiming at the cloud pixel, and the cloud pixel and the clear sky pixel need to be distinguished according to the cloud detection product. And (3) estimating the surface temperature under the cloud by adopting different methods according to whether the surface solar radiation corresponding to the pixel under the cloud is influenced by the cloud weakening effect, wherein the cloud shielding and the cloud coverage area need to be distinguished. It is therefore desirable to detect cloud shadows, cloud occlusions, and specific locations corresponding to cloud coverage areas. In this process, orthorectification of the cloud is a key step in correctly characterizing the solar-cloud-satellite geometric relationship (SHI et al, 2017). According to the geometric relation of the sun, the cloud layer, the zenith angle and the azimuth angle of the sun and the satellite and the position of the cloud layer on the remote sensing image are used as input data, the actual position of the cloud can be directly determined, and the surface shadow space position (WANG et al, 2017) can be deduced according to the orthoscopic position. The angle and the cloud height are derived from the geometric positioning file and the cloud height product. The orthonormal position of the cloud is determined using equations (1-9) and then the orthonormal projection position of the corresponding cloud shadow is estimated using equations (1-10) (WANG et al, 2017; WANG et al, 2019).

In the formula (X)_cld，Y_cld) Orthographic projection of the cloud on the ground surface; (X)_shw，Y_shw) The projection position of the cloud shadow on the ground surface; (x)_map,y_map) The cloud position in the remote sensing image is the projection position of the cloud cluster along the satellite observation direction; and H is the cloud top height above ground, calculated by subtracting the ground height (DEM product) from the cloud top height of MERSI-II (cloud height product). Theta_vAnd phi_vIs the satellite transit time sensor observation zenith angle and observation azimuth angle theta_sAnd phi_sRespectively the zenith angle and azimuth angle of the sun as observed by the sensor. When the offset position is calculated, the azimuth angle needs to be converted to-180 degrees, and the east-west direction, the south-east direction and the north-west direction are respectively as follows: 90 deg. -90 deg., 180 deg., 0 deg..

Step four, short wave net radiation inversion and observation geometric effect correction thereof

The earth's surface is often covered by clouds, which are one of the major factors in global or regional scale weather or climate change. The cloud has transmission, absorption and reflection effects on solar radiation, and emits long-wave radiation by itself. The cloud controls both solar incident radiation and atmospheric long-wave radiation. Thus, clouds have a decisive influence on ground radiant flux budget, surface temperature variations and surface energy balance. JIN et al (2000) proposes a neighboring pixel interpolation algorithm (NP) for estimating the pixel earth Surface temperature under the cloud based on a Surface Energy Balance model (SEB). The basic assumption of the algorithm is that cloud coverage pixels and similar clear sky pixels adjacent in space and time mainly cause earth surface temperature difference due to solar radiation difference. And estimating the surface temperature of the pixel under the cloud by establishing a linear relation between the surface long/short wave radiation, the latent/sensible heat flux and the surface temperature. The invention uses short wave net radiation as radiation characteristic variable for surface temperature estimation under cloud.

(1) Short wave net radiation inversion

The method adopts a parameterized algorithm to invert the short-wave net radiation, and only needs to input the solar zenith angle, the atmospheric water vapor, the relative azimuth angle, the atmospheric water vapor content and the atmospheric layer top albedo as input data in the calculation process to invert the short-wave net radiation.

Short wave net radiation inversion narrow band reflectivity rho is firstly carried out_iConverting the conversion into TOA short-waveband broadband albedo gamma, and then obtaining an empirical coefficient by using a parameterization method to calculate NSSR. Based on a MODTRAN model, a linear conversion formula of the narrow-band reflectivity and the wide-band albedo can be established:

γ＝b₀+b₁ρ₁+b₂ρ₂+b₃ρ₃+b₄ρ₄+b₅ρ₅+b₆ρ₆+b₇ρ₇ (1-11)

wherein gamma is the reflection rate of the short-wave broadband on the satellite, rho_iReflectivity of narrow band on satellite for MERSI-II band i, b_iFitting parameters corresponding to the MERSI-II wave band i. Variation of albedo with azimuthLess than the zenith angle variation, and therefore three coarse spacings ranging from 0 deg. -60 deg., 60 deg. -120 deg., and 120 deg. -180 deg. were chosen to represent approximate reflections in the forward, lateral, and reverse directions, respectively.

For a given pair of observed zenith angle (VZA) and Solar Zenith Angle (SZA), the obtained coefficient b is established using equation (5-20)₀～b₇And VZA relationship

In the formula, c_1i-c_4iAre constants given to SZA ═ 0 °, 10 °,20 °, 30 °, 40 °, 50 °, 60 °, and 70 °, respectively. The different SZAs, VZAs and relative azimuth conversion parameters used herein. And establishing a simulation database by using a radiation transmission model by using MODTRAN5.2 and taking typical atmospheric conditions and surface boundary parameters as input. According to the simulation database, establishing parameter lookup tables corresponding to different observation zenith angles, solar altitude angles and relative azimuth angles under the two conditions of clear sky and cloud, and obtaining the broadband albedo on the satellite by retrieving the lookup tables through linear interpolation.

