A method for deriving and screening land surface covariates for land surface temperature reconstruction

Abstract

This invention relates to the field of land surface temperature reconstruction technology, and discloses a method for deriving and screening land surface covariates for land surface temperature reconstruction. The method includes obtaining multi-source land surface covariates from all pixels in the study area; screening trend covariates from the multi-source land surface covariates based on the temporal correlation between each covariate and satellite land surface temperature; selecting residual covariates from the multi-source land surface covariates; constructing a trend variable set by the trend covariates and their time-series statistical derivatives, and calculating spatial and temporal derivatives for the residual covariates; constructing a residual variable set by the residual covariates and their temporal and spatial derivatives; and performing feature screening on the trend variable set and the residual variable set to generate optimized trend variable subsets and optimized residual variable subsets. This invention can effectively increase the number of available feature variables in a two-stage model while reducing the overlap of feature variables between the two-stage reconstruction models.

Description

Technical Field

[0001] This invention relates to the field of land surface temperature reconstruction technology, and more specifically to a method for the derivation and screening of land surface covariates for land surface temperature reconstruction. Background Technology

[0002] The temporal variation of land surface temperature is controlled by a periodic thermodynamic process driven by solar radiation, while short-term fluctuations are also influenced by factors such as weather systems, land cover, and human activities. Previous studies have proposed a two-stage reconstruction method that decomposes satellite land surface temperature (LST) into periodic and residual components. However, the use of surface covariates in the two-stage reconstruction of satellite LST still has the following shortcomings:

[0003] (1) The surface covariates are mainly actual observation results, and their number is limited. When the surface covariates are introduced into the two-stage reconstruction model of periodic components and residual components, it will lead to a high degree of overlap of covariates in the two-stage model, which will weaken the complementary advantages of the two-stage reconstruction.

[0004] (2) The same surface covariate cannot simultaneously satisfy the opposing requirements of the two-stage model, that is, the two-stage model has the difference between periodicity and heterogeneity;

[0005] (3) Surface covariates generally integrate multiple factors, some of which have low correlation with satellite surface temperature and lack interpretability. Summary of the Invention

[0006] The purpose of this invention is to provide a method for the derivation and screening of surface covariates for surface temperature reconstruction, in order to solve the technical problems in the prior art where the selection of covariates and derived variables in a two-stage reconstruction model lacks a clear method, is highly subjective, lacks clear physical guidance in the calculation of derived variables, and uses a relatively small number of derived variables, thus failing to fully explore the information of covariates in terms of the periodicity and heterogeneity of satellite surface temperature (LST).

[0007] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0008] A method for deriving and screening surface covariates for land surface temperature reconstruction includes the following steps:

[0009] Step 1: Based on satellite imagery and Earth observation data of the study area, obtain multi-source surface covariates for all pixels in the study area. The multi-source surface covariates are various environmental factors that affect the spatiotemporal distribution characteristics of surface temperature.

[0010] Step 2: Based on the temporal correlation between each covariate in the multi-source surface covariate and satellite surface temperature, select trend covariates from the multi-source surface covariates to characterize the annual cycle variation of surface temperature, and select residual covariates from the multi-source surface covariates to characterize the spatiotemporal comprehensive distribution characteristics of surface temperature.

[0011] Step 3: Calculate the time-series statistical derived variables of the trend covariate using a time-series sliding window, construct a trend variable set from the trend covariate and its time-series statistical derived variables, and calculate the spatial derived variables of the residual covariate using the spatial texture analysis method and the time-series first derivative method, construct a residual variable set from the residual covariate and its time and spatial derived variables.

[0012] Step 4: Perform feature filtering on the trend variable set and the residual variable set respectively to generate optimized trend variable subsets and optimized residual variable subsets.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] This invention uses mathematical transformations, spatiotemporal aggregation, or physical models to calculate new derived variable features from surface covariates. Introducing these derived variables improves the performance of the two-stage reconstruction model. On one hand, it effectively increases the number of available feature variables in the two-stage model, reducing the overlap of feature variables between the two stages while maintaining the number of features. On the other hand, it extracts information from surface covariates that is more relevant to satellite surface temperature (LST), including the periodic and heterogeneous characteristics of LST, thus meeting the feature variable requirements of the two-stage model and increasing its interpretability. Attached Figure Description

[0015] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the method for constructing covariate-derived features provided in an embodiment of this application.

[0017] Figure 2 A comparison chart showing the surface temperature reconstruction results using covariate-derived features and using only covariates, provided for embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] like Figure 1 As shown, this invention provides a method for the derivation and screening of surface covariates for land surface temperature reconstruction. This embodiment addresses the problem of insufficient surface covariates in two-stage land surface temperature reconstruction by using mathematical transformations, spatiotemporal aggregation, or physical model calculations to obtain new derived variable features from the surface covariates. Introducing these derived variables improves the performance of the two-stage reconstruction model. On one hand, it effectively increases the number of available feature variables in the two-stage model, reducing the overlap of feature variables between the two stages while maintaining the number of features. On the other hand, it extracts information from the surface covariates that is more relevant to land surface temperature, including the periodic and heterogeneous characteristics of land surface temperature, thus meeting the feature variable requirements of the two-stage model and increasing its interpretability.

