Three-stage high-precision near-surface air temperature remote sensing estimation method

Abstract

The invention discloses a three-stage high-precision near-surface air temperature remote sensing estimation method, which comprises the following steps of: performing time normalization on remote sensing surface temperature data to eliminate imaging time difference; dividing the research area into a plurality of sub-areas according to natural condition difference clustering, and estimating air temperature in each sub-area by using four machine learning models based on the surface temperature data after time normalization and space auxiliary variables to obtain a spatialized air temperature estimation result; and temperature estimation results of four machine learning models are integrated through a generalized additive model, so that the near-surface temperature with higher precision can be obtained. According to the method, the uncertainty caused by the time difference of the surface temperature is effectively eliminated, the model adaptability difference in different geographical environments is also considered, multiple single machine learning models are integrated based on the integration thought, and the remote sensing estimation precision of the air temperature is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of thermal infrared remote sensing, and in particular relates to a three-stage high-precision near-surface air temperature remote sensing estimation method. Background Art

[0002] Near-surface temperature is a key meteorological parameter that reflects the thermal environment and plays an important role in many fields, such as climate change monitoring, urban heat island effect research, public health, and agricultural production. Near-surface temperature data with high temporal and spatial resolution are essential for better understanding and simulating complex land-atmosphere processes. Although traditional ground-based observations can obtain accurate near-surface temperature observation data at a point scale, they have defects such as limited quantity and uneven spatial distribution, and cannot well describe the detailed changes in temperature over a large area.

[0003] The development of satellite remote sensing has provided a new means for thermal environment monitoring. Satellite thermal infrared remote sensing has been widely used in the estimation of near-surface air temperature due to its advantages of large-area and long-term continuous observation. There are four methods for estimating near-surface air temperature using remote sensing data: temperature-vegetation index method, energy balance method, statistical method and machine learning method. The temperature-vegetation index method estimates the air temperature based on the assumption that the air temperature in the vegetation canopy is equal to the surface temperature of the canopy. This method requires very few input parameters, all of which can be obtained through remote sensing, but it is not suitable for areas with sparse vegetation. The energy balance method estimates the relationship between the ground and air temperature based on the surface energy balance equation and then calculates the air temperature from the remotely sensed surface temperature. The method has clear physical principles and high portability and versatility, but some variables in the surface energy balance equation, such as dynamic impedance and surface roughness, cannot be obtained through remote sensing, which limits its large-scale application. The statistical method estimates the air temperature by establishing a regression equation between air temperature and environmental factors such as surface temperature. This method is easy to implement. For example, the patent with publication number CN113255148A discloses a method for estimating all-weather temperature and its spatiotemporal distribution based on MODIS product data, and the patent with publication number CN112016052B discloses a method, system and terminal for estimating the near-surface daily maximum temperature based on multi-source data. However, the regression model often cannot fit the complex relationship between variables. The machine learning method estimates the temperature based on a nonlinear machine learning model. Compared with the traditional regression method, the machine learning method can effectively handle complex nonlinear relationships and obtain higher accuracy. It is currently the most widely used temperature estimation method.

[0004] However, existing methods still ignore several issues: 1) Due to the inherent observation characteristics and wide field of view of polar-orbiting satellite sensors, the surface temperature products obtained from these sensors may have a time difference of up to two hours between different pixels on the same image, or between the same pixel on different dates within a revisit cycle. This time difference will bring great uncertainty when using remote sensing surface temperature products to estimate near-surface air temperature in large-scale regions or long-term studies. 2) Most existing studies develop a unified model for the entire region to estimate near-surface air temperature. However, due to the diversity of natural conditions in large-scale regions and the differences in the relationship between near-surface air temperature and environmental factors, the unified model cannot reflect the spatial heterogeneity in temperature estimation, which will bring bias in near-surface air temperature estimation. 3) Existing studies only use a single machine learning algorithm to estimate near-surface air temperature, or compare multiple machine learning algorithms and select the best performing model for temperature estimation. However, a single model usually has limitations and uncertainties because the performance of different algorithms varies due to changes in geographical location and environmental conditions. Summary of the invention

[0005] In response to the above-mentioned problems, the present invention proposes a three-stage high-precision near-surface temperature remote sensing estimation method, which not only effectively eliminates the uncertainty caused by the time difference of surface temperature, but also takes into account the differences in model adaptability in different geographical environments. It integrates multiple single machine learning models based on an integrated approach, thereby improving the accuracy of remote sensing estimation of temperature.

