A Machine Learning-Based Competitive Remote Sensing Method for Refining Land Surface Temperature

By constructing a high spatial resolution set of surface parameters from remote sensing images and performing pixel-by-pixel regression with a sliding window, and selecting the optimal machine learning model, the problem of identifying the heterogeneity of surface temperature and influencing factors in the spatial downscaling of remote sensing surface temperature was solved, and high-precision refinement of remote sensing surface temperature was achieved.

CN115660100BActive Publication Date: 2026-11-10INST OF GEOGRAPHICAL SCI & NATURAL RESOURCE RES CAS
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202211093293.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-09-08
Publication Date
2026-11-10
Estimated Expiration
2042-09-08

AI Technical Summary

Technical Problem

Existing spatial downscaling methods for remote sensing surface temperature cannot effectively identify the spatial heterogeneity of physical laws between surface temperature and influencing factors. As a result, the spatial downscaling of surface temperature in the thermal infrared band cannot achieve overall optimization and cannot fully utilize the multispectral information of remote sensing images.

Method used

A fine set of surface parameters is constructed using high spatial resolution bands of remote sensing images. The optimal surface parameters and machine learning model are determined by traversing pixel by pixel through a sliding window. The optimal model and parameters are selected using the competition rule of minimum absolute deviation, and a high spatial resolution predicted surface temperature is generated.

Benefits of technology

It improves the spatial accuracy of remote sensing of land surface temperature, reveals the local laws between land surface temperature and geographical environment, is applicable to the spatial downscaling of other land surface parameters, and has a high generalization ability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115660100B_ABST
    Figure CN115660100B_ABST
Patent Text Reader

Abstract

This invention provides a method for refining remotely sensed land surface temperature based on machine learning competition, comprising: constructing a fine set of land surface parameters affecting land surface temperature using high spatial resolution imagery and upscaling it to a coarse set of land surface parameters in the thermal infrared band spatial resolution, and creating a set of machine learning methods; using a sliding window to traverse and slide, determining the optimal coarse set of land surface parameters and machine learning model for each method in the method set per pixel through successive regression; selecting the optimal machine learning model and optimal set of land surface parameters per pixel based on a competition rule that minimizes the absolute deviation between the remotely sensed land surface temperature and the predicted land surface temperature; generating a high spatial resolution predicted land surface temperature based on the optimal set of land surface parameters and the optimal machine learning model per pixel; and refining the high spatial resolution predicted land surface temperature using the minimum deviation between the remotely sensed land surface temperature and the predicted land surface temperature per pixel. This invention can refine remotely sensed land surface temperature pixel by pixel and can be used for precise estimation of surface water and heat flux.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a method for spatial downscaling of remotely sensed surface temperature, and more particularly to a machine learning method for refining remotely sensed surface temperature in the thermal infrared band using multispectral bands of remotely sensed images. Background Technology

[0002] Land surface temperature is one of the most important state variables of the land surface, closely related to the exchange of water and heat energy between land and air, inseparable from regional and global climate change, and intimately connected to the reproduction and survival of the Earth's biosphere. Therefore, the measurement of land surface temperature is crucial for maintaining the sustainability of the global ecological environment. Calculating land surface temperature by measuring surface thermal infrared radiation energy using remote sensing technology is currently the only effective method for monitoring large-scale land surface temperatures. However, due to the difficulty in miniaturizing thermal infrared wavelength sensors, the spatial resolution of the thermal infrared band in remote sensing images is lower than that of other bands. The retrieved remote sensing land surface temperature often lacks detailed information about the temperature of different surfaces, limiting the practical application of remote sensing land surface temperature. How to utilize multi-band information from remote sensing images to enrich the details of remote sensing land surface temperature has become an effective means to improve the spatial quality of remote sensing land surface temperature. Therefore, the rapid, accurate, and convenient acquisition of high-quality remote sensing land surface temperature products by integrating multi-temporal, multi-scale, and multi-satellite remote sensing images has become a key technology for solving ecological and environmental health problems such as global warming and droughts and floods.

