A Downscaling Method of Land Surface Temperature Coupling Geographically Weighted and Random Forest

Abstract

The present invention relates to a method for downscaling land surface temperature by coupling geographically weighted and random forest, belonging to the technical field of remote sensing image processing. The method comprises the following steps: acquiring remote sensing data and preprocessing; selecting and calculating scale factors; establishing and optimizing parameters of a geographically weighted random forest downscaling model; inputting the preprocessed high-resolution scale factors and residuals into the geographically weighted random forest downscaling model to obtain high-resolution land surface temperature, and finally evaluating the model accuracy through the coefficient of determination, root mean square error, and mean absolute error. Based on the geographically weighted regression model, the present invention introduces a random forest regression model to express the non-stationarity and non-linearity between land surface temperature and scale factors.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and relates to a method for downscaling land surface temperature by coupling geographically weighted and random forest. Background Art

[0002] As one of the most important parameters describing the interaction between the atmosphere and the land surface, material cycling and energy exchange in regional and global-scale terrestrial ecosystems, land surface temperature (LST) plays a crucial role in the land surface energy balance. LST is widely used to estimate evaporation, analyze the impact of drought on vegetation, detect forest fires, and monitor the urban heat island effect. In recent years, the importance of LST has received increasing attention.

[0003] Currently, the main way to obtain land surface temperature over a large area is still through thermal sensors carried by satellites. Thermal infrared remote sensing can dynamically and macroscopically detect and describe the spatial differences and diversities of land surface temperature, while conventional ground-based site observations are difficult to obtain the spatio-temporal distribution of land surface temperature at regional and global scales. Due to the limitations of remote sensing satellite payloads and the bottleneck of sensor manufacturing technology, it is difficult to obtain thermal infrared data with fine resolution simultaneously in terms of spatial and temporal scales, and it is impossible to obtain land surface temperature data with high spatial resolution on a single satellite or sensor. At the same time, the influence of thermal mixing effects and cloud cover further reduces the available thermal infrared images. To obtain high-resolution spatio-temporal data, selected methods need to be used to process thermal infrared data. Downscaling land surface temperature is one of the effective and commonly used methods. This method explores the statistical regression relationship between the scale factor and land surface temperature on the premise of the assumption of scale invariance, and combines low-spatial-resolution land surface temperature with high-spatial-resolution auxiliary data to obtain high-spatial-resolution land surface temperature.

[0004] Although researchers have made efforts to improve the accuracy and reliability of LST downscaling algorithms through various statistical regression methods in heterogeneous regions, from linear to nonlinear, from global to local, there are still some problems that need further study. For example, most current downscaling methods are single in the selection of scale factors, which is insufficient to characterize the influence of different land surface cover types on land surface temperature, and at the same time, the generalization ability of the downscaling model is lacking. The most critical problem is that these statistical downscaling algorithms mainly solve the spatial non-stationarity between land surface temperature and scale factors, and rarely consider the spatial non-stationarity and nonlinear relationship between land surface temperature and scale factors simultaneously in regions with strong spatial heterogeneity such as cities. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method for downscaling land surface temperature by coupling geographically weighted and random forest.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for downscaling land surface temperature coupling geographically weighted and random forest, the method comprising the following steps:

[0008] S1: Obtain remote sensing data and preprocess it; obtain thermal infrared remote sensing images and multispectral remote sensing images within the research area, and perform preprocessing;

[0009] S2: Selection and calculation of scale factors; select the reflectance data of visible and near-infrared, select spectral factors that can characterize vegetation, bare soil, urban impervious surfaces, and water bodies according to the land cover type of the research area, as well as DEM data and its derived indices that characterize terrain, and then screen the scale factors with high correlation with land surface temperature through Pearson correlation coefficient as the scale factors of the final model;

[0010] S3: Establishment and parameter optimization of the geographically weighted random forest downscaling model; the parameters of the geographically weighted random forest downscaling model include the number of decision trees, the number of features used by each decision tree, and the local bandwidth size; use the land cover type data to geographically weight the data within the bandwidth to express the spatial non-stationarity of the sample points, establish a geographically weighted random forest model for each sample point and the data within its bandwidth, and by means of step-by-step optimization, use grid search to cyclically adjust the values of individual parameters, and use 10-fold cross-validation to obtain the optimal parameter values; after obtaining the optimal values of each parameter, retrain the data using out-of-bag error, calculate the generalization error, obtain the residuals at low spatial resolution according to the established regression relationship, and improve the residual resolution to the target high spatial resolution through ordinary Kriging interpolation;

[0011] S4: Input the preprocessed high-resolution scale factors into the geographically weighted random forest regression model GWRF, obtain the land surface temperature values at high spatial resolution according to the regression relationship between the land surface temperature at low spatial resolution and the scale factors, and evaluate the model accuracy through the coefficient of determination R 2 , root mean square error RMSE, and mean absolute error MAE.