[0006] The present invention discloses a three-stage high-precision near-surface air temperature remote sensing estimation method, the method comprising the following steps: S1, time normalization of remote sensing surface temperature data to eliminate imaging time differences; S2, the study area is clustered into multiple sub-areas according to differences in natural conditions. Based on the time-normalized surface temperature data and the spatial auxiliary variables corresponding to the meteorological stations in the sub-areas, four machine learning models are used to estimate the temperature in each sub-area to obtain spatial temperature estimation results; S3, integrates the temperature estimation results of four machine learning models through a generalized additive model to obtain a more accurate near-surface temperature.

[0007] Step S1 further comprises the following steps: The surface temperature normalization formula is used to perform time normalization on the surface temperature data collected by the MODIS remote sensing sensor: ; Where, T norm is the surface temperature normalized to the reference time, T ac is the actual surface temperature during the solar period, a and b are coefficients, t acis solar time, t norm is the base time; Hourly surface temperature was extracted based on ERA5_land reanalysis meteorological data, and masked with total cloud cover information of ERA5. Hourly monthly composite processing was performed on the surface temperature in the clear sky phase to obtain the spatiotemporally continuous monthly composite hourly ERA5_Land surface temperature data. The following formula was used to convert the UTC coordinated universal time corresponding to the ERA5_Land surface temperature data into solar time: ; In the formula, t solar_time is solar time, t UTC For Coordinated Universal Time, is the time difference between the standard time of the time zone where the study area is located and UTC, and Lon is the longitude value of the pixel; Based on the monthly composite hourly ERA5_Land surface temperature data and the corresponding solar time, the least squares fitting was performed pixel by pixel based on the surface temperature normalization formula to obtain the coefficients a and b values, and then resampled to 1 km resolution to be consistent with the resolution of MODIS surface temperature data; the surface temperature time normalization formula was applied pixel by pixel to perform time normalization on the MODIS surface temperature data to obtain the surface temperature data normalized to 11 o'clock.

[0008] Furthermore, in step S2, the process of clustering the study area into multiple sub-areas according to differences in natural conditions includes the following steps: The MODIS reflectivity product MOD09A1, MODIS albedo product MCD43A3, ERA5-Land and SRTM / DEM data were resampled to 1km resolution to be consistent with the MODIS surface temperature data resolution. The reflectivity of the red band, near infrared band, green band and shortwave infrared 1 band were extracted from the preprocessed MODIS reflectivity product MOD09A1; the albedo data and longitude and latitude data were extracted from the preprocessed MODIS albedo product MCD43A3; the solar downward radiation DSR was extracted from the preprocessed ERA5-Land data; the elevation and terrain index CTI were extracted from the preprocessed SRTM / DEM data.

[0009] The vegetation index NDVI and water body index MNDWI at a resolution of 500 m are calculated using the following formulas: ; ; In the formula, ρ NIR and ρ RED Represent the reflectance of near infrared band and red band respectively; ρ Greenis the green band reflectivity, ρ SWIR1 It is the reflectivity of shortwave infrared 1 band.

[0010] Furthermore, the albedo, elevation, downward solar radiation DSR, vegetation index NDVI and water body index MNDWI were annually averaged and normalized to eliminate the influence of value range differences in the classification process. Longitude and latitude were used as classification features, and the iterative self-organizing data analysis model was used for clustering and partitioning. The study area was divided into several sub-areas through iteration.

