A method for reconstructing hourly all-weather land surface temperature based on geostationary satellites

Abstract

The application provides a kind of based on stationary satellite hourly all-weather ground temperature reconstruction method, comprising the following steps: S1. collect ground temperature data and multi-source characteristic variable auxiliary data and construct sample data set;S2. construct temperature intra-day cycle model dynamic constraint kernel driven model, simultaneously realize spatial downscaling and ideal ground temperature reconstruction;S3. by hybrid neural network driven spatio-temporal fusion method regression and prediction, estimate weather interference and sensor system error;S4. superimpose weather interference and sensor system error on ideal ground temperature, reconstruct all-weather ground temperature.The method of the application organically combines kernel driven and spatio-temporal fusion method, realizes the reconstruction of all-weather ground temperature with high spatio-temporal resolution based on stationary satellite FY-4A ground temperature product.The method has good stability and interpretability, and the reconstruction result has high precision.

Description

Technical Field

[0001] This invention relates to the field of land surface temperature reconstruction, and more specifically, to a method for hourly, all-weather land surface temperature reconstruction based on geostationary satellites. Background Technology

[0002] Surface temperature is a core parameter reflecting the energy balance of the Earth's surface and has significant scientific value. It can be decomposed into the ideal temperature influenced by long-term or regular factors, and the unstable component caused by short-term weather disturbances. The latter causes surface temperature to exhibit significant spatiotemporal variability. Therefore, obtaining surface temperature data with high spatiotemporal resolution is of great importance. In recent years, geostationary satellites have shown great potential in high-frequency surface temperature monitoring applications. China's new-generation geostationary satellite, Fengyun-4A (FY-4A), can provide surface temperature products with a minimum interval of 15 minutes, but its spatial resolution is only about 4 km, and it suffers from a high rate of data loss due to cloud contamination, making it difficult to meet practical application needs. Therefore, there is an urgent need for a method to simultaneously improve the spatial resolution and spatial continuity of FY-4A surface temperature data to obtain all-weather surface temperature products with both high temporal and spatial resolution.

[0003] Existing spatial downscaling methods for geostationary satellites can be mainly categorized into kernel-driven algorithms and spatiotemporal fusion algorithms. The former achieves downscaling based on the assumption of scale invariance in the statistical relationship between the regression kernel and land surface temperature, allowing for flexible introduction of driving variables and thus widespread application. The latter leverages the correlation between similar pixels in the spatiotemporal neighborhood, fully utilizing the supervisory role of existing low-frequency, high-resolution land surface temperature products to achieve downscaling. All-weather land surface temperature reconstruction methods for geostationary satellites mainly include physical methods based on land surface energy balance and data fusion methods. The former offers strong physical interpretability but is complex and dependent on station observation data, making it difficult to generalize. The latter achieves missing reconstruction by organically combining various all-weather land surface temperature data, such as reanalysis data and passive microwave remote sensing data, offering high flexibility. However, most existing methods can only achieve spatial downscaling or all-weather reconstruction of geostationary satellite land surface temperature products individually, lacking methods that can simultaneously improve the spatial resolution and spatiotemporal continuity of geostationary satellite land surface temperature data to generate high-frequency, practical land surface temperature data. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method for reconstructing land surface temperature hourly and in all weather conditions based on geostationary satellites, so as to solve the problem that geostationary satellite land surface temperature products have low spatial resolution and high missing rate, making it difficult to meet the requirements of continuous land surface temperature monitoring with high spatiotemporal resolution.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] The present invention provides a method for reconstructing all-weather land surface temperature based on hourly geostationary satellite data, comprising the following steps: S1. Collecting land surface temperature data and auxiliary data of multi-source feature variables, and constructing a sample dataset; S2. Constructing a kernel-driven model with dynamic constraints of a daily temperature cycle model, simultaneously achieving spatial downscaling and ideal land surface temperature reconstruction; S3. Estimating weather interference and sensor system errors through regression and prediction using a spatiotemporal fusion method driven by a hybrid neural network; S4. Superimposing weather interference and sensor system errors onto the ideal land surface temperature to reconstruct the all-weather land surface temperature.

[0007] Optionally, in the above-mentioned method for reconstructing the all-weather land surface temperature based on hourly geostationary satellites, in step S1, the land surface temperature data includes the land surface temperature data of the Fengyun-4A geostationary satellite and the TRIMS LST all-weather land surface temperature dataset of 1km in mainland China and surrounding areas; the auxiliary data of multi-source feature variables include meteorological factor data, land surface radiation data, land surface attribute data and human activity data.