[0020] The method for constructing covariate-derived features specifically includes the following steps:

[0021] Step 1: Based on satellite imagery and Earth observation data of the study area, obtain multi-source surface covariates for all pixels in the study area. The multi-source surface covariates are various environmental factors that affect the spatiotemporal distribution characteristics of surface temperature.

[0022] The multi-source surface covariates specifically include vegetation index, multi-band dual-polarization microwave brightness temperature, DEM topographic data, 2m air temperature, transit cumulative downward shortwave radiation, and soil moisture.

[0023] The multi-band dual-polarization microwave brightness temperature data consists of eight types of data, namely, H and V polarization data at four frequencies: 18.7GHz, 23.8GHz, 36.5GHz, and 89.0GHz, which can be represented as: 18.7GHz_H, 23.8GHz_H, 36.5GHz_H, 89.0GHz_H, 18.7GHz_V, 23.8GHz_V, 36.5GHz_V, and 89.0GHz_V.

[0024] The specific implementation method for obtaining the above-mentioned multi-source surface covariates is as follows:

[0025] The daily land surface temperature (MYD11A1.061) and 16-day synthetic vegetation index (MOD13A2.061) of MODIS were acquired and preprocessed using the Google Earth Engine (GEE) platform.

[0026] Vegetation index information was extracted from NDVI data of MOD13A2.061 product, and then reconstructed using double-logistic function to obtain daily vegetation index time series. Land cover information was extracted from IGBP classification system data of MCD12Q1.061 product, and the original 5m data was aggregated to 1km using mode resampling to extract the dominant land cover attributes within the pixel and match the spatial resolution of land surface temperature.

[0027] The multi-band dual-polarization microwave brightness temperature data specifically utilizes data from the AMSR2 sensor aboard the GCOM-W1 satellite. The AMSR2 multi-band dual-polarization microwave brightness temperature data includes brightness temperatures from four horizontal and vertical polarization channels at 18.7 GHz, 23.8 GHz, 36.5 GHz, and 89.0 GHz. Data acquired daily at 13:30 during transit coincides with the acquisition time of the daily surface temperature (MYD11A1.061). This data has a spatial resolution of 0.1°, approximately 10 km. First, the data underwent a projection transformation to unify it to the WGS84 ellipsoidal geographic coordinate system; subsequently, the spatial resolution was resampled to 1 km.

[0028] The terrain data (DEM terrain data) is derived from the digital elevation model (DEM) data provided by the Advanced Land Observation Satellite (ALOS). First, the DEM data is projected and converted to the WGS84 ellipsoidal geographic coordinate system. Then, the original 30m resolution DEM data is resampled to 1km using cubic convolution interpolation to match the spatial resolution of the MODIS dataset.

[0029] The 2m air temperature and cumulative downward shortwave radiation (CSR) data from the land surface meteorological and hydrological parameter data were obtained from the Long-Term Series High-Resolution Surface Meteorological Element Driving Dataset (TPMFD) of the Third Pole Region. The 2m air temperature and CSR data were extracted for this study, and the daily mean of 2m air temperature and CSR were calculated. The spatial resolution of the calculated daily mean of 2m air temperature and CSR was 1 / 30°. Soil moisture data were obtained from the 1km Seamless Daily Soil Moisture Dataset of the Tibetan Plateau (BTCH). The above data underwent projection transformation and were unified to the WGS84 ellipsoidal geographic coordinate system. The daily mean of 2m air temperature and CSR were resampled to a 1km spatial resolution using bilinear interpolation.

[0030] When using a two-stage reconstruction paradigm that decomposes satellite surface temperature (LST) into periodic and residual components, a hybrid reconstruction framework is needed that couples an enhanced annual temperature cycle (EATC) model with spatial adaptive machine learning (Geographical-XGBoost): the EATC model is used to capture the inherent temporal evolution of LST; and on this basis, the Geographical-XGBoost algorithm is introduced to construct a residual correction model that can sense spatial location.

[0031] Since both the EATC model and the G-XGBoost model require the introduction of covariates, and there can be dozens of covariates or derived variables related to satellite Earth temperature (LST), introducing all covariates or derived variables indiscriminately into the two methods will not only reduce operational efficiency but also increase estimation errors after coupling. Therefore, it is necessary to classify the covariates used as much as possible to avoid or reduce the repeated use of covariates.