[0011] Furthermore, in step S2, the process of estimating the temperature in each sub-area using four machine learning models includes the following steps: Through the spatiotemporal matching of longitude, latitude and time, all spatial variables corresponding to the meteorological stations are extracted. The spatial variables include normalized surface temperature and spatial auxiliary variables including vegetation index NDVI, water body index MNDWI, albedo, elevation, downward solar radiation DSR, topographic index CTI, annual accumulated day DOY, longitude and latitude. Taking all spatial variables as independent variables and the measured temperature values ​​at meteorological stations as dependent variables, four machine learning methods, namely random forest, extreme gradient boosting, classification enhancement and Cubist, were used to construct temperature estimation models in each sub-area: ; In the formula, represents the estimated temperature at station i; represents the spatial coordinates at site i; is the intercept of the equation; is the regression coefficient of the corresponding independent variable; They are the normalized surface temperature, vegetation index NDVI, water body index MNDWI, albedo, elevation, downward solar radiation DSR, topographic index CTI, annual cumulative day DOY, longitude and latitude at station i respectively; is the random error term; The four constructed machine learning models were applied to the time-normalized surface temperature and spatial auxiliary variables including vegetation index NDVI, water body index MNDWI, albedo, elevation, downward solar radiation DSR, terrain index CTI, annual accumulated day DOY, longitude and latitude to obtain the spatial temperature estimation results.

[0012] Furthermore, in step S3, the temperature estimation results obtained by each machine learning model in step S2 are used as independent variables, and the measured temperature at the meteorological station is used as the dependent variable, and the generalized additive model GAM is used to construct an integrated model: ; In the formula, Ta To observe the near-surface air temperature; Ta model 1 、 model 2 、 model 3 、 model 4 are the temperature estimates of the four machine learning models: random forest, extreme gradient boosting, classification enhancement, and Cubist; S() represents the smooth function of the corresponding parameters; e i,j represents the residual error term for each observation; The constructed integrated model is applied to the estimated temperatures obtained by the four machine learning models to obtain the final integrated temperature estimate.

[0013] The advantages and beneficial effects of the present invention are: First, the three-stage high-precision near-surface air temperature remote sensing estimation method of the present invention time-normalizes the remote sensing surface temperature data based on reanalysis data, effectively eliminating the uncertainty caused by observation time differences.

[0014] Second, the three-stage high-precision near-surface temperature remote sensing estimation method of the present invention divides the study area into small areas based on the spatial clustering method, and constructs temperature estimation models in these smaller areas with more consistent environments. The adaptability of the model in different geographical environments is reduced through zoning modeling.

[0015] Third, the three-stage high-precision near-surface air temperature remote sensing estimation method of the present invention constructs an integrated model to integrate the estimation results of multiple single machine learning models. Through integration, the estimation results of various basic models complement each other, effectively reducing the deviation of a single model and generating higher-precision estimation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flow chart of the three-stage high-precision near-surface temperature remote sensing estimation method of the present invention; Figure 2 Comparison of MODIS LST before and after normalization on February 21, 2021: (a) is the schematic diagram of the original LST, and (b) is the schematic diagram of the LST after time normalization; Figure 3 This is a schematic diagram of the partition results; Figure 4a Scatter plot of the temperature estimated by the Cubist model and the temperature measured at the station; Figure 4bScatter plot of the temperature estimated by the Catboost model and the temperature measured at the station; Figure 4c Scatter plot of the temperature estimated by the XGboost model and the temperature measured at the site; Figure 4d Scatter plot of the temperature estimated by the RF model and the temperature measured at the station; Figure 5 Scatter plot of the final estimated temperature and the measured temperature at the station; Figure 6 This is the spatial distribution map of the average temperature in 2021 obtained by remote sensing estimation. DETAILED DESCRIPTION

[0017] The following examples will enable those skilled in the art to more fully understand the present invention, but are not intended to limit the present invention in any way.

[0018] The purpose of the present invention is to propose a three-stage high-precision near-surface temperature remote sensing estimation method, which includes three stages: in the first stage, time normalization processing is performed on the remote sensing surface temperature data to eliminate the uncertainty caused by the observation time difference; in the second stage, the spatial heterogeneity of the geographical environment within a large region is considered to divide it into multiple sub-regions with independent characteristics, and four machine learning models are used to estimate the temperature in each sub-region; in the third stage, the temperature estimation results of four single machine learning algorithms are integrated based on a generalized additive model to obtain high-precision near-surface temperature.