The short wave net radiation is expressed as follows:

NSSR＝a_sE₀μ/d² (1-13)

in the formula, E₀Is solar irradiance on the planet (1357W/m)²) And d is the earth-sun distance. Both parameters are expressed in astronomical units, a_sIs the surface absorption coefficient.

Surface absorption coefficient a_sCan be expressed as the equation of the wide wave band albedo r of the top of the atmosphere layer:

a_s(μ，ω，γ)＝α′-β′γ (1-14)

the intercept α 'slope β' in equations (1-14) are expressed as:

α′＝1-α₁μ^-1-a₂μ^-x-[1-exp(-μ)](a₃+a₄ω^y)μ^-1 (1-15)

in the formula, a₁-a₇Is a constant for various surfaces, μ is the cosine of SZA, and the values of x, y, and z are 0.5.

(2) Short wave net radiation geometry effect correction

And estimating the surface temperature under the cloud based on a surface energy balance model, wherein short wave net radiation is a key input parameter. The geometrical structure of the sun, the cloud and the satellite has universality on the influence of surface radiation, and short wave net radiation data are influenced by the geometric effect of the SCSG, so that the phenomenon of obvious overestimation or underestimation is generated. The distribution position of the cloud has a certain offset on the remote sensing image due to the influence of the SCSG effect. The NSSR geometric effect correction steps are as follows: firstly, determining cloud and shadow range thereof by pixel mainly by using equations 1-9 and 1-10, and calibrating the position relation between each cloud pixel and shadow projection thereof. The NSSR for shadow region (region a) image elements and cloud coverage region (region B) cloud image elements is designated as their corresponding cloud projection region NSSR. The above process is first iteratively performed pixel by pixel for all cloud covered area pixels in the image. For the cloud-shaded region (region C), these picture elements can be illuminated by the sun, so the NSSR of these picture elements should be very similar to clear sky picture elements in close range. In the text, the cloud occlusion area pixels NSSR are filled by interpolation calculation or HANTS algorithm of adjacent clear sky pixels, and finally, short-wave radiation data NSSR corrected by geometric effect is generated_{corrected-cld}。

Step five, selecting and preprocessing characteristic variables

(1) Feature variable selection

The temperature characteristic factor is necessary input data of training data and regression reconstruction cloud pixel LST and is used for realizing the whole reconstruction process. The surface temperature is synergistically affected by geographic location, vegetation, terrain, microtopography, surface active radiation, meteorological information, and the like.

The method selects characteristic variables of regression operation of the surface temperature under the cloud from the aspects of surface properties, solar radiation, atmospheric conditions and sensors. Among all the factors related to the surface temperature, elevation is the first important influence factor, elevation and temperature correlation are quantifiable parameters, the influence of elevation on the surface temperature change is particularly obvious in plateaus and mountainous areas, and the influence on the radiation energy distribution is weakened when the elevation is reduced. DEM, gradient and slope direction are selected as surface terrain characterization parameters. The vegetation selectively absorbs and reflects solar radiation, latent heat and sensible heat exchange are adjusted to influence earth surface thermal radiation information, vegetation parameters are indispensable, and NVI, EVI and LAI are selected as the vegetation parameters. Furthermore, the LST daily profile is related to the short wave net radiant flux absorbed and generated by the surface, and a close relationship between the daily surface temperature cycle and the surface solar net radiation (short wave net radiation) has also been determined. Therefore, in the aspect of solar radiation, the surface short wave net radiation and the solar altitude are considered. According to the short-wave net radiation inversion process, input data of surface short-wave net radiation inversion comprise solar constants, solar altitude angles, solar azimuth angles, atmospheric Water Vapor (WVC), sensor observation angles and the like, and the characteristic variables are comprehensive representations of geographic positions, solar radiation, sensor observation conditions and atmospheric conditions. Table 1 gives detailed information of the time resolution, unit, and the like of each feature variable.

TABLE 1 estimation of characteristic variables for subsurface surface temperature

(2) Harmonic optimization processing of NVI, EVI and LAI data

NVI, EVI and LAI are important variables indicating the covering characteristics of the earth surface vegetation, and due to the influences of factors such as atmospheric conditions, aerosol and sensor faults and the like in the observation of remote sensing data, certain data loss and low-quality data exist in a MERSI-II LAI data product, and the model training application is seriously influenced in the estimation of the temperature of the earth surface under clouds. In order to obtain a high-precision mode simulation result, missing pixels of NVI, EVI and LAI data are filled by using an HANTS algorithm, low-quality data caused by cloud pollution or other factors are smoothed and filtered, and continuous and complete long-time sequence high-quality training data are generated.

(3) Characteristic variable collinearity analysis

In order to select high quality feature variables, it is necessary to check whether co-linearity problems exist between the variables. And calculating a PEARSON characteristic correlation matrix heat map among characteristic variables based on a Seaborn statistical analysis visualization module, and removing the variables with larger correlation.