[0003] Currently, spatial downscaling methods for remotely sensed land surface temperature (LST) fall into two categories: 1) statistical regression and 2) data fusion. Statistical regression assumes that the spatial scale of the relationship between LST and LST parameters remains unchanged, extending the statistical relationship between LST and LST parameters obtained at low spatial resolution to high spatial resolution, thereby enhancing the spatial downscaling of LST. Data fusion uses mathematical methods to extract spatial increments for improving LST from remote sensing products of different spatial resolutions across multiple time periods, effectively fusing spatial details from multi-temporal images with LST data to achieve spatial downscaling. Statistical regression often considers factors influencing LST such as NDVI, albedo, and reflectance, employing methods like DisTrad, TsHARP, HUTS, geographic weighted regression, artificial neural networks, and genetic algorithms. While statistical regression has a simple mechanism and clear logic, different considerations regarding LST influencing factors can lead to different spatial downscaling results for remotely sensed LST. Data fusion can fully extract the spatial details contained in multi-temporal remote sensing images, but the mathematical methods used are biased towards data mining and cannot reconstruct the physical relationships between remote sensing multispectral bands. In addition, both of these methods ignore the differences in the physical relationships between surface temperature and influencing factors caused by the spatial heterogeneity of the surface, and the spatial downscaling of surface temperature in the thermal infrared band cannot achieve overall optimization.

[0004] The shortcomings of the two methods mentioned above mainly stem from their inability to identify the spatial heterogeneity of physical laws governing surface temperature and its influencing factors, resulting in insufficient revelation of the evolution of surface temperature under changing surface environments. To fully consider the spatial heterogeneity of surface parameters affecting surface temperature, to effectively utilize the surface parameters revealed by the visible light band, and to further improve the spatial accuracy of remotely sensed surface temperature, a machine learning-based competitive method for refining remotely sensed surface temperature has been invented. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for refining remote sensing surface temperature based on machine learning competition.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A remote sensing surface temperature refinement method based on machine learning competition includes the following features:

[0008] P1: Construct a fine set of surface parameters affecting surface temperature using high spatial resolution bands of remote sensing images, and then scale it up to a coarse set of surface parameters with thermal infrared band spatial resolution, and create a set of machine learning methods.

[0009] P2: By using a sliding window to traverse and slide, the optimal coarse set of surface parameters and machine learning model for each pixel in the method set are determined through stepwise regression;

[0010] P3: The optimal machine learning model and the optimal set of surface parameters per pixel are selected based on the competition rule of minimizing the absolute deviation between the remotely sensed surface temperature and the predicted surface temperature.

[0011] P4: Generate high spatial resolution predicted land surface temperature based on the pixel-by-pixel optimal set of surface parameters and the optimal machine learning model;

[0012] P5: Refine the high spatial resolution predicted land surface temperature by utilizing the minimum pixel-by-pixel deviation between remotely sensed land surface temperature and predicted land surface temperature.

[0013] In one embodiment, P1 includes the following steps:

[0014] P11: A fine set of surface parameters affecting surface temperature is constructed using high spatial resolution bands of remote sensing imagery, specifically:

[0015] Ω={NDVI, FCVI, AWEIsh, ATSAVI, AFRI, ARVI, BAI, BCC, BWDRVI, WDRVI, CVI, EBBI, EVI, GARI, GLI, GVMI, MCARI, NDBI, NDGI, NDPI, SAVI, MWIR, RDVI, MCARI, SARVI, MSAVI, NLI, EBSI, EMBI, ALB}

[0016] In the formula, Ω represents the high spatial resolution fine set of surface parameters affecting surface temperature, NDVI represents the normalized difference vegetation index, FCVI represents the fluorescence-corrected vegetation index, AWEIsh represents the automatic water extraction index for shading elimination, ATSAVI represents the adjusted conversion soil-regulated vegetation index, AFRI represents the aerosol-free vegetation index, ARVI represents the atmospheric resistance vegetation index, BAI represents the burn area index, BCC represents the blue color coordinate, BWDRVI represents the blue wide dynamic range vegetation index, WDRVI represents the wide range dynamic vegetation index, CVI represents the chlorophyll vegetation index, EBBI represents the enhanced building exposure index, EVI represents the enhanced vegetation index, and GARI represents green atmospheric resistant vegetation. The indices are as follows: GLI (Green Leaf Index), GVMI (Gross Vegetation Moisture Index), MCARI (Modified Chlorophyll Absorption Index), NDBI (Normalized Difference Bareness Index), NDGI (Normalized Difference Greenness Index), NDPI (Normalized Difference Phenological Index), SAVI (Soil-Regulated Vegetation Index), MWIR (Modified Multispectral Water Index), RDVI (Renormalized Difference Vegetation Index), MCARI (Modified Chlorophyll Absorption Ratio Index), SARVI (Soil and Atmospheric Resistance Vegetation Index), MSAVI (Modified Soil-Adjusted Vegetation Index), NLI (Nonlinear Vegetation Index), EBSI (Enhanced Bare Soil Index), EMBI (Enhanced and Improved Bare Soil Index), and ALB (Albedo).