[0012] Optionally, perform radiometric correction, atmospheric correction, and geometric correction on the remote sensing data. The remote sensing data preprocessing includes resampling, cropping, and reprojection of the multispectral remote sensing images to the same research area; the obtained remote sensing data includes MODIS land surface temperature data and Landsat 8 surface reflectance data, DEM digital elevation data, where the MODIS land surface temperature data comes from the temperature product MOD11A1 with a spatial resolution of 1000 meters; the Landsat 8 surface reflectance data has a spatial resolution of 30 meters; use the radiative transfer equation method to invert the temperature of the Landsat 8 thermal infrared remote sensing image to obtain the land surface temperature of the research area as the cross-validation data of the model.

[0013] Optionally, the Landsat 8 surface reflectance data is used for scale factors and calculating spectral indices, specifically including the Normalized Difference Vegetation Index (NDVI), the Soil-Adjusted Vegetation Index (SAVI), the Normalized Difference Water Index (NDWI), the Modified Normalized Difference Water Index (MNDWI), the Normalized Difference Built-up Index (NDBI), the Normalized Difference Bare and Built-up Index (NDBBI), the Urban Index (UI), the Soil-Adjusted Vegetation Index (SAVI), the Normalized Difference Drought Index (NDDI). The calculation of spectral indices is derived from the combination of single or multiple visible and near-infrared bands and shortwave infrared bands; the Digital Elevation Model (DEM) is used to calculate the Sky View Factor (SVF), slope, and aspect.

[0014] Optionally, the Geographically Weighted Random Forest Regression Model (GWRF) couples the idea of geographical weighting into the random forest model, and performs local non-linear modeling on the geographical space change characteristics in the form of multiple local spatial weighted random forests, enabling GWRF to have the ability to handle the relationship between spatial non-stationarity and non-linearity of spatial variables between the response variable and the predictor variable in the geographical space; GWRF constructs a local training set within an adaptive bandwidth or a fixed bandwidth through the geographical spatial position of the data. The data within the bandwidth uses the idea of geographical weighting and distance weighting for spatially unbalanced data with different coverage types, and then uses the random forest model to perform non-linear fitting on the non-stationary spatial data. The optimal bandwidth is determined by minimizing the out-of-bag error E of the random forest model. OOB Determine the hyperparameters of the coupled model, including the number of trees in the random forest and the number of predictor variables for each tree, using a traversal search method.

[0015] When applying the geographically weighted and random forest coupled model to downscale the land surface temperature, for areas with coverage types, the random forest coupled model non-linearly decomposes the land surface temperature through multiple spatial scale factors, and dynamically adjusts the spatial geographical weights according to the importance of scale factors for the land surface temperature under different coverage types, enabling the coupled model to fully characterize the influence of spatial scale factors on the land surface temperature in different regions; the scale factors and the land surface temperature are initially screened for features using the Pearson correlation coefficient, and then the feature importance is ranked through the global random forest regression model to select appropriate features as downscaling factors; the construction of the random forest is optimized by multiple decision trees to minimize the out-of-bag error for feature selection. The feature importance is obtained by calculating the out-of-bag error by changing the feature selection, that is, the greater the error before and after the change, the more important the feature; the out-of-bag error E OOB The formula is as follows:

[0016]

[0017] n is the number of out-of-bag samples: Based on the given sample X i is the output data of the random forest regression model; Y i is the observed data;

[0018] When fitting a non - linear relationship with a random forest regression model, a local random forest regression model is established for each sample point according to the bandwidth, and different weights are assigned to each sample point considering spatial heterogeneity. The weight function calculates the weights based on the geospatial distance between two points and considering the land cover type of the sample points. The weight calculation is as follows:

[0019]

[0020] This weight function considers different land cover types. That is, when the sample points within the bandwidth are of the same cover type, the weight is larger. If the number of sample points within the bandwidth is too small, different cover types are considered, but the sample points with the closest distance are considered, and different weights are considered for sample points of different cover types. Among them, lc i is the land cover type of the i - th sample point, d ij is the spatial distance from sample point i to sample point j, b is the bandwidth, α is a multiple of the distance, and a number greater than 1 is taken;