Step six, model establishment and parameter optimization

(1) Determination of training data set

In order to accurately represent the complex relation between the LST and the characteristic variables, a fitting regression model from each characteristic variable to the earth surface temperature is constructed, high-quality clear sky and shadow data are selected as a training data set, the earth surface temperature of cloud pixels is used as a target variable, and the regression model is established. The third-party modules Scikit-left and XGboost based on Python utilize three regression models of RF, GBRT and XGboost to carry out learning training on surface temperature data sets of clear sky and shadow pixels, utilize a train _ test _ split function to randomly extract and divide sample data, and two parts of 80% training partitions and 20% testing partitions are respectively applied to a training model and a testing model.

(2) Parameter setting and optimization

To prevent the overfitting phenomenon, the number of model regression trees (max-depth) parameters in the three models were adjusted and scores of the training data were checked by 10-fold cross validation. To view model performance, their training times were viewed for different training samples.

Seventhly, training is carried out on clear sky and shadow regions based on an integrated learning model, and a regression prediction model between the characteristic variables and the LST is established

The method is characterized in that the characteristic that the surface temperature under the cloud has multi-factor coupling is considered, the integrated regression algorithm has excellent performance for the characteristic excavation of a nonlinear transformation process and a non-stable characteristic, on the basis of fully considering the surface temperature and the radiated solar-cloud-satellite geometric effect, three integrated regression algorithms of RF, GBRT and XGboost are adopted, and a target variable (FY-3D MERSI-II surface temperature) and a preprocessed characteristic variable are used for carrying out characteristic training simulation to establish an LST regression prediction model aiming at a clear area and a shadow area. And meanwhile, calculating the Importance scores and the sequence of the characteristic variables based on a Python third-party extension package scimit-lean characteristic variable Importance calculation method (recommendation opportunity).

Step eight, inputting an image to be calculated, and performing cloud occlusion and cloud coverage area under-cloud LST estimation by utilizing a regression model

And predicting the surface temperature of all pixels of the cloud shielding area and the cloud coverage area in the invention area by adopting the under-cloud LST regression model established in the previous step, and performing precision evaluation on the estimated under-cloud surface temperature by utilizing CLDAS assimilation LST data.

Compared with other methods for estimating the temperature of the subsurface surface under the cloud, the method has the advantages that:

1. the integrated regression algorithm is used for estimating the surface temperature under the cloud, the calculation speed is high, and the data processing efficiency is improved.

2. The algorithm is not influenced by factors such as the type of the underlying surface, the size of the cloud coverage range and the like, and the applicability of the cloud pixel LST estimation is improved.

3. The cloud earth surface temperature estimation is carried out by adopting the integrated regression models of the three decision tree types, the model fitting accuracy can be improved, and the overfitting phenomenon can be prevented through parameter optimization.

Drawings

FIG. 1 is a flow chart of an integrated regression algorithm for surface temperature under cloud with consideration of the surface temperature and radiative SCSG effects

FIG. 2 is a schematic view of a sun-cloud-satellite observation geometry

FIG. 3 FY-3D MERSI-II cloud shadow detection and cloud blocking and cloud coverage area pixel statistics

FIG. 4 FY-3D MERSI-II four seasons SCSG effect pixel classification results

FIG. 5 is a schematic diagram of NSSR geometry correction

FIG. 6 short wave net radiance data after geometric effect correction

FIG. 7 machine learning regression model Pearson feature correlation matrix heatmap

FIG. 82020 year 5 month 2 day clear sky and LST density distribution diagram of each input characteristic variable and LST in shadow region

FIG. 9 maximum Tree depth and model validation set score

FIG. 10 time consumption for different training sample numbers for the three models

FIG. 11 training accuracy versus time used

FIG. 12 importance of features of regression operation of surface temperature under the cloud for RF, GBRT and Xgboost algorithms

FIG. 13 is a FY-3D MERSI-II cloud-based subsurface surface temperature image map based on three integrated regression models

FIG. 14 is a scatterplot of sub-cloud surface temperature and CLDAS temperature based on FY-3D MERSI-II data

Detailed Description

To implement the proposed method, we take FY-3D mid-resolution imager MERSI-II as an example, and perform an experiment of the estimation of the surface temperature under the cloud. We need to perform geometric registration between the target image and the reference image and determine whether the pixel is clear sky or cloud covered. The process comprises three parts: radiometric calibration and geometric correction of a reference image, inversion of clear sky surface temperature, cloud detection and identification of cloud occlusion, cloud coverage pixels, shadow pixels and clear sky pixels, selection and pretreatment of characteristic variables, model establishment and training, and estimation of cloud-to-ground surface temperature.

The Medium Resolution Spectral Imager II (Medium Resolution Spectral Imager, MERSI-II) is one of core instruments carried on the FY-3D, integrates the functions of the original wind cloud three-satellite Imager MERSI and VIRR, and is provided with 25 channels in total, wherein the channels 1-19 are solar reflection channels (0.4-2.1 mu m), the channels 20-25 are infrared emission channels (3.8-12.5 mu m), the spatial resolutions of the channels 1-4 and 24-25 are 250 meters, and the spatial Resolution of the channels 5-23 is 1000 m. FY-3D MERSI-II mainly provides three-level data products of L1, L2 and L3 at present, 25 data sets are in total, and the data storage format is HDF5 scientific data set format.