[0017] P12: Upscale the fine set of surface parameters to a coarse set of surface parameters with spatial resolution in the thermal infrared band, specifically:

[0018]

[0019] e i =U(E i ) 1≤i≤I E i ∈Ω

[0020] In the formula, U represents the spatial downscaling model, and E i Let e ​​represent the i-th fine surface parameter in Ω. i This means using U to transfer E iSpatial upscaling is a coarse surface parameter with spatial resolution in the thermal infrared band, where I represents the number of elements in Ω. Indicates by {e i} 1≤i≤I A coarse set of surface parameters;

[0021] P13: The collection of methods for creating machine learning tools is as follows:

[0022] M={OLS, SVR, RFR, CNR, PNN, XGB, LAS, PCR}

[0023] In the formula, OLS represents least squares, SVR represents support vector regression, RFR represents random forest, CNR represents convolutional network, PNN represents perceptron neural network, XGB represents extreme gradient boosting tree, LAS represents lasso algorithm, PCR represents principal component regression, and M represents the set of machine learning methods. Let J represent the j-th element in M, where 1 ≤ j ≤ J, and J represents the number of elements in M.

[0024] In one embodiment, the specific steps of P2 are as follows:

[0025] P21: For each element in M Perform steps P22-P25;

[0026] P22: Targeting Generate a set of windows that slide across the remotely sensed surface temperature, specifically: A j ={h×w, h×(w+1),…, h×W, (h+1)×w, (h+1)×(w+1),…, (h+1)×W,…, H×W}

[0027] In the formula, A j Indicating targeting The set of all sliding windows, h×w, represents the operation on remotely sensed surface temperature. The smallest window, H×W, represents the row and column of the remotely sensed surface temperature;

[0028] P23: Using A j Each element in the data is traversed pixel by pixel along the row and column directions of the remotely sensed surface temperature, for example, A. j The area covered by each element in each swipe Perform step P24;

[0029] P24: Using successive regression from 2 of I elements I - From the -1 combinations, the optimal coarse set of surface parameters is selected. The criterion is that the root mean square error of the difference between the predicted surface temperature and the remotely sensed surface temperature is the smallest. Specifically:

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] In the formula, express exist From The optimal coarse set of surface parameters selected from the data. express Surface parameters in various parts of China The position in the middle, yes against exist The concrete model on, express use exist The above predictions of surface temperature, express The predicted surface temperature, Represents any given middle set, express Surface parameters in various parts of China The position in the middle, yes against exist The concrete model on, express use exist The above predictions of surface temperature, express The predicted surface temperature, T jk express For remote sensing of land surface temperature, RMSE represents the root mean square error of the regional average. express With T jk The difference, || represents the absolute value taken pixel by pixel. express absolute value per pixel;

[0036] P25: Select the optimal coarse set of surface parameters and the corresponding machine learning model from all regions, pixel by pixel, specifically as follows:

[0037]

[0038]

[0039] In the formula, (x,y) represents the location of the remotely sensed surface temperature pixel, κ jxy This indicates that at (x,y) Predict the area with the smallest absolute deviation in surface temperature. express When predicting the surface temperature with the smallest absolute deviation at (x,y), select The set of element positions express The minimum absolute deviation of the predicted surface temperature at (x,y). express All sliding windows iterate through the set of all regions containing (x, y) within the sliding window, where Υ represents... Any given element in Indicating targeting The coarse set of optimal surface parameters at (x,y) express The optimal machine learning model at (x,y).

[0040] In one embodiment, P3 specifically refers to:

[0041]

[0042]

[0043]

[0044] In the formula, ι represents the method subscript for predicting the surface temperature at (x,y) with the minimum absolute deviation in M, and κ represents the method subscript for predicting the surface temperature at (x,y). ιxy This indicates the area number in M ​​that has the smallest absolute deviation in surface temperature forecast at (x,y). When the surface temperature at (x,y) in the forecast is selected in M, it represents the position with the smallest absolute deviation. Location of the optimal surface parameters. This represents the coarse set of optimal surface parameters at (x,y). Let represent the optimal machine learning model at (x,y).

[0045] In one embodiment, P4 specifically refers to:

[0046]

[0047] In the formula, This represents a specific model that predicts the surface temperature at (x,y) with the minimum absolute deviation. express In the region κ ιxy The selected optimal coarse set of surface parameters The optimal fine set of surface parameters composed of various surface parameters in China. express use The predicted surface temperature at (x,y) with high spatial resolution.