[0032] Therefore, the hybrid reconstruction method proposed in this embodiment first classifies the covariates, that is, distinguishes between the trend covariates of the EATC model and the residual covariates used in the G-XGBoost model. The trend covariates are covariates that can represent the periodic variation characteristics of land surface temperature; the residual covariates are covariates that can represent the temporal fluctuations and spatial heterogeneity of land surface temperature.

[0033] Then, the derived variables of the trend covariate and the derived variables of the residual covariate are calculated separately. The trend covariate and its derived variables are constructed into a trend variable set, and the residual covariate and its derived variables are constructed into a residual variable set.

[0034] Step 2: Based on the temporal correlation between each covariate in the multi-source surface covariate and satellite surface temperature, select trend covariates from the multi-source surface covariates to characterize the annual cycle variation of surface temperature, and select residual covariates from the multi-source surface covariates to characterize the spatiotemporal comprehensive distribution characteristics of surface temperature.

[0035] Furthermore, based on the temporal correlation between each covariate and satellite surface temperature, the specific steps for selecting trend covariates include:

[0036] Step 21: Calculate the correlation between each covariate of the multi-source surface covariate and the true value of satellite surface temperature using the Pearson correlation coefficient.

[0037] Step 22: Take the mean absolute value of the correlation coefficient of each covariate across all pixels in the entire study area as the overall correlation strength between the covariate and the surface temperature.

[0038] Step 23: Select the top three covariates with the highest overall correlation as trend covariates. The trend covariates include: 2m air temperature, vegetation index and cumulative downward shortwave radiation.

[0039] Step 24: Take all covariates from the multi-source surface covariates, namely vegetation index, multi-band dual-polarization microwave brightness temperature data, DEM topographic data, 2m air temperature, transit cumulative downward shortwave radiation, and soil moisture, as residual covariates.

[0040] Specifically, in step 21, the method for calculating the correlation between a multi-source surface covariate and satellite surface temperature using the Pearson correlation coefficient is as follows:

[0041] First, a two-dimensional coordinate system is constructed for the entire study area, and the pixel positions within the entire study area are represented in two-dimensional coordinate form. Then, the Pearson correlation coefficient is calculated pixel by pixel for the entire study area. The formula for calculating the Pearson correlation coefficient is as follows:

[0042] ;

[0043] in, is the Pearson correlation coefficient between each multi-source surface covariate and surface temperature in (m,n) pixels of the study area;

[0044] k is the number of times the true value of the land surface temperature is extracted from the (m,n) pixel throughout the year. When the true value of the land surface temperature is extracted from the (m,n) pixel, it is set as a clear sky pixel.

[0045] For the i-th time when a pixel (m,n) is used as a clear sky pixel, the specific value of a certain multi-source surface covariate corresponding to its pixel location;

[0046] The true value of the surface temperature at the location of pixel (m,n) when it is used as a clear sky pixel for the i-th time.

[0047] Let (m,n) be the annual mean of a certain multi-source surface covariate when the (m,n) pixel is used as a clear sky pixel throughout the year;

[0048] Let (m,n) be the annual average of the true value of the land surface temperature corresponding to the clear sky pixel throughout the year;

[0049] Let (m,n) be the standard deviation of a multi-source surface covariate when the pixel is used as a clear sky pixel throughout the year.

[0050] Let be the standard deviation of the true value of the surface temperature when the (m,n) pixel is used as a clear sky pixel throughout the year.

[0051] Given a study area containing M*N pixels, the absolute value of the Pearson correlation coefficient for a certain multi-source surface covariate at each pixel location is taken. The mean of the absolute values ​​of the multi-source surface covariate across the M*N pixels in the study area is calculated, and this mean is used as the correlation between the multi-source surface covariate and surface temperature. The correlation calculation formula is as follows:

[0052] ;

[0053] R represents the correlation between a certain multi-source surface covariate and surface temperature. The larger the R value, the stronger the correlation between the multi-source surface covariate and surface temperature.

[0054] The mean Pearson correlation coefficient between each multi-source surface covariate and surface temperature was obtained in all pixels within the study area. The correlation between each multi-source surface covariate and surface temperature was calculated. Based on the correlation R value, the top three covariates that are strongly correlated with surface temperature were selected as trend covariates. The trend covariates included: 2m air temperature, vegetation index and transit cumulative downward shortwave radiation.

[0055] In the field of remote sensing, the annual temperature cycle (ATC) model is often used to simulate and fill in missing values ​​of land surface temperature (satellite surface temperature (LST)) over time using limited thermal infrared remote sensing data. However, traditional ATC models are usually based on a smooth sine function and cannot effectively capture rapid fluctuations in land surface temperature caused by short-term factors such as weather.