The invention develops an FY-3D MERSI-II cloud-below-surface temperature integrated regression algorithm by taking a solar-cloud-satellite geometric effect as a starting point. Firstly, starting from the influence of solar radiation, the earth surface and atmospheric conditions on the earth surface temperature, selecting 7 variables as earth surface thermal radiation correlation factors, mainly correcting the geometric effect of a solar-cloud-satellite on key characteristic variable short wave net radiation, and carrying out pretreatment such as co-linear analysis and the like on each characteristic variable. Three integrated regression models of a Random Forest (RF), a Gradient regression Tree (GBRT) and an autoregressive Extreme Gradient Boosting algorithm (XGboost) are adopted, FY-3D MERSI-II clear sky and cloud shadow region data are used as training data sets, cloud shielding and cloud coverage area surface temperature are used as target variables to carry out simulation prediction, and a set of under-cloud LST integrated regression algorithm fully considering the surface temperature and the radiant flux sun-cloud-satellite geometric effect is established. Based on the algorithm, the data of four days in different seasons in the North China area in 2020 is adopted to predict the surface temperature under the cloud. Performing clear sky and sub-cloud surface temperature estimation based on FY-3D MERSI-II requires MERSI-II geometric positioning data and partial data products (cloud-high, cloud detection, NVI and LAI ten day products) as auxiliary data. The present invention will be described in further detail with reference to the accompanying drawings. Table 2 shows the FY-3D MERSI-II dataset used for all-weather surface temperature estimation.

TABLE 2 FY-3D MERSI-II dataset for all-weather surface temperature estimation

Step one, FY-3D MERSI-II data preprocessing

Acquiring satellite image data of level L1 of FY-3D MERSI-II according to the selected North China invention district range. The part of work mainly comprises three parts of reflection channel on-satellite reflectivity calculation, thermal emission channel brightness temperature calculation and geometric correction.

1. On-board reflectivity calculation

The MERSI-II L1 level data adopts unsigned integer data to store the original observation pixel so as to reduce the data size, and the gray value is output after normalization and correction by multiple probes. Radiometric calibration is a necessary step for quantitative remote sensing application, and converts a dimensionless DN value recorded by a sensor original observation into spectral reflectivity or radiance data with actual physical significance.

The reflectivity of the sun reflection wave band CH 1-19 on the satellite is calculated, and firstly, the DN value needs to be converted into the visible light calibration reflectivity Ref. The data sets EV _1KM RefSB and EV _250 Aggr.1KM RefSB are normalized DN values of 1KM reflection channels (CH 5-19) and 250m reflection channels (CH 1-4), respectively, and the scaled reflectivity conversion formula is as follows:

dn＝DN*Slope+Intercept (1-17)

Ref＝Cal₂*dn²+Cal₁*dn+Cal₀ (1-18)

in the formula, Cal₀、Cal₁And Cal₂Scale coefficients (corresponding to

columns

1, 2 and 3, respectively) for corresponding channels in the data set VIS _ Cal _ Coeff, Slope and Intercept are attributes of the data set EV _1KM _ RefSB and EV _250_ aggr.1km _ RefSB.

Then converting the visible light scaling reflectivity Ref into the entrance pupil radiance L_toa：

Combined with the entrance pupil radiance L_toaI.e. convertible to apparent reflectance ρ_toa：

In the formula, E₀For tunnel extraterrestrial Solar Irradiance, the attribute Solar _ Irradiance, DES can be taken from the L1 data set attribute, EarthSunDistanceRatio, L1 data set. Mu is cos theta, theta is the solar zenith angle, and can be obtained from the SolarZenith attribute of the MERSI-IIL1 scientific data set, and as the storage mode is scaled integer data, the real solar zenith angle theta is calculated by using a corresponding slope coefficient and an Intercept coefficient according to the following formula.

θ＝SolarZenith×Slope+Intercept (1-21)

2. Luminance temperature calculation

FY-3D MERSI-II data EV _1KM _ Emissive and EV _250_ Aggr.1KM _ Emissive are respectively the amplified radiance values of 1KM emission channels (CH 20-23) and 250m emission channels (CH 24-25), gain (scales) and offset (offsets) of each waveband I are obtained according to attributes, and the radiance I of the waveband I is calculated according to the following formula_i(W·m^-2·sr^-1·μm^-1) (xuna et al, 2019):

I_i＝(DN-offset)·scales (1-22)

converting the radiance of the thermal infrared channel (CH 24-25) into the blackbody brightness temperature, and obtaining the equivalent blackbody brightness temperature T through inverse transformation calculation of the equivalent Center WaveLength effect _ Center _ waveLength (unit mum) and the channel radiance by using a Planck function_eiThe calculation formula is as follows:

in the formula, λ_iIs the effective Center WaveLength of the i (i-24, 25) th band, which can be obtained according to the attribute variable effect _ Center _ WaveLength (unit μm) in the L1 file; as can be seen from Table 3-1, the center wavelengths λ of the 24 th and 25 th channels of MERSI-II_iRespectively take lambda₂₄11.8 μm and λ₂₅＝12.0μm；C₁And C₂Respectively 1 st and 2 nd spectral constants, respectively taking C₁＝1.19104356×10^-16W/m²And C₂＝1.4387685×10⁴μ m. multidot.K. Finally, T is corrected by utilizing a channel brightness temperature correction coefficient (TbbCorr _ Coeff)_eiConversion into channel black body brightness temperature T_i：