[0048] In one embodiment, P5 specifically refers to:

[0049]

[0050] In the formula, D represents the spatial downscaling model. This represents the minimum pixel-by-pixel deviation between the remotely sensed surface temperature and the predicted surface temperature. Indicates the use of Correction The results of remote sensing of Earth's surface temperature were refined.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] 1) The localized nature of the relationship between surface temperature and geographical environment was revealed by the surface parameters related to vegetation, soil, water bodies and impermeable surfaces derived from the high spatial resolution bands of remote sensing images.

[0053] 2) A machine learning competition scheme for pixel-by-pixel optimal refinement of remote sensing surface temperature was proposed;

[0054] 3) The method proposed in this invention has high generalization ability and is applicable to the spatial downscaling of other surface parameters such as soil moisture and hydrothermal flux. Attached Figure Description

[0055] Figure 1 Flowcharts of embodiments provided by the present invention;

[0056] Figure 2 P3 flowchart provided for an embodiment of the present invention Detailed Implementation

[0057] The technical features, beneficial effects, and implementation process of the present invention will be described in detail below with reference to the accompanying drawings used in the embodiments. It should be noted that the embodiments are only some embodiments of the technical solution of the present invention and should not be regarded as a limitation on the scope of the present invention.

[0058] Please see Figure 1 , Figure 1 The technical flowchart used in a specific embodiment of the present invention includes the following steps:

[0059] P1: Acquire and preprocess remote sensing images and remote sensing surface temperature of the study area;

[0060] P2: Construct a fine set of surface parameters affecting surface temperature using high spatial resolution bands of remote sensing images, and then scale it up to a coarse set of surface parameters with thermal infrared band spatial resolution, and create a set of machine learning methods.

[0061] P3: Using a sliding window to traverse and slide, determine the optimal coarse set of surface parameters and machine learning model for each pixel in the method set through stepwise regression;

[0062] P4: Select the optimal machine learning model and the optimal set of surface parameters per pixel based on the competition rule of minimizing the absolute deviation between remotely sensed surface temperature and predicted surface temperature.

[0063] P5: Generate high spatial resolution predicted land surface temperature based on the pixel-by-pixel optimal set of surface parameters and the optimal machine learning model;

[0064] P6: Refine the high spatial resolution predicted land surface temperature by using the minimum pixel-by-pixel deviation between remotely sensed land surface temperature and predicted land surface temperature.

[0065] In P1, the remote sensing image set of the study area includes near-infrared band, red band, green band, blue band, short-wave infrared, and thermal infrared band. Each band of the remote sensing image set can be composed of remote sensing images from multiple stars or multiple periods.

[0066] For example, the remote sensing image set may be Sentinel 3 Land-Sea Colorimeter (OLCI) and Land-Sea Surface Temperature Radiometer (SLSTR) images, or Landsat 8 Land Imager Sensor (OLI) and Thermal Infrared Sensor (TIRS) images, or ASTER sensor images, which can be downloaded via ftp: / / ladsweb.modaps.eosdis.nasa.gov;

[0067] In P1, the remote sensing surface temperature of the study area can be obtained by inverting from the remote sensing image set using a surface temperature inversion algorithm. For example, the surface temperature inversion algorithm can be selected from the temperature emissivity separation method, the thermal radiation transfer equation, the empirical statistical method, or the split window algorithm. The atmospheric water vapor content data required for surface temperature inversion can be selected from the MOD021KM product, and the surface emissivity can be selected from the ASTER GEDv3 product, both of which can be downloaded from ftp: / / ladsweb.modaps.eosdis.nasa.gov.

[0068] In P1, the remote sensing land surface temperature of the study area can also be downloaded from the website. For example, MOD11_L2, MOD11A1, MOD11B1, or MOD11A2 land surface temperature products can be downloaded from ftp: / / ladsweb.modaps.eosdis.nasa.gov. These products have global coverage, with spatial resolutions of 1 km, 1 km, 6 km, and 1 km, and temporal resolutions of instantaneous, daily, daily, and eight-day, respectively. The geographic coordinates are latitude and longitude, sine, sine, and sine. These products are currently widely used, and the data is updated daily. They have undergone many corrections during the inversion process and have wide applications in the fields of resource environment, climate change, and terrestrial ecosystems. They are important data sources for land surface temperature-related research.

[0069] In P1, the preprocessing for remote sensing image sets includes radiometric calibration, atmospheric correction, geometric correction, and projection transformation. Radiometric calibration converts image pixel values ​​into physical quantities such as emissivity or reflectivity of the corresponding pixels based on the gain and offset of each band in the remote sensing image header file, and is applied to the thermal infrared band and multispectral band respectively. Atmospheric correction is generally applied to the multispectral band of remote sensing images and can be completed using the dark target method. Geometric correction is used to locate the remote sensing image set on geographic coordinates and correct the geometry of the remote sensing image set to match the base map, and can be completed using software such as ENVI, ERDAS, or ArcGIS. Projection transformation converts the remote sensing image set to a unified projection, and can be completed using software such as ENVI, ERDAS, or ArcGIS. For remote sensing surface temperature, preprocessing includes data format conversion, reprojection, cropping and stitching, and can be performed using the specialized software tool HEG.