[0056] To address this deficiency, the enhanced EATC model introduces three new key variables: ① Near-surface air temperature (SATs): reflecting the short-term direct impact of weather changes on temperature; ② Satellite vegetation indices (NDVIs): quantifying changes in surface vegetation cover, as vegetation conditions significantly affect surface temperature; ③ Transiting cumulative downward shortwave radiation: after the Earth's surface absorbs solar radiation, its temperature rises, and it releases energy in the form of thermal radiation (longwave radiation), sensible heat, and latent heat. Therefore, transiting cumulative downward shortwave radiation significantly affects surface temperature. Through these improvements, the EATC model can more accurately depict the daily dynamic changes in surface temperature.

[0057] In the two-stage reconstruction of satellite surface temperature (LST), the use of surface covariates still has the following shortcomings: (1) The surface covariates are mainly actual observation results, and their number is limited, which will lead to high overlap of covariates in the two-stage model, weakening the complementary advantages of the two-stage reconstruction; (2) The same surface covariate cannot simultaneously meet the opposing requirements of the two-stage model, namely periodicity and heterogeneity; (3) Surface covariates generally integrate multiple factors, some of which have low correlation with satellite surface temperature (LST) and lack interpretability.

[0058] To address the aforementioned issues, it is necessary to introduce derived variables from surface covariates. This can effectively increase the number of available feature variables in the two-stage model, while maintaining the number of features in the sub-models and reducing the overlap of feature variables between the sub-models.

[0059] Step 3: Calculate the time-series statistical derivatives of the trend covariate using a time-series sliding window, construct a trend variable set from the trend covariate and its time-series statistical derivatives, and calculate the spatial derivatives of the residual covariate using spatial texture analysis and the time-series first derivative method, constructing a residual variable set from the residual covariate and its time and spatial derivatives.

[0060] The calculation of existing derived variables lacks clear physical guidance and uses a relatively small number of derived variables, failing to fully explore the information of covariates regarding the periodicity and heterogeneity of satellite surface temperature (LST). To address the above technical issues, the specific implementation method for calculating time-series statistical derived variables using a time-series sliding window for trend covariates is as follows:

[0061] The derived calculations for each trend covariate are performed using a time-series sliding window to obtain the derived variables for each trend covariate. These derived variables specifically include the mean, maximum, minimum, and linear slope of each trend covariate within the time-series sliding window.

[0062] Suppose that the value of a certain trend covariate of the (m, n) pixels in the study area on day t is... The annual time series data of a certain trend covariate at pixel (m, n) is used as a vector. , where t = (1, ..., 365);

[0063] Assume the time-series sliding window size is 2k+1, {1, 2}; The calculation result Result of the time-series sliding window is represented as:

[0064] ;

[0065] ;

[0066] ;

[0067] in, for One of the functions, For calculating sets The mean, For calculating sets The maximum value, For calculating sets The minimum value, For calculating sets The linear slope is calculated using the least squares method for the point set. Calculate the set of calculations The linear slope;

[0068] The mean, maximum, minimum and linear slope of the covariate within each time-series sliding window are used as the derived variable values ​​of the pixel on the date at the center of the window.

[0069] The trend covariate and its corresponding derived variable together constitute the trend variable set.

[0070] For each trend covariate, each cell in each period's data can be used to calculate four derived variables through a time-series sliding window. The time-series sliding window size can be set to two, so for each trend covariate data, two windows are used to calculate the mean, maximum, minimum, and linear slope.

[0071] Specifically, the size of the time series sliding window is set to 3 and 5 (for example, when the time series sliding window is 3, it means that the calculation involves the previous day, the current day and the next day). When performing time series sliding window calculations at the beginning and end of the year, the covariate data of the previous year and the next year are supplemented to meet the calculation requirements.

[0072] The derived variables of the residual covariates include time-derived variables and space-derived variables. The method for calculating the derived variables of the above residual covariates is as follows:

[0073] The time-derived variables corresponding to all residual covariates other than DEM topographic data were calculated using the first-order time-series method. Specifically, the time-derived variables corresponding to vegetation index, 2m air temperature, cumulative downward shortwave radiation, multi-band dual-polarized microwave brightness temperature data, and soil moisture were calculated.

[0074] Spatial derived variables corresponding to each residual covariate were calculated using spatial texture analysis methods. Specifically, spatial derived variables corresponding to vegetation index, 2m air temperature, cumulative downward shortwave radiation, multi-band dual-polarized microwave brightness temperature data, soil moisture, and DEM topographic data were calculated.

[0075] Spatial derived variables of DEM topographic data are calculated by combining slope and aspect.

[0076] The implementation method for calculating the time-derived variables of vegetation index, 2m air temperature, cumulative downward shortwave radiation, multi-band dual-polarized microwave brightness temperature data, and soil moisture, excluding DEM topographic data, using the time-series first-order derivative method is as follows:

[0077] Alignment and stacking of the time dimension: All multi-period raster data of residual covariates (vegetation index, 2m air temperature, cumulative shortwave radiation, polarization brightness temperature of each frequency band, and soil moisture) are sorted by acquisition time to ensure that each variable has corresponding observations at the same (or very close) time. If the time steps of different variables are inconsistent, time resampling (such as linear interpolation to a unified time point sequence) is required to obtain a common time label [t1, t2, ..., tT]. Microwave brightness temperature is split into multiple independent variables according to frequency band and polarization mode. For each residual covariate X, a three-dimensional array of size M rows × N columns × T time phases is obtained.