T_i＝A*T_ei+B (1-24)

In the formula, T_iIs the luminance temperature of the i (i-24, 25) th band of the MERSI-II, i.e. the luminance temperature T corresponding to the thermal infrared channels 24 and 25₂₄And T₂₅(ii) a A and B are attribute variables TBB _ Trans _ Coefficient _ A and TBB _ Trans _ Coefficient _ B, respectively, inside the L1 file.

3. Geometric correction

The FY-3DMERSI-II data adopts a multi-element detector parallel scanning technology, and the image geometric distortion phenomenon is caused by factors such as the imaging geometric characteristics of a sensor, terrain fluctuation, surface curvature, satellite jitter and the like. The invention adopts a Geographic Lookup Table (GLT) method to carry out system positioning work. Specifically, the geographic position lookup table is generated through the geometric positioning information of the corresponding data input by the geometric positioning file. The actual geographic position of each pixel in the original data can be calculated through a lookup table. However, due to the influence of many factors such as terrain, surface and atmospheric refraction, the geometric positioning error of 2-10 pixels still exists in the primary system correction result, and the application requirement cannot be met. Therefore, the method adopts a remote sensing image automatic matching technology based on a Moravec angular point detection algorithm to carry out geometric fine correction on the preliminarily corrected data. The algorithm effectively reduces the local geometric distortion of the image by automatically selecting control points, improves the geometric positioning precision of data, basically controls the error in 3 pixels, and basically meets the requirement on geometric correction precision. The accurate geometric positioning lays a good geometric foundation for the smooth development of all-weather surface temperature estimation work of subsequent FY-3DMERSI-II data. The accurate geometric positioning lays a good geometric foundation for the smooth development of all-weather surface temperature estimation work of subsequent FY-3DMERSI-II data. And preparing data such as NDVI, LAI, DEM, gradient, slope, solar altitude angle, satellite observation angle and the like of the corresponding region and time of the data, and ensuring the projection and resolution of the data to be consistent.

Step two, FY-3D MERSI-II clear sky surface temperature inversion

Linearly developing the Planck function, proposing a two-factor split window algorithm (TFSWA) through a series of deductions, estimating the LST by using a series of linear combinations of two brightness temperature channels, corresponding waveband average surface emissivity and atmospheric transmittance, wherein the LST is between-50 and 50 DEG, and the MERSI-II TFSWA algorithm waveband 24 constant term a₂₄＝-53.477，b₂₄0.3951; the constant term of the wave band 25 is a₂₅＝-57.087，b₂₅＝0.4292。

Step three, detecting cloud shadow, cloud occlusion and cloud coverage area

The MERSI-II cloud detection product mainly provides MERSI-II satellite data cloud and clear sky distribution information, and is basic auxiliary data for underground surface temperature estimation research. The invention uses the cloud detection result obtained by comprehensive analysis of a multi-band (3, 4, 6, 7, 19, 20, 21, 24, 25) characteristic threshold method of an FY-3D MERSI-II cloud detection product. According to whether the surface solar radiation corresponding to the under-cloud pixels is affected by the cloud weakening effect or not, different methods are adopted to estimate the under-cloud surface temperature, and specific positions corresponding to cloud shadows, cloud shelters and cloud coverage areas need to be detected. Therefore, by combining the angle and cloud height data of FY-3DMERSI-II, cloud shading, cloud coverage and cloud shadow detection are carried out on the data of 4 days in different seasons in 2020 in North China by using the method in step three. Fig. 3 shows the statistical results of FY-3D mers i-II cloud shadow, cloud cover and clear sky pixel, and fig. 4 shows the classification results of four image pixels in different seasons of 2020.

Step four, short wave net radiation inversion and geometric effect correction

Short wave net radiation is an important characteristic parameter in the inversion of the surface temperature under the cloud. And inverting FY-3D MERSI-II short wave net radiation by adopting a parameterized algorithm, and correcting the SCSG geometric effect of the short wave net radiation. The MERSI-II has two groups of visible light wave bands, wherein the bandwidth of the wave band is 8-14 narrow (20nm), the MERSI-II is mainly used for ocean color, plankton and the like, the wave bands are 1-7 designed for terrestrial application, and the seven reflected solar Radiation (RSB) channels and 1 water vapor absorption channel are mainly used for inversion of the MERSI-II on-satellite albedo (Table 3).