[0070] In P2, the step of constructing a fine set of surface parameters affecting surface temperature using high spatial resolution bands of remote sensing images and then scaling it up to a coarse set of surface parameters with thermal infrared band spatial resolution, as well as creating a set of machine learning methods, includes the following steps:

[0071] P21: A fine set of surface parameters affecting surface temperature is constructed using high spatial resolution bands of remote sensing imagery, specifically:

[0072] Ω={NDVI, FCVI, AWEIsh, ATSAVI, AFRI, ARVI, BAI, BCC, BWDRVI, WDRVI, CVI, EBBI, EVI, GARI, GLI, GVMI, MCARI, NDBI, NDGI, NDPI, SAVI, MWIR, RDVI, MCARI, SARVI, MSAVI, NLI, EBSI, EMBI, ALB}

[0073] In the formula, Ω represents the high spatial resolution fine set of surface parameters affecting surface temperature, NDVI represents the normalized difference vegetation index, FCVI represents the fluorescence-corrected vegetation index, AWEIsh represents the shadow removal automatic water extraction index, ATSAVI represents the adjusted conversion soil-regulated vegetation index, AFRI represents the aerosol-free vegetation index, ARVI represents the atmospheric resistance vegetation index, BAI represents the burn area index, BCC represents the blue color coordinate, BWDRVI represents the blue wide dynamic range vegetation index, WDRVI represents the wide range dynamic vegetation index, CVI represents the chlorophyll vegetation index, EBBI represents the enhanced building exposure index, EVI represents the enhanced vegetation index, GARI represents the green atmospheric resistance vegetation index, GLI represents the green leaf index, GVMI represents the global vegetation moisture index, and MCARI represents the corrected The chlorophyll uptake index, NDBI (Normalized Difference Bareness Index), NDGI (Normalized Difference Greenness Index), NDPI (Normalized Difference Phenological Index), SAVI (Soil-Adjusted Vegetation Index), MWIR (Modified Multispectral Water Index), RDVI (Renormalized Difference Vegetation Index), MCARI (Modified Chlorophyll Absorption Ratio Index), SARVI (Soil and Atmospheric Resistance Vegetation Index), MSAVI (Modified Soil-Adjusted Vegetation Index), NLI (Nonlinear Vegetation Index), EBSI (Enhanced Bare Soil Index), EMBI (Enhanced Improved Bare Soil Index), and ALB (Albedo) are all indices. For their calculation formulas, please refer to https: / / awesome-ee-spectral-indices.readthedocs.io / en / latest / .

[0074] P22: Upscaling the fine set of surface parameters to a coarse set of surface parameters with spatial resolution in the thermal infrared band, specifically:

[0075]

[0076] e i =U(E i )1≤i≤I E i ∈Ω

[0077] In the formula, U represents the spatial upscaling model, and E i Let e ​​represent the i-th fine surface parameter in Ω. i This means using U to transfer E i Spatial upscaling is a coarse surface parameter with spatial resolution in the thermal infrared band, where I represents the number of elements in Ω. Indicates by {e i} 1≤i≤I A coarse set of surface parameters;

[0078] For example, the spatial upscaling model U can be selected as a low-pass filter, specifically:

[0079]

[0080]

[0081] In the formula, (x,y) represents the surface parameter e with spatial resolution in the thermal infrared band. i The mid-pixel location (x+c, y+d) represents the high spatial resolution surface parameter E. i The middle pixel position, 2*W+1 represents the size of the rectangular low-pass filter, which is determined by the ratio of the spatial resolution of the thermal infrared band to the high spatial resolution. (c,d) represents the position of the filter value in the rectangular low-pass filter. Γ represents the low-pass filter, and the filter value is determined by the distance of its position in the window from the center of the window. The width of the low-pass filter σ determines the smoothness. The larger σ is, the wider the bandwidth of Γ and the higher the smoothness.

[0082] P23: A collection of methods for creating machine learning tools, specifically:

[0083] M={OLS, SVR, RFR, CNR, PNN, XGB, LAS, PCR}

[0084] In the formula, OLS represents least squares, SVR represents support vector regression, RFR represents random forest, CNR represents convolutional network, PNN represents perceptron neural network, XGB represents extreme gradient boosting tree, LAS represents lasso algorithm, PCR represents principal component regression, and M represents the set of machine learning methods. Let J represent the j-th element in M, where 1 ≤ j ≤ J, and J represents the number of elements in M.