[0008] Optionally, in the above-mentioned method for reconstructing land surface temperature hourly and all-weather based on geostationary satellites, in step S1, the collected land surface temperature data and auxiliary data of multi-source feature variables are unified to a spatial resolution of 4km and 1km using an appropriate resampling method; time registration between non-same-source data is achieved based on linear interpolation; and pixel-level latitude and longitude and timestamps are extracted from the data to construct a spatiotemporal index and build a sample dataset.

[0009] Optionally, in the above method for reconstructing land surface temperature hourly and all-weather based on geostationary satellites, step S2 includes:

[0010] S2-1: Constructing a lightweight gradient boosting model at a 4km scale to characterize the relationship between land surface temperature and feature variables for regression: The relationship between ideal land surface temperature and each feature variable at a 4km scale is constructed using a quantized gradient boosting model, expressed as:

[0011]

[0012] in This indicates the FY-4A surface temperature under clear skies, selected using cloud coverage data. The surface attribute information represents the 4km spatial resolution, including normalized vegetation index, normalized water index, surface albedo, land cover type, climate zone, elevation, slope, and aspect. Information on human activity with a spatial resolution of 4km, including nighttime light and population density; This represents a spatiotemporal index, including longitude, latitude, date category variables, date values, time category variables, time sine values, and time cosine values; The regression residuals are derived from weather disturbances that cannot be explained by static variables.

[0013] S2-2. Reconstructing the Ideal 1km Land Surface Temperature: Input the feature variables at the 1km scale into the lightweight gradient boosting model at the 4km scale constructed in step S2-1 to achieve prediction. Discard the regression residuals to achieve spatial downscaling. The result at this stage is the ideal land surface temperature. This process is represented as follows:

[0014]

[0015] in, The LST reconstruction results are for ideal surface temperature, with a spatial resolution of 1 km.

[0016] S2-3. An improved six-parameter intraday temperature cycle model (GOT09) is introduced to dynamically fit the results, thereby enhancing the physical plausibility and dynamic consistency of the results over time.

[0017]

[0018] in For the final ideal LST reconstruction results, This is a set of intraday temperature cycle parameters.

[0019] Optionally, in the above-mentioned method for reconstructing the all-weather surface temperature based on geostationary satellites hourly, in step S3, based on the spatiotemporal fusion method, a hybrid neural network model is constructed at the reference time using low-frequency, high spatial resolution all-weather data, meteorological factors, and radiation information as auxiliary data. The characteristic variables of the target time are input into the hybrid neural network model to estimate weather interference and sensor system errors.

[0020] Optionally, in the above method for reconstructing land surface temperature hourly and all-weather based on geostationary satellites, step S3 includes:

[0021] S3-1. At the reference time, calculate the sensor system error and weather interference between the low temporal resolution, high spatial resolution all-weather surface temperature auxiliary data and the ideal surface temperature constructed in S2. The formula is expressed as:

[0022]

[0023] in, and The images show the surface temperature at the same time, obtained from different sensors. It is a systematic error;

[0024] Introducing weather disturbances, at the reference time, the low temporal resolution, high spatial resolution all-weather auxiliary surface temperature data and the ideal surface temperature constructed by S2 exhibit the following relationship:

[0025]

[0026] in Indicates reference time. This provides low-frequency, high spatial resolution, all-weather land surface temperature data. The ideal surface temperature constructed in step S2, It is a systematic error. Due to weather disturbances;

[0027] S3-2. At the reference time, construct a hybrid neural network model representing weather interference, sensor system errors, and feature variables to achieve regression: Utilize a hybrid neural network (CNN) and a self-attention deep learning model (Transformer) to construct the relationship between feature parameters and sensor system errors and weather interference:

[0028]

[0029] in, Meteorological factors include cloud cover, fog monitoring, 2m dew point temperature, 2m air temperature, soil temperature, snow cover, total evaporation, 10m U / V component wind, surface air pressure, and precipitation. It is radiation information, including incident solar radiation at the ground and long-wave radiation rising from the Earth's surface. It represents land surface attribute information, including normalized difference vegetation index, normalized difference water index, surface albedo, land cover type, climate zone, elevation, slope, and aspect; To supplement surface properties with parameters that represent soil type; Information indicating human activity, including nighttime light and population density; This represents a spatiotemporal index, including longitude, latitude, date category variables, date values, time category variables, time sine values, and time cosine values;

[0030] S3-3. Based on the principle of spatiotemporal fusion, Input the characteristic variables of each time point into the model to estimate sensor system errors and weather interference at other times of the day:

[0031] ,

[0032] in, Indicates other times of the day.