TABLE 3 MERSI-II reflected solar radiation band for NSSR inversion

Short wave net radiation inversion first narrow band reflectivity rho_iConverting the conversion into TOA short-waveband broadband albedo gamma, and then obtaining an empirical coefficient by using a parameterization method to calculate NSSR. Based on a MODTRAN model, a linear conversion formula of the narrow-band reflectivity and the wide-band albedo can be established:

γ＝b₀+b₁ρ₁+b₂ρ₂+b₃ρ₃+b₄ρ₄+b₅ρ₅+b₆ρ₆+b₇ρ₇+b₁₉ρ₁₉ (1-25)

wherein gamma is the reflection rate of the satellite short wave broadband, rho_iReflectivity of narrow band on satellite for MERSI-II band i, b_iFitting parameters corresponding to the MERSI-II wave band i. The albedo change with azimuth angle is less than the zenith angle change, so three coarse intervals ranging from 0-60 °, 60-120 ° and 120-180 ° are selected to represent the approximate reflection in the forward, lateral and reverse directions, respectively.

Table 4 main parameter settings in MODTRAN5.2 atmospheric radiation transmission simulation

And F Y-3D MERSI-II short wave net radiation can be estimated according to the parameterization process of the step four.

Step five, selecting and preprocessing characteristic variables

(1) Feature variable selection

Data of products in the ten-day days of the MERSI-II LAI leaf area index and products in the ten-day days of the NVI vegetation index are mainly used for searching similar pixels and used as characteristic variables of an integrated regression algorithm, and the two data sets are respectively longitude and latitude projection and Hammer projection of GLL and the like and are stored by adopting the global 10 degrees multiplied by 10 degrees and 0.01 degrees spatial resolution.

The MERSI-II vegetation index ten-day product data comprises NDVI, EVI, sensor and satellite angle information, quality control information and the like. The MERSI-II leaf area index ten-day product mainly comprises a leaf area index and corresponding data quality control information. The cloud top height refers to the height of the cloud layer top, the unit is hPa or m, and all data are uniformly converted into geometric height units and then used for calculating the cloud shadow position and correcting the geometric effect of the radiation data. All auxiliary data are unified into a WGS84 geographic coordinate system, 0.01-degree resolution and other latitude and longitude projections.

(2) Variable pre-processing

a. Short wave net radiation geometry effect correction

And in consideration of the universality of the SCSG effect of the radiation data, carrying out SCSG geometric effect correction on the short-wave net radiation of the key characteristic variable for estimating the temperature of the subsurface cloud, and referring to the step four in the concrete method. Fig. 5 shows a schematic diagram of the geometric effect of the short-wave radiation data, which shows MERSI-iinsssr in north china observed on 5, month and 2 days 2020, wherein fig. 5(a) is an RGB true color image map, and two straight lines m and n are used for interpolation profile curves before and after NSSR correction in the invention. FIG. 5(b) shows the classification result of each type of pixel: the cloud shadow area in the area A, the cloud coverage area in the area B, the cloud shielding area in the area C and the clear empty area in the area D. To better characterize the effect of sun-cloud-satellite geometry on NSSR, raw NSSR data were subtracted from the SCSG geometry-corrected NSSR, detailed in fig. 5 (c). From the difference map, it can be seen that the shadow region (region a) in the image is negative and the cloud-shielded region (region C) is positive, indicating that the NSSR is underestimated in the cloud-shielded region (region C) and overestimated in the shadow region (region a) if the geometric effect of the sun-cloud-satellite is neglected. The specific magnitude of this difference varies with the physical and optical characteristics of the cloud. In the example given, this difference can be up to about 800W/m²Approximately 80% of NSSR in clear sky. FIGS. 5(d) and (e) are NSSR differences (NSSR) at two lines of m and n_corrected-cld-NSSR_clear) In the cross-sectional view, the left dotted line corresponds to a cloud-masked region of the original data, which is an underestimated region, and the right dotted line corresponds to a shaded region of the original data, which belongs to an NSSR overestimated region. Fig. 6 shows an image map after geometric effect correction of short wave net radiation data in four seasons (20200218, 20200502, 20200715 and 20201113), and corresponding cloud detection results, namely pixel classification results.

Harmonic optimization processing of NVI, EVI and LAI data

In order to obtain a high-precision mode simulation result, missing pixels of FY-3DMERSI-IINVI, EVI and LAI data are filled by using an HANTS algorithm, low-quality data caused by cloud pollution or other factors are smoothed and filtered, and continuous and complete long-time sequence high-quality training data are generated.

c. Characteristic variable collinearity analysis

In order to select high quality feature variables, it is necessary to check whether co-linearity problems exist between the variables. Based on the Seaborn statistical analysis visualization module, the PEARSON feature correlation matrix heat map (fig. 7) among the above 9 feature variables is calculated, and it can be seen that the correlation between SZA and NSSR is above 0.7, and the collinearity problem exists, so the SZA variables are removed. Other variables do not present significant co-linearity problems. In addition, the feature importance analysis was performed to find that the slope score was the lowest, and therefore, in order to reduce the noise effect, the slope feature variables were removed, and therefore, 7 effective feature variables were determined. In 2 days of 5 months in 2020, 802468 clear sky and shadow pixels exist; the pixels are widely distributed to cover different land covering types, vegetation covering degrees, terrain features, solar radiation conditions and the like. Based on the observation data of 5, month and 2 days in 2020, the relation between the temperature and each characteristic variable of the LST (local surface temperature) of the clear sky and cloud shadow area, namely the remote sensing visual area, is analyzed. Fig. 8 shows the density distribution of LST and 7 characteristic variables determined in clear sky and shaded areas on day 5, month 2 of 2020. As can be seen from the graph, the surface temperature and the distribution of each parameter are more concentrated, and there is no obvious linear relationship between the surface temperature and the parameters under the cloud. This complex relationship confirms that the surface temperature is commonly affected by many surface variables, and the coupling effect between factors contributes to this complexity.