[0085] In P3, the process of using a sliding window to iterate and determine the optimal coarse set of surface parameters and machine learning model for each pixel in the method set through stepwise regression is described in the following section. Figure 2 The specific steps are as follows:

[0086] P31: For each element in M Perform steps P32-P35;

[0087] P32: Targeting Generate a set of windows that slide across the remotely sensed surface temperature, specifically: A j ={h×w, h×(w+1),…, h×W, (h+1)×w, (h+1)×(w+1),…, (h+1)×W,…, H×W}

[0088] In the formula, A j Indicating targeting The set of all sliding windows, h×w, represents the operation on remotely sensed surface temperature. The smallest window, H×W, represents the row and column of the remotely sensed surface temperature;

[0089] P33: Using A j Each element in the data is traversed pixel by pixel along the row and column directions of the remotely sensed surface temperature, for example, A. j The area covered by each element in each swipe Perform step P34;

[0090] P34: Using successive regression from 2 of I elements I - From the -1 combinations, the optimal coarse set of surface parameters is selected. The criterion is that the root mean square error of the difference between the predicted surface temperature and the remotely sensed surface temperature is the smallest. Specifically:

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] In the formula, express exist From The optimal coarse set of surface parameters selected from the data. express Surface parameters in various parts of China The position in the middle, yes against exist The concrete model on, express use exist The above predictions of surface temperature, express The predicted surface temperature, Represents any given middle set, express Surface parameters in various parts of China The position in the middle, yes against exist The concrete model on, express use exist The above predictions of surface temperature, express The predicted surface temperature, T jk express For remote sensing of land surface temperature, RMSE represents the root mean square error of the regional average. express With T jk The difference, || represents the absolute value taken pixel by pixel. express absolute value per pixel;

[0097] P35: Select the optimal coarse set of surface parameters and the corresponding machine learning model from all regions, pixel by pixel, specifically as follows:

[0098]

[0099]

[0100] In the formula, (x,y) represents the location of the remotely sensed surface temperature pixel, κ jxy This indicates that at (x,y) Predict the area with the smallest absolute deviation in surface temperature. express When predicting the surface temperature with the smallest absolute deviation at (x,y), select The set of element positions express The minimum absolute deviation of the predicted surface temperature at (x,y). express The set of all sliding windows that traverse the sliding window and include all regions (x, y) is given, where Υ represents any given region. Chinese elements Indicating targeting The coarse set of optimal surface parameters at (x,y) express The optimal machine learning model at (x,y).

[0101] In P4, the selection of the optimal machine learning model and optimal set of surface parameters per pixel based on the competition rule of minimizing the absolute deviation between the remotely sensed surface temperature and the predicted surface temperature specifically involves:

[0102]

[0103]

[0104]

[0105] In the formula, ι represents the method subscript for predicting the surface temperature at (x,y) with the minimum absolute deviation in M, and κ represents the method subscript for predicting the surface temperature at (x,y). ιxy This indicates the area number in M ​​that has the smallest absolute deviation in surface temperature forecast at (x,y). When the surface temperature at (x,y) in the forecast is selected in M, it represents the position with the smallest absolute deviation. Location of the optimal surface parameters. This represents the coarse set of optimal surface parameters at (x,y). Let represent the optimal machine learning model at (x,y).

[0106] In P5, the step of generating a high spatial resolution predicted land surface temperature based on the optimal set of surface parameters per pixel and the optimal machine learning model specifically includes:

[0107]

[0108] In the formula, This represents a specific model that predicts the surface temperature at (x,y) with the minimum absolute deviation. express In the region κ ιxy The selected optimal coarse set of surface parameters The optimal fine set of surface parameters composed of various surface parameters in China. express use The predicted surface temperature at (x,y) with high spatial resolution.

[0109] In P6, the refinement of the high spatial resolution predicted land surface temperature using the minimum pixel-by-pixel deviation between the remotely sensed land surface temperature and the predicted land surface temperature specifically involves:

[0110]

[0111] In the formula, D represents the spatial downscaling model. This represents the minimum pixel-by-pixel deviation between the remotely sensed surface temperature and the predicted surface temperature. Indicates the use of Correction The results of remote sensing of Earth's surface temperature were refined.

[0112] For example, the spatial downscaling model D can be selected as one of polynomial interpolation, spline interpolation, or bilinear interpolation, which can be implemented using ENVI, ERDAS, or ArcGIS.