[0019] 1. Time normalization of remotely sensed surface temperature Since the MODIS remote sensing sensor has a swath width of more than 2000 km and an orbit repetition period of 16 days, the observed solar time within the same scanning strip and between adjacent scanning strips varies greatly. In a 16-day revisit period, the observed solar time of the same pixel on different dates also varies greatly, with the maximum time difference being up to two hours. Considering that the surface temperature is greatly affected by solar time, there will be a large uncertainty in directly using the original remote sensing surface temperature to estimate the near-surface temperature.

[0020] In order to eliminate the influence of different solar observation times in MODIS surface temperature data, it is necessary to perform time normalization processing. The time normalization equation for surface temperature is: (1); Where, T norm is the surface temperature normalized to the reference time, T ac is the actual surface temperature during the solar period, a and b are coefficients, t ac is solar time, t norm The Terra / MODIS daytime imaging time range is 9.9 ~ 12.8 (solar time), so tnorm Set to 11.0.

[0021] The coefficients a and b in the time normalization equation are related to the geographical location and environment, and need to be solved pixel by pixel. Hourly surface temperature is extracted based on ERA5_land reanalysis meteorological data, masked with ERA5 total cloud information, and then the clear sky phase surface temperature is taken for hourly monthly synthesis to obtain the spatiotemporally continuous monthly synthetic hourly ERA5_Land surface temperature data. The time corresponding to the ERA5_Land surface temperature data is UTC Coordinated Universal Time, which needs to be converted to solar time. The conversion formula is as follows: (2); In the formula, t solar_time is solar time, t UTC For Coordinated Universal Time, is the time difference between the standard time of the time zone where the study area is located and UTC, Tzone_offset in China is 8, and Lon is the longitude value of the pixel.

[0022] Based on the monthly synthetic hourly ERA5_Land surface temperature data and the corresponding solar time, the coefficients a and b are obtained by least squares fitting based on the form of formula (1) for each pixel, and resampled to 1 km to be consistent with the resolution of MODIS surface temperature data. Then, the MODIS temperature product data is time-normalized by applying formula (1) pixel by pixel to obtain the surface temperature data normalized to 11 o'clock.

[0023] 2. Temperature estimation based on regional modeling The spatial auxiliary variables used in the present invention are vegetation index NDVI, water body index MNDWI, albedo, elevation, solar downward radiation DSR, terrain index CTI, annual accumulation of days DOY, longitude and latitude. MODIS reflectivity product MOD09A1, MODIS albedo product MCD43A3, MODIS albedo product MCD43A3, ERA5-Land and SRTM / DEM data are resampled to a resolution of 1 km to be consistent with the resolution of MODIS surface temperature data. Extract the reflectance of the red band, near infrared band, green band and shortwave infrared 1 band from the MODIS reflectance product MOD09A1, and use formula (3) and formula (4) to calculate NDVI and MNDWI respectively; extract albedo data and longitude and latitude data from the MODIS albedo product MCD43A3; extract downlink shortwave radiation DSR from ERA5-Land; obtain elevation data and CTI from SRTM / DEM. Among them, the calculation formulas of NDVI and MNDWI are: (3); (4); In the formula, ρ NIR and ρ RED Represent the reflectance of near infrared band and red band respectively, ρ Green is the green band reflectivity, ρ SWIR1 It is the reflectivity of shortwave infrared 1 band.

[0024] Albedo, elevation, DSR, NDVI and MNDWI were annually averaged and normalized to eliminate the influence of value range differences in the classification process. Longitude and Latitude were used as classification features, and the iterative self-organizing data analysis model (ISODATA) was used for clustering and partitioning. The minimum and maximum values ​​of the number of categories in the clustering process were set to 5 and 12 respectively. The study area was divided into several sub-areas through iteration.