Step six, model establishment and parameter optimization

(1) Determination of training data set

In order to accurately represent the complex relation between the LST and the characteristic variables, a fitting regression model from each characteristic variable to the earth surface temperature is constructed, high-quality clear sky and shadow data in 5-month and 2-day 2020 is selected as a training data set, the earth surface temperature of cloud pixels is used as a target variable, and the regression model is established. The third-party module scimit-spare and XGboost base based on Python utilizes three regression models of RF, GBRT and XGboost to carry out learning training on surface temperature data sets of clear sky and shadow pixels, and in the process, sample data is randomly extracted and divided by utilizing a train _ test _ split function, and two parts of an 80% training partition and a 20% testing partition are respectively applied to a training model and a testing model.

(2) Parameter setting and optimization

In order to prevent the overfitting phenomenon, the number of model regression trees (max-depth) parameters in the three models are adjusted, scores of training data are checked through 10-fold cross validation, and as a result, as shown in fig. 9, the validation scores of the models are increased as max-depth is increased from 0 to 20, and the scores of the models reach extreme values when RF, GBRT and XGBoost are respectively 8 max-depth, 5 and 4 max-depth. Max _ depth is chosen to be 8, 5, 4 as the optimal parameters for RF, GBRT and XGBoost. In order to check the model performance, the training time (fig. 10) of different training samples is checked, and the time consumed by the same number of training samples, namely GBRT, is the shortest, and the XGboost time is the longest. Fig. 11 shows the time taken by each model to reach different accuracies, and from the graph analysis, the time consumed by the random forest and the XGBoosting algorithm of the three methods of the integrated tree class is relatively long, but the relatively high accuracy is reached.

Seventhly, performing regression operation on the surface temperature under the cloud

Aiming at the eastern region in Asia, a plurality of images are selected in four seasons in 2020 to carry out a cloud subsurface temperature reconstruction experiment, and the proposed method is evaluated. And calculating the Importance score and the sequence of each feature variable based on a Python third-party extension package scidit-leann feature variable Importance calculation method (recommendation import). FIG. 11 shows the results of feature importance calculations for three regression models. As can be seen from the graph, the characteristic variables of the first three of the parameters used by the three models are consistent, the parameter with the largest influence on the temperature of the subsurface cloud surface is DEM, the characteristic importance is from 0.87 of XGboost to 0.95 of RF, NWDI (0.76-0.82) and NSSR (0.35-0.39) are arranged in the second place, and finally the vegetation index and the leaf area index are arranged in the last place. Different from the DEM, the optical path length of solar radiation changes, and in addition, the thickness of the atmosphere through which the sunlight passes when reaching the ground is higher, the thinner the atmosphere is, and the stronger the radiation is. The elevation condition determines the energy distribution, and therefore has a large influence on the surface heat radiation in a local range. Short wave net radiation is a comprehensive characterization of solar radiation, cloud and earth surface, and is one of the leading factors of earth surface temperature change. Generally, within the invention zone, the conditions that have the greatest effect on the subsurface temperature under clouds are terrain conditions, followed by NDWI and short wave net ground radiation, and finally vegetation coverage conditions. FIG. 12 is a shadow map of the FY-3D MERSI-II cloud-based subsurface temperature distribution estimated based on three machine-learned regression model algorithms.

The method is used for carrying out precision verification on the algorithm by utilizing surface temperature Data of a China weather service Data Assimilation System (CMA Land Data Assimilation System, CLDAS). FIG. 14 shows a plot of sub-cloud surface temperature versus CLDAS temperature scatter based on FY-3D MERSI-II data. Analysis in comparison to CLDAS surface temperature data showed that the RMSE accuracy of the method was about 2.84K to 5.43K. The result shows that the earth surface temperature estimated by the cloud-below-ground surface temperature integrated regression algorithm considering the solar-cloud-satellite observation geometric effect has high accuracy, and the influence of the terrain and vegetation on the earth surface temperature can be well described.