[0078] Preprocessing of time series: For the time series of each pixel, if there are missing individual phases, linear interpolation between the preceding and following phases is used to fill in the gaps and make the sequence continuous;

[0079] Calculation of the first-order temporal derivative (calculating the numerical value reflecting the rate of change for each pixel across T time phases): T-1 rate of change layers are generated through time-series differencing, and the formula for calculating the first-order temporal derivative is as follows:

[0080] ;

[0081] In the formula, Let be the time-series first derivative of the residual covariates other than the DEM topographic data in the (m, n) pixels of the entire study area on day t.

[0082] For the (m, n) pixels in the entire study area on day t, the values ​​of other residual covariates excluding the DEM topographic data are given. For the (m, n) pixels in the entire study area on day t+1, the values ​​of other residual covariates excluding the DEM topographic data are given. The value is 1;

[0083] Calculate the time-series first derivatives of the residual covariates (excluding DEM topographic data) in all pixels of the entire study area on day t, using the time-series first derivatives.

[0084] For each original trend covariate, output a set of time-varying layers and incorporate them into the residual feature set according to the time correspondence.

[0085] The method for calculating the spatial derived variables of all residual covariates using the characteristic statistics of the gray-level co-occurrence matrix is ​​as follows:

[0086] For each residual covariate data in the residual variable set, the gray-level co-occurrence matrix and the feature statistics of the window center pixel are calculated sequentially using a spatial sliding window. The window size is 3×3. The feature statistics include: correlation, contrast, entropy, and dissimilarity. When calculating the gray-level co-occurrence matrix and feature statistics of pixels at the boundary of the study area using a spatial sliding window, each residual covariate data at the boundary of the study area is spatially expanded by 8 neighborhoods with an expansion radius of 1.

[0087] The calculation steps of the gray-level co-occurrence matrix are as follows: the pixels within the spatial sliding window are quantized into gray levels, and the floating-point residual covariance is compressed to a finite number of integer gray levels to adapt to the calculation of the gray-level co-occurrence matrix and suppress noise. The number of gray levels is set to 32, and the offset is set to (1,1).

[0088] The feature statistics of the gray-level co-occurrence matrix calculated based on the sliding window are: correlation, contrast, entropy, and dissimilarity. The calculation formulas are as follows:

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] In the formula, This represents the value in the i-th row and j-th column of the gray-level co-occurrence matrix; , These are the row edge distribution and the column edge distribution, respectively. , These are the row mean and column mean, respectively. , These are the row standard deviation and the column standard deviation, respectively. It is a square root function. For correlation, For contrast, Let be entropy, and Dis be dissimilarity.

[0100] For each pixel of each period's data for each residual covariate, four characteristic statistics are calculated through a spatial sliding window. Therefore, for each residual covariate data, four spatial derived variables are calculated: correlation, contrast, entropy, and dissimilarity.

[0101] The method for calculating spatial derived variables of DEM topographic data using slope and aspect is as follows:

[0102] The gradient and aspect of the DEM topographic data for each pixel in the study area were calculated using a 3×3 spatial sliding window, and the slope angle was calculated based on the gradient of the DEM topographic data of the study area.

[0103] The gradient calculation method is as follows: Using the Horn algorithm, for each central pixel, the elevation values ​​within a 3×3 window surrounding it are taken, and the elevation change rates for x and y are calculated separately. Specifically, the gradient is calculated using a weighted average of the elevation differences between adjacent pixels, expressed as a central difference. z / x and z / y;

[0104] Slope angle calculation: Combine the gradients in the x and y directions to obtain the maximum slope angle, slope = arctan Then multiply by the conversion factor to get the slope angle;

[0105] The slope aspect calculation method is as follows: The elevation change rate of each central pixel is calculated based on the Horn algorithm to obtain... z / x and z / y, calculate the aspect angle: aspect = arctan2( z / y - z / x);

[0106] Output the slope raster layer and aspect raster layer corresponding to all pixels in the study area. The slope raster layer of the study area has a value range of 0° to 90°, representing the local surface dip angle, and the aspect raster layer of the study area has a value range of 0° to 360°.

[0107] In other words, the trend variables include 2m air temperature, vegetation index, and cumulative downward shortwave radiation, as well as derived variables. If all variables are introduced without selection, some variables may be highly correlated (e.g., 2m air temperature and cumulative downward shortwave radiation are highly coupled on a seasonal scale), leading to multicollinearity. That is, too many covariates may still cause the periodic component analysis model to overfit clear-sky data, reducing its ability to generalize to missing values ​​below clouds.