[0025] By matching longitude, latitude and time, we extracted the normalized surface temperature, NDVI, MNDWI, Albedo, elevation, DSR, CTI, DOY, Longitude and Latitude corresponding to the meteorological station. With these spatial variables as independent variables and the measured temperature value of the meteorological station as the dependent variable, we used four machine learning methods, namely, random forest (RF), extreme gradient boosting (XGBoost), classification boosting (CatBoost) and Cubist, to build temperature estimation models in each sub-area: (5); In the formula, represents the estimated temperature at station i; represents the spatial coordinates at site i; is the intercept of the equation; is the regression coefficient of the corresponding independent variable; are the normalized surface temperature, NDVI, MNDWI, Albedo, elevation, DSR, CTI, DOY, Longitude and Latitude at site i, respectively; is a random error term.

[0026] The four constructed machine learning models were applied to spatial independent variables such as surface temperature, NDVI, MNDWI, Albedo, elevation, DSR, CTI, DOY, Longitude and Latitude after time normalization to obtain spatial temperature estimation results.

[0027] 3. Temperature estimation based on integrated model The estimated temperature obtained by each machine learning model in the second stage is used as the independent variable, and the measured temperature at the meteorological station is used as the dependent variable. The generalized additive model GAM is used to construct an integrated model: (6); In the formula, Ta To observe the near-surface air temperature; Ta model 1 、 model 2 、 model 3 、 model 4 are the estimated values ​​of temperature by Cubist, XGBoost, Catboost and RF models respectively; S() represents the smooth function of the corresponding parameters; e i,j Represents the residual error term for each observation.

[0028] The constructed integrated model was applied to the estimated temperatures obtained by the four machine learning methods to obtain the final integrated temperature estimate.

[0029] Examples The present invention is to estimate the near-surface temperature with high precision based on remote sensing data. The following are the specific implementation steps of the example. Figure 1 .

[0030] 1) The area between 95°-110°E and 29°-37°N was taken as the study area. The surface temperature and cloud cover information were extracted from the ERA5_Land and ERA5 data, respectively. Cloud mask processing and monthly synthesis processing were performed to generate the temporally and spatially continuous monthly hourly ERA5_Land surface temperature data. The UTC time corresponding to the ERA5_Land surface temperature was converted to solar time using formula (2). Then, the coefficients a and b in formula (1) were obtained by least squares fitting pixel by pixel, and resampled to a resolution of 1 km to be consistent with the MODIS surface temperature.

[0031] 2) The clear sky land surface temperature data and the corresponding solar time were extracted from the MODIS land surface temperature product MOD11A1, and the MODIS land surface temperature values ​​were time normalized for each pixel using formula (1). Figure 2 A comparison of MODIS surface temperature before and after time normalization on February 21, 2021 is given. It can be seen that before time normalization (see Figure 2 In (a) of the figure, the surface temperature at the edge of the scanning strip is obviously inconsistent due to the difference in satellite transit time. After time normalization correction (see Figure 2 (b) in the figure significantly reduces the inconsistency, and the spatial distribution of surface temperature is more continuous and uniform, effectively eliminating the temperature deviation caused by different imaging of the sun.

[0032] 3) Resample the MODIS reflectance product MOD09A1, MODIS albedo product MCD43A3, ERA5-Land and SRTM / DEM data to 1 km resolution to match the MODIS surface temperature data resolution. Extract the red band, near-red band, green band and shortwave infrared 1 band reflectance from the MODIS surface reflectance product MOD09A1, and calculate the vegetation index NDVI and water body index MNDWI using formula (3) and formula (4) respectively; extract the albedo and the corresponding longitude and latitude from the MODIS albedo product MCD43A3; extract the downlink shortwave radiation DSR from ERA5-Land; extract the elevation and CTI data from SRTM / DEM.

[0033] 4) Annual average processing of Albedo, DEM, DSR, NDVI, and MNDWI was performed, and then used together with Longitude and Latitude as classification features, and the ISODATA iterative self-organizing method was used for classification to divide the study area into several sub-areas (see Figure 3 ). In the ISODATA clustering process, the minimum and maximum number of categories are set to 5 and 12 respectively.