Claims

1. A method for estimating the temperature of an underground cloud surface based on a sun-cloud-satellite observation geometry is characterized by comprising the following steps:

s1: acquiring and preprocessing L1-grade satellite image data of FY-3D MERSI-II;

s2: inverting the surface temperature in clear sky;

s3: cloud shadow, cloud occlusion and cloud coverage area detection;

s4: inverting FY-3D MERSI-II short wave net radiation by adopting a parameterized algorithm, and correcting SCSG geometric effect on the short wave net radiation;

s5: selecting characteristic variables of cloud-below-ground surface temperature regression operation and preprocessing the characteristic variables;

s6: setting and optimizing parameters for the three integrated regression models;

s7: aiming at the LST in the clear sky and the shadow area, establishing an integrated regression model from each characteristic variable to the surface temperature under the cloud;

s8: and inputting an image to be calculated, and carrying out LST estimation on the cloud occlusion and the cloud coverage area by adopting the established regression model.

2. The method for estimating the temperature of the subsurface of the cloud based on the solar-cloud-satellite observation geometry according to claim 1, wherein the step S1 specifically comprises the following steps:

s11: calculating the reflectivity on the satellite;

s12: calculating the brightness temperature;

s13: based on the fact that a Geographic Lookup Table (GLT) method is adopted to carry out system positioning work, and a Moravec corner detection algorithm is used for carrying out geometric fine correction on the preliminarily corrected data.

3. The method for estimating the temperature of the subsurface of the cloud based on the solar-cloud-satellite observation geometry according to claim 1, wherein the step S2 specifically comprises the following steps:

the LST is estimated using a series of equations for the two luminance temperature channels, the corresponding band average surface emissivity and the atmospheric transmittance.

4. The method for estimating the temperature of the subsurface of the cloud based on the solar-cloud-satellite observation geometry according to claim 1, wherein the step S3 specifically comprises the following steps:

determining the orthonormal position of the cloud:

the orthographic projection positions of the corresponding cloud shadows are then estimated:

in the formula (X)_cld，Y_cld) At the surface as cloudsOrthographic projection; (X)_shw，Y_shw) The projection position of the cloud shadow on the ground surface; (x)_map,y_map) The cloud position in the remote sensing image is the projection position of the cloud cluster along the satellite observation direction; h is the height of the cloud top above the ground, and is calculated by subtracting the height of the ground from the height of the cloud top of MERSI-II; theta_vAnd

is the satellite transit time sensor observation zenith angle and observation azimuth angle theta_sAnd

respectively the zenith and azimuth angles of the sun as observed by the sensor.

5. The method for estimating the temperature of the subsurface of the cloud based on the solar-cloud-satellite observation geometry according to claim 1, wherein the step S4 specifically comprises the following steps:

short wave net radiation inversion first narrow band reflectivity rho_iConverting the conversion into TOA short-waveband broadband albedo gamma, and then obtaining an empirical coefficient by using a parameterization method to calculate NSSR; based on an MODTRAN model, a linear conversion formula of the narrow-band reflectivity and the wide-band albedo can be established:

γ＝b₀+b₁ρ₁+b₂ρ₂+b₃ρ₃+b₄ρ₄+b₅ρ₅+b₆ρ₆+b₇ρ₇+b₁₉ρ₁₉ (1-11)

wherein gamma is the reflection rate of the satellite short wave broadband, rho_iReflectivity of narrow band on satellite for MERSI-II band i, b_iFitting parameters corresponding to the MERSI-II wave band i;

then, determining the cloud and the shadow range thereof pixel by pixel, calibrating the position relation between each cloud pixel and the shadow projection thereof, generating a cloud shielding area pixel NSSR by adopting Hants algorithm interpolation calculation, carrying out geometric position correction on the NSSR of the cloud coverage area, and finally generating short wave radiation data subjected to geometric effect correction.

6. The method for estimating the temperature of the subsurface earth based on the solar-cloud-satellite observation geometry according to claim 1, wherein the characteristic variable of the regression operation of the temperature of the subsurface earth can be a combination of one or more of the following variables: leaf area index, normalized vegetation index, enhanced vegetation index, digital ground elevation, slope direction, normalized water body index, solar elevation angle, and short wave net radiation.

7. The method for estimating the temperature of the subsurface of the cloud based on the solar-cloud-satellite observation geometry according to claim 1, wherein the step S6 specifically comprises:

the method comprises the steps of learning and training temperature data sets of clear sky and shadow pixels by three regression models of RF, GBRT and XGboost, randomly extracting sample data by a train _ test _ split function, dividing the sample data into a training partition and a testing partition, and applying the training partition and the testing partition to a training model and a testing model respectively.

8. The method for estimating the temperature of the subsurface of the cloud based on the solar-cloud-satellite observation geometry according to claim 1, wherein the step S7 specifically comprises:

the method comprises the steps of adopting three integrated regression algorithms of RF, GBRT and XGboost, carrying out regression operation by taking FY-3D MERSI-II in-cloud surface temperature as a target variable, and carrying out precision evaluation on the estimated in-cloud surface temperature by utilizing CLDAS assimilation LST data.

9. The method for estimating the temperature of the subsurface of the cloud based on the solar-cloud-satellite observation geometry according to claim 1, wherein the step S8 specifically comprises:

and predicting the surface temperature under the clouds in the invention area by adopting the constructed regression model, and combining the predicted cloud pixel LST and clear sky LST to obtain a seamless all-weather LST image of the invention area.

10. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the method according to any one of claims 1-9.