[0113] The remote sensing surface temperature refinement method based on machine learning competition proposed in this invention has the following characteristics:

[0114] 1) The data requirements are simple; only multispectral bands of remote sensing images are needed to refine the remote sensing surface temperature, and the execution requirements are easy to meet.

[0115] 2) It enhanced the differences in surface temperature with environmental changes, and the changes in surface environment revealed by remote sensing multispectral bands covered as many factors affecting surface temperature as possible.

[0116] 3) Pixel-by-pixel refinement optimization: The root mean square error of pixel-by-pixel land surface temperature prediction under thermal infrared band spatial resolution was minimized through competitive learning among machine learning methods.

[0117] 4) It has strong generalizability and applicability. With appropriate modifications to the proposed method, the detailed set of surface elements can be used to refine other surface parameters.

[0118] It should be noted that those skilled in the art can implement this invention in various other ways by modifying the construction of the fine set of surface elements, replacing the spatial scaling / descaling model, and changing the form of the set of machine learning methods. However, these embodiments without creative effort do not depart from the protection scope of this invention.

Claims

1. A method for refining remotely sensed land surface temperature based on machine learning competition, characterized in that, Includes the following steps: P1: Construct a fine set of surface parameters affecting surface temperature using high spatial resolution bands of remote sensing images, and then scale it up to a coarse set of surface parameters with thermal infrared band spatial resolution, and create a set of machine learning methods. P2: By using a sliding window to traverse and slide, the optimal coarse set of surface parameters and machine learning model for each pixel in the method set are determined through stepwise regression; P3: The optimal machine learning model and the optimal set of surface parameters per pixel are selected based on the competition rule of minimizing the absolute deviation between the remotely sensed surface temperature and the predicted surface temperature. P4: Based on the pixel-by-pixel optimal set of surface parameters and the optimal machine learning model, a high spatial resolution predicted surface temperature is generated, specifically: In the formula, This represents a specific model that predicts the surface temperature at (x, y) with the smallest absolute deviation. express In the area The optimal fine set of surface parameters composed of various surface parameters. express use The predicted surface temperature at (x, y) with high spatial resolution; P5: Refine the high spatial resolution predicted land surface temperature by utilizing the minimum pixel-by-pixel deviation between remotely sensed land surface temperature and predicted land surface temperature.

2. The method according to claim 1, characterized in that, The P1 includes the following steps: P11: A fine set of surface parameters affecting surface temperature is constructed using high spatial resolution bands of remote sensing imagery, specifically: Ω={NDVI, FCVI, AWEIsh, ATSAVI, AFRI, ARVI, BAI, BCC, BWDRVI, WDRVI, CVI, EBBI, EVI, GARI, GLI, GVMI, MCARI, NDBI, NDGI, NDPI, SAVI, MWIR, RDVI, MCARI, SARVI, MSAVI, NLI, EBSI, EMBI, ALB} In the formula, Ω represents the high spatial resolution fine set of surface parameters affecting surface temperature, NDVI represents the normalized difference vegetation index, FCVI represents the fluorescence-corrected vegetation index, AWEIsh represents the automatic water extraction index for shading elimination, ATSAVI represents the adjusted conversion soil-regulated vegetation index, AFRI represents the aerosol-free vegetation index, ARVI represents the atmospheric resistance vegetation index, BAI represents the burn area index, BCC represents the blue color coordinate, BWDRVI represents the blue wide dynamic range vegetation index, WDRVI represents the wide range dynamic vegetation index, CVI represents the chlorophyll vegetation index, EBBI represents the enhanced building exposure index, EVI represents the enhanced vegetation index, and GARI represents green atmospheric resistant vegetation. The indices are as follows: GLI (Green Leaf Index), GVMI (Gross Vegetation Moisture Index), MCARI (Modified Chlorophyll Absorption Index), NDBI (Normalized Difference Bareness Index), NDGI (Normalized Difference Greenness Index), NDPI (Normalized Difference Phenological Index), SAVI (Soil-Regulated Vegetation Index), MWIR (Modified Multispectral Water Index), RDVI (Renormalized Difference Vegetation Index), MCARI (Modified Chlorophyll Absorption Ratio Index), SARVI (Soil and Atmospheric Resistance Vegetation Index), MSAVI (Modified Soil-Adjusted Vegetation Index), NLI (Nonlinear Vegetation Index), EBSI (Enhanced Bare Soil Index), EMBI (Enhanced and Improved Bare Soil Index), and ALB (Albedo). P12: Upscale the fine set of surface parameters to a coarse set of surface parameters with spatial resolution in the thermal infrared band, specifically: ℧={ndvi, fcvi, awesome, atsavi, afri, arvi, bai, bcc, bwdrvi, wdrvi, cvi, ebbi, evi, gari, gli, gvmi, mcari, ndbi, ndgi, ndpi, savi, mwir, rdvi, mcari, sarvi, msavi, nli, ebsi, embi, alb} yes i =U(E i ) 1≤i≤I e i ∈ E i ∈Ω In the formula, U represents the spatial upscaling model, and E i Let e ​​represent the i-th fine surface parameter in Ω. i This means using U to transfer E i Spatial upscaling is a coarse surface parameter with spatial resolution in the thermal infrared band, where I represents the number of elements in Ω. Indicates by {e i } 1≤i≤I A coarse set of surface parameters; P13: The collection of methods for creating machine learning tools is as follows: M={OLS, SVR, RFR, CNR, PNN, XGB, LAS, PCR} In the formula, OLS represents least squares, SVR represents support vector regression, RFR represents random forest, CNR represents convolutional network, PNN represents perceptron neural network, XGB represents extreme gradient boosting tree, LAS represents lasso algorithm, PCR represents principal component regression, and M represents the set of machine learning methods. Let J represent the j-th element in M, where 1 ≤ j ≤ J, and J represents the number of elements in M.