[0034] 5) Through the spatiotemporal matching of location, time and region, the normalized surface temperature, NDVI, MNDWI, Albedo, DSR, CTI, elevation, Longitude, Latitude and DOY values ​​corresponding to the station observed temperature were extracted to construct the temperature estimation sample set. The measured temperature at the station was taken as the dependent variable, and the normalized surface temperature, NDVI, MNDWI, Albedo, DSR, CTI, elevation, Longitude, Latitude and DOY of the corresponding pixels were taken as the independent variables. The near-surface temperature estimation model was constructed in each sub-area using Cubist, XGBoost, RF and Catboost methods (see formula (5)). In order to optimize the estimation of the model, cross-validation was used to adjust the model parameter values ​​during the fitting process. Figures 4a to 4d A scatter plot of the temperature estimation accuracy of four machine learning models is given.

[0035] 6) The four machine learning models constructed for each sub-area were applied to the normalized surface temperature, NDVI, MNDWI, Albedo, DSR, CTI, elevation, Longitude, Latitude and DOY spatial variables to calculate the estimated temperature of each sub-area. The estimated temperatures of each sub-area obtained by the four machine learning models were mosaicked to obtain the estimated temperatures of the four machine learning models for the entire study area.

[0036] 7) The estimated temperatures of the four machine learning models for the corresponding pixels of each meteorological station were extracted respectively. The measured temperature of the station was taken as the dependent variable and the estimated temperatures of the four machine learning models for the corresponding pixels were taken as the independent variable. The generalized additive model was used to construct an integrated model for estimating the temperature using the four machine learning methods (see formula (6)).

[0037] 8) Apply the constructed integrated model to the four machine learning models obtained in step 7 to estimate the temperature, and calculate the final remote sensing estimation results of the near-surface temperature in the study area. Figure 5 The verification scatter plot of the integrated results is given. The sample points are mainly distributed near the 1:1 line. The determination coefficient R 2 The average absolute error (MAE) is 0.99, the bias is approximately 0, indicating that there is no obvious overestimation or underestimation. Compared with the four single machine learning methods, the temperature estimation accuracy of the integrated method is significantly improved. Figure 6 The spatial distribution map of the average temperature in the study area in 2021 obtained by applying the method of the present invention is given. The annual average temperature is mainly between -8 and 27°C, showing significant spatial differences. The temperature in the southeast region is higher, usually between 15°C and 25°C; the temperature in the northwest region is lower, generally below 0°C.

[0038] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A three-stage high-precision near-surface temperature remote sensing estimation method, characterized in that: The method comprises the following steps: S1, time normalization of remote sensing surface temperature data to eliminate imaging time differences; S2, the study area is clustered into multiple sub-areas according to differences in natural conditions. Based on the time-normalized surface temperature data and the spatial auxiliary variables corresponding to the meteorological stations in the sub-areas, four machine learning models are used to estimate the temperature in each sub-area to obtain spatial temperature estimation results; S3, integrates the temperature estimation results of four machine learning models through a generalized additive model to obtain a more accurate near-surface temperature.

2. The three-stage high-precision near-surface air temperature remote sensing estimation method according to claim 1 is characterized in that: Step S1 further comprises the following steps: The surface temperature normalization formula is used to perform time normalization on the surface temperature data collected by the MODIS remote sensing sensor: ； Where, T norm is the surface temperature normalized to the reference time, T ac is the actual surface temperature during the solar period, a and b are coefficients, t ac is solar time, t norm is the base time; Hourly surface temperature was extracted based on ERA5_land reanalysis meteorological data, and masked with total cloud cover information of ERA5. Hourly monthly composite processing was performed on the surface temperature in the clear sky phase to obtain the spatiotemporally continuous monthly composite hourly ERA5_Land surface temperature data. The following formula was used to convert the UTC coordinated universal time corresponding to the ERA5_Land surface temperature data into solar time: ； Where, t solar_time is solar time, t UTC For Coordinated Universal Time, is the time difference between the standard time of the time zone where the study area is located and UTC, and Lon is the longitude value of the pixel; Based on the monthly composite hourly ERA5_Land surface temperature data and the corresponding solar time, the least squares fitting was performed pixel by pixel based on the surface temperature normalization formula to obtain the coefficients a and b values, and then resampled to 1 km resolution to be consistent with the resolution of MODIS surface temperature data; the surface temperature time normalization formula was applied pixel by pixel to perform time normalization on the MODIS surface temperature data to obtain the surface temperature data normalized to 11 o'clock.