[0021] The geographically weighted random forest regression model of the scale factor and land surface temperature established at low resolution when applying the geographically weighted and random forest coupled model to land surface temperature downscaling is as follows:

[0022]

[0023] Among them, T b (W·X i ) represents the b - th decision tree in the random forest model, W represents the spatial distance weight, GWRF i represents the coupled model constructed for the i - th data point, represents the land surface temperature of data point i at low spatial resolution, that is, the MODIS land surface temperature, is the k - th scale factor of data point i at low spatial resolution, is the model prediction value, is the residual between the MODIS land surface temperature and the model prediction value at low resolution;

[0024] Based on scale invariance, the regression relationship at low spatial resolution is applied to high resolution to calculate the land surface temperature value at high spatial resolution. The calculation formula is:

[0025]

[0026] Among them, represents the predicted land surface temperature of data point i at high spatial resolution, is the k-th scale factor of data point i at high spatial resolution, is the residual at high resolution after ordinary Kriging interpolation.

[0027] Optionally, after S4, S5 is further included: evaluating the downscaled high-resolution land surface temperature data with the high-resolution land surface temperature results retrieved from the thermal infrared band of Landsat8, and comprehensively examining the accuracy of the downscaling results using the coefficient of determination R and the root mean square error RMSE index

[0028] The beneficial effects of the present invention are as follows: The present invention proposes a new method for downscaling land surface temperature by coupling GWR and random forest. Multiple spectral indices, reflectivities, and land surface form indices are selected as scale candidate factors, and the spatial resolution of MODIS LST is increased from 1000 meters to 100 meters. Based on the geographically weighted regression model, the algorithm introduces a random forest regression model to express the non-stationarity and non-linearity between land surface temperature and scale factors, and can better describe the relationship between LST and scale factors.

[0029] The main innovation of the present invention is to consider the spatial heterogeneity of the study area and the complex non-linearity between land surface temperature and scale factors, and propose a new land surface temperature downscaling model - Geographically Weighted Random Forest Model (GWRF). Compared with the prior art, the present invention first selects scale factors with a relatively large correlation with land surface temperature according to a semi-empirical method, and then uses the insensitive characteristics of random forest to variables to represent land surface temperature with multiple scale factors, realizing the adaptive selection of scale factors and enhancing the generalization of the downscaling model for different land cover areas. Spectral indices including Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (MNDWI), Building Index (NDBI), Normalized Difference Bare Land and Built-up Index (NDBBI), Urban Index (UI), Soil Adjusted Vegetation Index (SAVI), Normalized Difference Drought Index (NDDI) are selected. The calculation of spectral indices is derived from the combination of single or multiple visible and near-infrared bands and short-wave infrared bands; the sky view factor (SVF), slope, and aspect are calculated using the Digital Elevation Model (DEM); reflectivities including red band reflectivity, blue band reflectivity, green band reflectivity, short-wave infrared reflectivity (SWIR), etc. are used as downscaling factors to downscale the spatial resolution of MODIS temperature products to a finer resolution, obtaining high-precision and high-resolution land surface temperature under cloud-free conditions.

[0030] Other advantages, objects, and features of the present invention will be set forth in part in the following description, and in part will be obvious to those skilled in the art upon examination of the following, or may be learned by practice of the present invention. The objects and other advantages of the present invention may be realized and obtained by the following description of the specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0032] Figure 1 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] The following specific examples illustrate the embodiments of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0034] Among them, the accompanying drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as a limitation to the present invention; in order to better illustrate the embodiments of the present invention, some components in the accompanying drawings will be omitted, enlarged, or reduced, and do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the accompanying drawings may be omitted.