[0108] The residual variable set contains all multi-source surface covariates and their derived variables, such as vegetation index, microwave brightness temperature, cumulative downward shortwave radiation, soil moisture, as well as time derivatives (such as the first derivative of the vegetation index) and spatial texture features (such as GLCM contrast and entropy of DEM topographic data). These variables are highly correlated (for example, the vegetation index and soil moisture change together in space). If they are not filtered, a lot of redundant information will be introduced, which will reduce the efficiency of model training and may lead to overfitting.

[0109] Therefore, although the existing two-stage reconstruction studies of LST have introduced derived variables, the following problems still exist: (1) There is no clear method for selecting application parameters of covariates and derived variables in the two-stage reconstruction model of LST, and the subjectivity is strong; (2) In the two-stage reconstruction model of LST, covariates and derived variables have a strong correlation, which increases the risk of overfitting and reduces the generalization ability of the model. In order to solve the above problems, this implementation provides the following step 4.

[0110] Step 4: Perform feature filtering on the trend variable set and the residual variable set respectively to generate optimized trend variable subsets and optimized residual variable subsets.

[0111] In step 4, the feature selection is performed on the trend variable set and the residual variable set respectively:

[0112] Normalize each covariate and derived variable in the trend variable set and the residual variable set respectively, so that each variable is normalized to [0,1].

[0113] A stability selection algorithm based on elastic networks is used to screen features from the trend variable set, selecting 3 to 5 feature variables to form an optimized trend variable subset;

[0114] A recursive feature elimination algorithm is used to screen features in the residual variable set. An optimized subset of residual variables is selected from the residual variable set, and 8 to 10 features are selected from the residual covariates to form the optimized subset of residual variables.

[0115] The specific implementation method is as follows:

[0116] A stability selection algorithm is used to iteratively screen the covariates and their derived variables in the trend variable set. The importance of the variables is ranked based on the selection frequency. The variables selected according to importance are used to construct an optimized trend variable subset. Among them, a certain number (3-5) of variables are retained to construct the optimized trend variable subset. That is, in this implementation, the first three covariates are selected from the multi-source surface covariates as trend covariates based on the Pearson correlation coefficient, namely: 2m air temperature, vegetation index and cumulative downwave radiation (CSR).

[0117] The three covariates mentioned above are placed into the trend variable set, and time series analysis is used to perform derivative calculations on them. These derivative variables are calculated using a sliding window and include: multi-day averages, multi-day maximum / minimum values, and multi-day linear slopes. The multi-day calculations utilize two time spans: 3 days and 5 days. For the sliding window calculations at the beginning and end of the year, further supplementary covariate data from the previous and following years are needed. The total number of derived variables for the three covariates is 3 * 4 * 2 = 24, such as the 5-day average of 2-meter temperature and the 3-day linear slope of NDVI. The covariates and derived variables together constitute the trend variable set of 27.

[0118] Since the derived variables have strong autocorrelation, a stability selection algorithm and selected frequency are used for feature selection. After screening 21 features, the 3-day average of 2m temperature, vegetation index, CSR and the 3-day linear slope of CSR are selected. The above four covariates and the covariate derivatives together constitute the optimized trend variable subset.

[0119] The covariates and their derived variables in the residual variable set are iteratively screened based on the recursive feature elimination algorithm, and the variables selected based on feature importance ranking are used to construct an optimized subset of residual variables.

[0120] The derived variables in the residual variable set are mainly used to extract the temporal fluctuations and spatial heterogeneity of covariates, including: the first derivative of the time series, the correlation, contrast, entropy, and dissimilarity of the spatial 3*3 window, and the spatial derived variables of DEM slope and aspect. The residual variable set contains a total of 80 features. A recursive feature elimination algorithm and feature importance were used for feature selection, and the optimal subset of residual variables selected consisted of 8 features.

[0121] This implementation proposes to iteratively screen the trend variable set and the residual variable set, construct an optimized subset of trend variables and an optimized subset of residual variables, and retain the most representative variables of LST periodic changes and spatial heterogeneity, respectively, to ensure that the two-stage reconstruction process focuses on the main driving factors and significantly reduces training time.

[0122] Therefore, in the two-stage reconstruction of land surface temperature, new derived variables are obtained by mathematical transformation, spatiotemporal aggregation, or physical model calculation of covariates. Introducing these derived variables into the two-stage reconstruction of land surface temperature can effectively increase the number of available feature variables in the two-stage model. While maintaining the number of features in the sub-models, it reduces the overlap of feature variables between the sub-models. On the other hand, it can extract information from the surface covariates that is more relevant to land surface temperature, including the periodic and heterogeneous characteristics of land surface temperature, while meeting the feature variable requirements of the two-stage model and increasing the interpretability of the model.