3. The three-stage high-precision near-surface air temperature remote sensing estimation method according to claim 1 is characterized in that: In step S2, the process of clustering the study area into multiple sub-areas according to differences in natural conditions includes the following steps: The albedo, elevation, solar downward radiation DSR, vegetation index NDVI and water body index MNDWI were annually averaged and normalized to eliminate the influence of value range differences in the classification process. Longitude and latitude were used as classification features, and the iterative self-organizing data analysis model was used for clustering and partitioning. The study area was divided into several sub-areas through iteration.

4. The three-stage high-precision near-surface air temperature remote sensing estimation method according to claim 3 is characterized in that: Resample the MODIS reflectance product MOD09A1, MODIS albedo product MCD43A3, ERA5-Land and SRTM / DEM data to 1km resolution to be consistent with the resolution of MODIS surface temperature data; extract the reflectance of the red band, near infrared band, green band and shortwave infrared 1 band from the preprocessed MODIS reflectance product MOD09A1; extract the albedo data and longitude and latitude data from the preprocessed MODIS albedo product MCD43A3; extract the solar downward radiation DSR from the preprocessed ERA5-Land; extract the elevation and terrain index CTI from the preprocessed SRTM / DEM data; The vegetation index NDVI and water body index MNDWI at a resolution of 500 m are calculated using the following formulas: ； ； In the formula, ρ NIR and ρ RED Represent the reflectance of near infrared band and red band respectively; ρ Green is the green band reflectivity, ρ SWIR1 It is the reflectivity of shortwave infrared 1 band.

5. The three-stage high-precision near-surface air temperature remote sensing estimation method according to claim 1 is characterized in that: In step S2, the process of estimating the temperature in each sub-area using four machine learning models includes the following steps: Through the spatiotemporal matching of longitude, latitude and time, all spatial variables corresponding to the meteorological stations are extracted. The spatial variables include normalized surface temperature and spatial auxiliary variables including vegetation index NDVI, water body index MNDWI, albedo, elevation, downward solar radiation DSR, topographic index CTI, annual accumulated day DOY, longitude and latitude. Taking all spatial variables as independent variables and the measured temperature values ​​at meteorological stations as dependent variables, four machine learning methods, namely random forest, extreme gradient boosting, classification enhancement and Cubist, were used to construct temperature estimation models in each sub-area: ； In the formula, represents the estimated temperature at station i; represents the spatial coordinates at site i; is the intercept of the equation; is the regression coefficient of the corresponding independent variable; They are the normalized surface temperature, vegetation index NDVI, water body index MNDWI, albedo, elevation, downward solar radiation DSR, topographic index CTI, annual cumulative day DOY, longitude and latitude at station i respectively; is the random error term; The four constructed machine learning models were applied to the time-normalized surface temperature and spatial auxiliary variables including vegetation index NDVI, water body index MNDWI, albedo, elevation, downward solar radiation DSR, terrain index CTI, annual accumulated day DOY, longitude and latitude to obtain the spatial temperature estimation results.

6. The three-stage high-precision near-surface air temperature remote sensing estimation method according to claim 1 is characterized in that: In step S3, the temperature estimation results obtained by each machine learning model in step S2 are used as independent variables, and the measured temperature at the meteorological station is used as the dependent variable. The generalized additive model GAM is used to construct an integrated model: ； In the formula, Ta To observe near-surface air temperature; Ta model 1 、 model 2 、 model 3 、 model 4 are the temperature estimates of the four machine learning models: random forest, extreme gradient boosting, classification enhancement, and Cubist; S() represents the smooth function of the corresponding parameters; e i,j represents the residual error term for each observation; The constructed integrated model is applied to the estimated temperatures obtained by the four machine learning models to obtain the final integrated temperature estimate.

Images