[0035] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the accompanying drawings are only for illustrative purposes and should not be construed as a limitation to the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0036] Such as Figure 1As shown, a method for downscaling land surface temperature by coupling geographically weighted and random forest is proposed. The random forest (RF) is introduced into the geographically weighted regression (GWR). By considering the spatial non-stationarity among regions and the non-linear relationship between land surface temperature and scale factors, the relationship between land surface temperature and scale factors can be better described. The specific steps are as follows:

[0037] S1: Obtain remote sensing data and preprocess it; obtain thermal infrared remote sensing images and multispectral remote sensing images within the study area and preprocess them;

[0038] S2: Select and calculate scale factors; introduce urban morphological factors based on traditional spectral indices; upscale each scale factor to 1000 meters and 100 meters respectively;

[0039] S3: Establish and optimize the parameters of the geographically weighted random forest regression model; the parameters of the geographically weighted random forest regression model include the number of decision trees, the number of features used in each decision tree, and the local bandwidth size; use land cover type data to weight the data within the bandwidth, and by gradually tuning, adopt grid search to cyclically adjust the values of individual parameters, and use 10-fold cross-validation to obtain the optimal parameter values; after obtaining the optimal values of each parameter, retrain the data using out-of-bag error, calculate the generalization error, and finally obtain the geographically weighted random forest regression model. According to the established regression relationship, obtain the residuals at low spatial resolution, and improve the residuals to the target high spatial resolution through ordinary Kriging interpolation;

[0040] S4: Input the preprocessed high-resolution scale factors into the geographically weighted random forest regression model, and obtain the land surface temperature values at high spatial resolution according to the regression relationship between land surface temperature and scale factors at low spatial resolution. Evaluate the model accuracy through the coefficient of determination R 2 , root mean square error RMSE, and mean absolute error MAE.

[0041] Preferably, in this study, the remote sensing data used is radiometrically corrected, atmospherically corrected, and geometrically corrected. The preprocessing of the present invention mainly includes resampling, cropping, and reprojection of the multispectral remote sensing images to the same study area. The obtained remote sensing data includes MODIS land surface temperature data and Landsat 8 surface reflectance data, DEM digital elevation data. Among them, the MODIS land surface temperature data comes from the temperature product MOD11A1 with a spatial resolution of 1000 meters; the Landsat 8 surface reflectance data has a spatial resolution of 30 meters. Then, the radiative transfer equation method is used to invert the temperature of the Landsat 8 thermal infrared remote sensing image to obtain the land surface temperature of the study area as the cross-validation data of the model.

[0042] Preferably, calculating spectral indices using Landsat 8 surface reflectance data mainly includes Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Water Index (NDWI), Modified Normalized Difference Water Index (MNDWI), Normalized Difference Built-up Index (NDBI), Normalized Difference Bare and Built-up Index (NDBBI), Urban Index (UI), Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Drought Index (NDDI). The calculation of spectral indices is derived from the combination of single or multiple visible and near-infrared bands and short-wave infrared bands; calculating Sky View Factor (SVF), slope, and aspect using Digital Elevation Model (DEM). Based on the excellent performance of random forest in high-dimensional feature processing, various scale factors are introduced without considering the problem of feature collinearity. And the random forest regression model can derive feature importance according to the degree of influence of features on the out-of-bag error. The scale factors and land surface temperature are initially screened for features using Pearson correlation coefficient, and then the appropriate features are selected as downscaling factors through the global random forest regression model for feature importance ranking.

[0043] Preferably, when fitting the non-linear relationship with the random forest regression model, the imbalance of the sample points themselves is considered. That is, the first law of geography states that in the spatial relationship of geographical variables, everything is spatially correlated, and things closer in distance have greater spatial correlation than those farther away. So, the observed data closer to the prediction point should be weighted more than those farther away, and the weights given to the observed points at different positions vary with their distances from the observation point. A local random forest regression model is established for each sample point according to the bandwidth, and different weights are assigned to each sample point considering spatial heterogeneity. The weight function calculates the weight based on the geographical spatial distance between two points. The weight calculation is as follows:

[0044]

[0045] This weight function considers different land surface cover types. That is, when the sample points within the bandwidth are of the same cover type, the weight is greater. If the number of sample points within the bandwidth is too small, different cover types but the nearest sample points are considered, and different weights are considered for sample points of different cover types, where lc i is the land surface cover type of the i-th sample point, d ij is the spatial distance from sample point i to sample point j, b is the bandwidth, α is a multiple of the distance, and a natural number greater than 1 is taken.

[0046] The construction of the random forest is to optimize the feature selection of decision trees by multiple decision trees according to the minimization of the out-of-bag error, and the feature importance is obtained by calculating the out-of-bag error by changing the feature selection. That is, the greater the error before and after the change, the more important the feature, which provides a basis for feature selection in the present invention. Out-of-bag error (Error of out of bag, EOOB ) The formula is as follows:

[0047]

[0048] n is the number of samples outside the bag: is based on the given sample X i Output data based on the random forest regression model; Y i is the observed data.