[0123] like Figure 2As shown, the third column, Covariate-only, represents the surface temperature reconstruction image generated when only covariates are used, and the fourth column, Reference LST, represents the surface temperature reconstruction reference image generated when both covariates and derived variables are used. The red circles indicate the differences in surface temperature reconstruction between the two methods: texture or color.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. All such modifications or substitutions should be covered within the protection scope of this application, and should not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for deriving and screening surface covariates for surface temperature reconstruction, characterized in that, Includes the following steps: Step 1: Based on satellite imagery and Earth observation data of the study area, obtain multi-source surface covariates for all pixels in the study area. The multi-source surface covariates are various environmental factors that affect the spatiotemporal distribution characteristics of surface temperature. Step 2: Based on the temporal correlation between each covariate in the multi-source surface covariate and satellite surface temperature, select trend covariates from the multi-source surface covariates to characterize the annual cycle variation of surface temperature, and select residual covariates from the multi-source surface covariates to characterize the spatiotemporal comprehensive distribution characteristics of surface temperature. Step 3: Calculate the time-series statistical derived variables of the trend covariate using a time-series sliding window, construct a trend variable set from the trend covariate and its time-series statistical derived variables, and calculate the spatial derived variables of the residual covariate using the spatial texture analysis method and the time-series first derivative method, construct a residual variable set from the residual covariate and its time and spatial derived variables. Step 4: Perform feature filtering on the trend variable set and the residual variable set respectively to generate optimized trend variable subsets and optimized residual variable subsets.

2. The method according to claim 1, characterized in that, In step 1, the multi-source surface covariates specifically include vegetation index, multi-band dual-polarization microwave brightness temperature, DEM topographic data, 2m air temperature, transit cumulative downward shortwave radiation, and soil moisture. The multi-band dual-polarization microwave brightness temperature data consists of eight types of data, namely, H and V polarization data at four frequencies: 18.7GHz, 23.8GHz, 36.5GHz, and 89.0GHz, which can be represented as: 18.7GHz_H, 23.8GHz_H, 36.5GHz_H, 89.0GHz_H, 18.7GHz_V, 23.8GHz_V, 36.5GHz_V, and 89.0GHz_V.

3. The method according to claim 1, characterized in that, In step 1, the correlation between each covariate of the multi-source surface covariate and satellite surface temperature is calculated using the Pearson correlation coefficient. The top three covariates that are strongly correlated with satellite surface temperature are selected from the multi-source surface covariates as trend covariates. The trend covariates include: 2m air temperature, vegetation index and transit cumulative downward shortwave radiation. All covariates from the multi-source surface covariates, namely vegetation index, multi-band dual-polarization microwave brightness temperature data, DEM topographic data, 2m air temperature, transit cumulative downward shortwave radiation, and soil moisture, are used as residual covariates.

4. The method according to claim 3, characterized in that, In step 2, the method for calculating the correlation between a multi-source surface covariate and satellite surface temperature using the Pearson correlation coefficient is as follows: First, a two-dimensional coordinate system is constructed for the entire study area, and the pixel positions within the entire study area are represented in two-dimensional coordinate form. Then, the Pearson correlation coefficient is calculated pixel by pixel for the entire study area. The formula for calculating the Pearson correlation coefficient is as follows: ； in, is the Pearson correlation coefficient between each multi-source surface covariate and surface temperature in (m,n) pixels of the study area; k represents the number of times clear sky data is detected for pixel (m,n) throughout the year. When clear sky data is detected for pixel (m,n), it is set as a clear sky pixel. For the i-th time when a pixel (m,n) is used as a clear sky pixel, the specific value of a certain multi-source surface covariate corresponding to its pixel location; The true value of the surface temperature at the location of pixel (m,n) when it is used as a clear sky pixel for the i-th time. Let (m,n) be the annual mean of a certain multi-source surface covariate when the (m,n) pixel is used as a clear sky pixel throughout the year; Let (m,n) be the annual average of the true value of the land surface temperature corresponding to the clear sky pixel throughout the year; Let (m,n) be the standard deviation of a multi-source surface covariate when the pixel is used as a clear sky pixel throughout the year. Let (m,n) be the standard deviation of the true value of the land surface temperature when the (m,n) pixel is used as a clear sky pixel throughout the year; Given a study area containing M*N pixels, the absolute value of the Pearson correlation coefficient for a certain multi-source surface covariate at each pixel location is taken. The mean of the absolute values ​​of the multi-source surface covariate across the M*N pixels in the study area is calculated, and this mean is used as the correlation between the multi-source surface covariate and surface temperature. The correlation calculation formula is as follows: ； R represents the correlation between a certain multi-source surface covariate and surface temperature. The larger the R value, the stronger the correlation between the multi-source surface covariate and surface temperature.