3. The method according to claim 1, characterized in that, The specific steps for P2 are as follows: P21: For each element in M Perform steps P22-P25; P22: Targeting Generate a set of windows that slide across the remotely sensed surface temperature, specifically: A j ={h×w,h×(w+1),…,h×W,(h+1)×w,(h+1)×(w+1),…,(h+1)×W,…,H×W} In the formula, A j Indicating targeting The set of all sliding windows, h×w, represents the operation on remotely sensed surface temperature. The smallest window, H×W, represents the row and column of the remotely sensed surface temperature; P23: Using A j Each element in the data is traversed pixel by pixel along the row and column directions of the remotely sensed surface temperature, for example, A. j The area covered by each element in each swipe Proceed to step P24; P24: Using successive regression from 2 of I elements I - From the -1 combinations, the optimal coarse set of surface parameters is selected. The criterion is that the root mean square error of the difference between the predicted surface temperature and the remotely sensed surface temperature is the smallest. Specifically: In the formula, express exist From The optimal coarse set of surface parameters selected from the data. express Surface parameters in various parts of China The position in the middle, yes against exist The concrete model on, express use exist The above predictions of surface temperature, express The predicted surface temperature, Represents any given middle set, express Surface parameters in various parts of China The position in the middle, yes against exist The concrete model on, express use exist The above predictions of surface temperature, express The predicted surface temperature, T jk express For remote sensing of land surface temperature, RMSE represents the root mean square error of the regional average. express With T jk The difference, || represents the absolute value taken pixel by pixel. express absolute value per pixel; P25: Select the optimal coarse set of surface parameters and the corresponding machine learning model from all regions, pixel by pixel, specifically as follows: In the formula, (x, y) represents the location of the remotely sensed surface temperature pixel, κ jxy This indicates that at (x, y) Predict the area with the smallest absolute deviation in surface temperature. express When predicting the surface temperature with the smallest absolute deviation at (x, y), select... The set of element positions express The minimum absolute deviation of the predicted surface temperature at (x, y). express The set of all sliding windows that traverse the sliding window and contain all regions (x, y) is ϒ. Any given element in Indicating targeting The optimal coarse set of surface parameters at (x, y) express The optimal machine learning model at (x, y).

4. The method according to claim 1, characterized in that, Specifically, P3 refers to: In the formula, The subscript representing the method for predicting the surface temperature at (x, y) with the smallest absolute deviation in M ​​is denoted by . This indicates the area number in M ​​that has the smallest absolute deviation in surface temperature forecast at (x, y). The term "predicted surface temperature with the smallest absolute deviation" in M ​​is used when M is used. Location of the optimal surface parameters. Let (x, y) represent the coarse set of optimal surface parameters. Let represent the optimal machine learning model at (x, y).

5. The method according to claim 1, characterized in that, The P5 mentioned above is specifically: In the formula, D represents the spatial downscaling model. This represents the minimum pixel-by-pixel deviation between the remotely sensed surface temperature and the predicted surface temperature. Indicates the use of Correction The results of remote sensing of Earth's surface temperature were refined.

Citation Information

Patent Citations

  • Random forest-based multifactor remote sensing surface temperature space downscaling method

    CN107748736A

  • Method for improving water reserve resolution and precision based on machine learning spatial downscaling

    CN114218853A