[0049] The geographically weighted random forest regression model of the scale factor and land surface temperature established at low resolution is as follows:

[0050]

[0051] Among them, represents the land surface temperature of data point i at low spatial resolution, that is, the MODIS land surface temperature, is the k-th scale factor of data point i at low spatial resolution, is the model prediction value, is the difference between the MODIS land surface temperature and the model prediction value at low resolution.

[0052] Furthermore, based on scale invariance, the regression relationship at low spatial resolution is applied to high resolution to calculate the land surface temperature value at high resolution. The calculation formula is:

[0053]

[0054] Among them, represents the land surface temperature of data point i at high spatial resolution, that is, the MODIS land surface temperature, is the k-th scale factor of data point i at high spatial resolution, is The residual at high resolution after ordinary cokriging interpolation.

[0055] Preferably, the high-resolution land surface temperature data after downscaling is used to evaluate the accuracy with the high-resolution land surface temperature results retrieved from the thermal infrared band of Landsat 8. The accuracy of the downscaling results is comprehensively evaluated using the determination coefficient R and the root mean square error RMSE indicators.

[0056] The present invention proposes a new method for downscaling land surface temperature based on the GWRF model, selects various spectral indices, reflectance, and land surface morphology indices as scale candidate factors, and improves the spatial resolution of MODIS LST from 1000 meters to 100 meters. By introducing a random forest regression model on the basis of the geographically weighted regression model, this algorithm can express the non-stationarity and non-linearity between land surface temperature and scale factors, and can better describe the relationship between LST and scale factors.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A downscaling method for land surface temperature coupling geographically weighted and random forest, characterized in that: This method includes the following steps: S1: Obtain remote sensing data and preprocess it; obtain thermal infrared remote sensing images and multispectral remote sensing images within the study area and preprocess them; S2: Selection and calculation of scale factors; select the reflectance data in the visible and near-infrared bands, select spectral factors that can represent vegetation, bare soil, urban impervious surfaces, and water bodies according to the land cover type of the study area, as well as DEM data and its derived indices that represent terrain, and then screen the scale factors with high correlation with land surface temperature through the Pearson correlation coefficient as the scale factors for the final model; S3: Establishment and parameter optimization of the geographically weighted random forest downscaling model; the parameters of the geographically weighted random forest downscaling model include the number of decision trees, the number of features used in each decision tree, and the local bandwidth size; use the land cover type data to geographically weight the data within the bandwidth to express the spatial non-stationarity of the sample points, establish a geographically weighted random forest model for each sample point and the data within its bandwidth, and by means of step-by-step optimization, use grid search to cyclically adjust the values of individual parameters, and use 10-fold cross-validation to obtain the optimal parameter values; after obtaining the optimal values of each parameter, retrain the data using the out-of-bag error, calculate the generalization error, obtain the residuals at low spatial resolution according to the established regression relationship, and improve the residual resolution to the target high spatial resolution through ordinary Kriging interpolation; When applying the geographically weighted and random forest coupled model to the downscaling of land surface temperature, for areas with coverage types, the random forest coupled model nonlinearly decomposes the land surface temperature through multiple spatial scale factors, and dynamically adjusts the spatial geographical weights according to the importance of scale factors for the land surface temperature under different coverage types, enabling the coupled model to fully characterize the influence of spatial scale factors on the land surface temperature in different regions; the scale factors and the land surface temperature are initially screened for features using the Pearson correlation coefficient, and then the feature importance is ranked through a global random forest regression model to select appropriate features as downscaling factors; the construction of the random forest is based on multiple decision trees optimizing the feature selection of the decision tree according to the out-of-bag error minimization, and the feature importance is obtained by calculating the out-of-bag error by changing the feature selection, that is, the greater the error before and after the change, the more important the feature; the out-of-bag error E OOB The formula is as follows: n is the number of out-of-bag samples: is based on the given sample X i Output data of the random forest regression model; Y i is the observed data; When fitting the non-linear relationship with the random forest regression model, establish a local random forest regression model for each sample point according to the bandwidth, and consider spatial heterogeneity to assign different weights to each sample point. The weight function calculates the weights based on the geographical spatial distance between two points and considering the land cover type of the sample point. The weight calculation is as follows: The weight function considers different land cover types. That is, when the sample points within the bandwidth are of the same cover type, the weight is greater. If the number of sample points within the bandwidth is too small, different cover types are considered, but the sample points with the closest distance are considered. Different weights are considered for sample points of different cover types, where lc i is the land cover type of the i-th sample point, d ij is the spatial distance from sample point i to sample point j, b is the bandwidth, α is a multiple of the distance, and a number greater than 1 is taken; When applying the geographically weighted and random forest coupling model to the downscaling of land surface temperature, the geographically weighted random forest regression model between the scale factors and land surface temperature established at low resolution is as follows: where T b (W·X i ) represents the b-th decision tree in the random forest model, W represents the spatial distance weight, and GWRF i represents the coupling model of the i-th data point constructed, represents the land surface temperature of data point i at low spatial resolution, i.e., the MODIS land surface temperature, is the k-th scale factor of data point i at low spatial resolution, is the model prediction value, is the residual between the MODIS land surface temperature and the model prediction value at low resolution; Based on scale invariance, apply the regression relationship at low spatial resolution to high resolution, and calculate the land surface temperature value at high spatial resolution. The calculation formula is: Among them, represents the predicted surface temperature of data point i at high spatial resolution, is the k-th scale factor of data point i at high spatial resolution, is the residual after ordinary Kriging interpolation at high resolution; S4: Input the preprocessed high-resolution scale factor into the geographically weighted random forest regression model GWRF, and obtain the land surface temperature value at high spatial resolution according to the regression relationship between the land surface temperature and the scale factor at low spatial resolution. Evaluate the model accuracy through the coefficient of determination R 2 , root mean square error RMSE, and mean absolute error MAE.