5. The method according to claim 4, characterized in that, The Pearson correlation coefficient between each multi-source surface covariate and surface temperature in each pixel of the study area was obtained sequentially, and the correlation between each multi-source surface covariate and surface temperature was calculated. The larger the correlation R value, the stronger the correlation. The top three covariates that are strongly correlated with surface temperature are selected as trend covariates. The trend covariates include: 2m air temperature, vegetation index and cumulative downward shortwave radiation.

6. The method according to claim 1, characterized in that, The method for calculating the derived variables of the trend covariates and constructing the trend variable set is as follows: In step 3, the derived calculations of each trend covariate are performed through a time series sliding window to obtain the derived variables of each trend covariate. The derived variables specifically include the mean, maximum value, minimum value, and linear slope of each trend covariate within the time series sliding window. Suppose that the value of a certain trend covariate of the (m, n) pixels in the study area on day t is... The annual time series data of a certain trend covariate at pixel (m, n) is used as a vector. , where t=(1,...,365); Assume the time-series sliding window size is 2k+1, {1,2}; The result of the time-series sliding window calculation is represented as: ； ； ； in, for One of the functions, For calculating sets The mean, For calculating sets The maximum value, For calculating sets The minimum value, For calculating sets The linear slope, The calculation specifically involves using the least squares method to calculate the point set. The linear slope; Each trend covariate and its corresponding derived variables together constitute the trend variable set.

7. The method according to claim 1, characterized in that, In step 3, the derived variables of the residual covariates include time-derived variables and spatial-derived variables. Specifically, the time-derived variables corresponding to vegetation index, 2m air temperature, cumulative downward shortwave radiation, multi-band dual-polarization microwave brightness temperature data, and soil moisture, excluding DEM topographic data, are calculated using the first-order time derivative. Spatial derived variables corresponding to vegetation index, 2m air temperature, cumulative downward shortwave radiation, multi-band dual-polarized microwave brightness temperature data, soil moisture, and DEM topographic data were calculated using spatial texture analysis methods. And combine slope and aspect to calculate spatial derived variables of DEM topographic data; The formula for calculating the first-order derivative of the time series is as follows: ； In the formula, Let be the time-series first derivative of the residual covariates other than the DEM topographic data in the (m, n) pixels of the entire study area on day t. For the (m, n) pixels in the entire study area on day t, the values ​​of other residual covariates excluding the DEM topographic data are given. For the (m, n) pixels of the entire study area on day t+1, excluding the DEM topographic data, the values ​​of other residual covariates are given. When t=365, The values ​​of the other residual covariates besides the DEM topographic data for the first day of the following year. The value is 1; Calculate the time-series first derivatives of the residual covariates (excluding DEM topographic data) in all pixels of the entire study area on day t, using the time-series first derivatives.

8. The method according to claim 7, characterized in that, The spatial texture analysis method specifically utilizes the gray-level co-occurrence matrix characteristic statistics. The implementation method for calculating the spatially derived variables of all residual covariates using these statistics is as follows: For each residual covariate data in the residual variable set, the gray-level co-occurrence matrix and the feature statistics of the window center pixel are calculated sequentially using a spatial sliding window. The window size is 3×3, and the feature statistics include: correlation, contrast, entropy, and dissimilarity. The specific calculation formula is as follows: ； ； ； ； ； ； ； ； ； ； In the formula, This represents the value in the i-th row and j-th column of the gray-level co-occurrence matrix; , These are the row edge distribution and the column edge distribution, respectively. , These are the row mean and column mean, respectively. , These are the row standard deviation and the column standard deviation, respectively. It is a square root function. For correlation, For contrast, Let be entropy, and Dis be dissimilarity. For each pixel of each period's data for each residual covariate, four characteristic statistics are calculated through a spatial sliding window. Therefore, for each residual covariate data, four spatial derived variables are calculated: correlation, contrast, entropy, and dissimilarity.

9. The method according to claim 7, characterized in that, The method for calculating spatial derived variables of DEM topographic data using slope and aspect is as follows: The Horn algorithm was used to calculate the elevation change rate of each pixel in the study area in the north-south and east-west directions using a 3×3 window. The maximum slope angle is synthesized based on the elevation change rate and converted into a slope value to generate a slope raster layer. The slope aspect angle is calculated based on the elevation change rate, and a slope aspect raster layer is generated.

10. The method according to claim 1, characterized in that, In step 4, the step of performing feature filtering on the trend variable set and the residual variable set respectively: Normalize each covariate and derived variable in the trend variable set and the residual variable set respectively, so that each variable is normalized to [0,1]. The trend variable set is subjected to feature selection using a stability selection algorithm based on elastic networks, and 3 to 5 feature variables are selected from the trend variable set to form an optimized trend variable subset; The recursive feature elimination algorithm is used to screen features in the residual variable set to select an optimized subset of residual variables. 8 to 10 features are selected from the residual covariates to form the optimized subset of residual variables.

Images