2. A downscaling method for land surface temperature coupling geographically weighted and random forest according to claim 1, characterized in that: Perform radiometric correction, atmospheric correction, and geometric correction on the remote sensing data. The preprocessing of the remote sensing data includes resampling, cropping, and reprojection of the multispectral remote sensing images to the same study area; the obtained remote sensing data includes MODIS land surface temperature data and Landsat 8 surface reflectance data, DEM digital elevation data, where the MODIS land surface temperature data comes from the temperature product MOD11A1 with a spatial resolution of 1000 meters; the Landsat 8 surface reflectance data has a spatial resolution of 30 meters; use the radiative transfer equation method to invert the temperature of the Landsat 8 thermal infrared remote sensing image to obtain the land surface temperature of the study area as the cross-validation data of the model.

3. A downscaling method for land surface temperature coupling geographically weighted and random forest according to claim 2, characterized in that: The Landsat 8 surface reflectance data is used for scale factors and calculating spectral indices, specifically including the Normalized Difference Vegetation Index (NDVI), the Soil-Adjusted Vegetation Index (SAVI), the Normalized Difference Water Index (NDWI), the Modified Normalized Difference Water Index (MNDWI), the Normalized Difference Built-up Index (NDBI), the Normalized Difference Bare and Built-up Index (NDBBI), the Urban Index (UI), the Soil-Adjusted Vegetation Index (SAVI), the Normalized Difference Drought Index (NDDI). The calculation of spectral indices is derived from the combination of single or multiple visible and near-infrared bands and short-wave infrared bands; the Digital Elevation Model (DEM) is used to calculate the Sky View Factor (SVF), slope, and aspect.

4. A downscaling method for land surface temperature coupling geographically weighted and random forest according to claim 1, characterized in that: The geographical weighted random forest regression model GWRF couples the idea of geographical weighting into the random forest model, and conducts local non-linear modeling of geographical space change characteristics in the form of multiple local spatial weighted random forests, enabling GWRF to have the ability to handle the relationships between spatial non-stationarity and spatial variable non-linearity simultaneously between the response variable and the predictor variable in geographical space; GWRF constructs a local training set within an adaptive bandwidth or a fixed bandwidth based on the geographical spatial positions among data. The data within the bandwidth uses the idea of geographical weighting and distance weighting for spatially unbalanced data with different coverage types, and then uses the random forest model to perform non-linear fitting on the non-stationary spatial data. The optimal bandwidth is determined by minimizing the out-of-bag error E of the random forest model. OOB Determine the hyperparameters of the coupled model, including the number of trees in the random forest and the number of predictor variables for each tree, by using the method of exhaustive search.

5. A downscaling method for land surface temperature coupling geographically weighted and random forest according to claim 1, characterized in that: after S4, it further includes S5: evaluating the high-resolution land surface temperature data after downscaling with the high-resolution land surface temperature results retrieved from the Landsat 8 thermal infrared band, and comprehensively examining the accuracy of the downscaling results using the coefficient of determination R and the Root Mean Square Error (RMSE) index.

Images