[0033] Optionally, in the above method for reconstructing all-weather land surface temperature based on hourly geostationary satellites, step S4 includes: superimposing the ideal land surface temperature constructed in S2 with the system error and weather interference estimated in S3 to achieve all-weather land surface temperature reconstruction, expressed as:

[0034]

[0035] in, This is a reconstructed surface temperature result for all weather conditions, with a spatial resolution of 1 km and a temporal resolution of 1 hour. The ideal surface temperature constructed in step S2, It is a systematic error. Due to weather disturbances, Indicates other times of the day.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] This invention presents an hourly all-weather land surface temperature reconstruction method based on geostationary satellites. This method organically combines kernel-driven and spatiotemporal fusion approaches to reconstruct high spatiotemporal resolution all-weather land surface temperatures using FY-4A geostationary satellite land surface temperature products. The method exhibits good stability and interpretability, resulting in high-accuracy reconstruction results. Furthermore, this invention incorporates a deep learning model with interpretability, demonstrating good transferability and generalization ability alongside high accuracy. Therefore, it can achieve high-accuracy reconstruction results in any region and time, making it highly practical. Attached Figure Description

[0038] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below.

[0039] Figure 1 This is a flowchart of the steps of the method for reconstructing the Earth's surface temperature hourly and all-weather based on geostationary satellites according to the present invention;

[0040] Figure 2 This is a flowchart of the invention for reconstructing high spatiotemporal resolution all-weather surface temperature by combining kernel-driven and spatiotemporal fusion methods;

[0041] Figure 3 It refers to the accuracy of the reconstruction results at different ground observation stations;

[0042] Figure 4 This is the reconstruction result in the test area. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0044] The present invention discloses an hourly all-weather land surface temperature reconstruction method based on geostationary satellites, comprising the following steps: S1. Collecting land surface temperature data and multi-source auxiliary data of characteristic variables, and constructing a sample dataset; S2. Constructing a kernel-driven model with dynamic constraints of a diurnal temperature cycle (DTC) model to simultaneously achieve spatial downscaling and ideal land surface temperature reconstruction; S3. Estimating weather interference and sensor system errors through a spatiotemporal fusion method driven by a hybrid neural network; S4. Superimposing weather interference and sensor system errors onto the ideal land surface temperature to reconstruct the all-weather land surface temperature. This invention combines kernel-driven and spatiotemporal fusion methods, introduces multi-source auxiliary data, and achieves high spatiotemporal resolution all-weather land surface temperature reconstruction based on FY-4A geostationary satellite land surface temperature products. The method exhibits good stability and interpretability, and the reconstruction results have high accuracy.

[0045] like Figure 1 and Figure 2 As shown, the method for reconstructing the Earth's surface temperature hourly and all-weather based on geostationary satellites according to the present invention includes the following steps:

[0046] S1. Collect surface temperature data and auxiliary data of multi-source characteristic variables, and construct a sample dataset;

[0047] The surface temperature data includes surface temperature data from the Fengyun-4A geostationary satellite (FY-4A) and the TRIMS LST dataset, a 1-kilometer all-weather surface temperature dataset for mainland China and surrounding areas. Multi-source auxiliary data includes meteorological data, surface radiation data, surface attribute data, and human activity data. Meteorological data include cloud cover and fog monitoring products from FY-4A, as well as datasets from the European Centre for Medium-Range Weather Forecasts' Generation 5 Land Reanalysis Data (ERA5-Land) for 2-meter air temperature, 2-meter dew point temperature, surface pressure, precipitation, snow cover, soil temperature, 10-meter wind speed, and total evaporation. Surface radiation data include FY-4A surface incident solar radiation products and surface ascending longwave radiation products (ELITE). Surface attribute data include Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) calculated based on the Moderate Resolution Imaging Spectroradiometer (MODIS) surface reflectance product (MCD43A4), as well as the reflectance from the MODIS surface albedo product (MCD43A3), the China Land Cover Dataset (CLCD), Köppen-Geiger Climate Classification Maps, and the Space Shuttle Radar Topography Mission Digital Elevation Model (SRTM). The datasets include DEM and ERA5-Land soil type datasets; human activity data include the nighttime light dataset (NPP-VIIRS-like NLT) and the population density dataset (WorldPop).

[0048] All data (collected surface temperature data and auxiliary data of multi-source feature variables) were unified to 4km and 1km spatial resolution using appropriate resampling methods; temporal registration between non-similar data was achieved based on linear interpolation; and pixel-level latitude and longitude and timestamps were extracted from the data to construct a spatiotemporal index and build a sample dataset.

[0049] S2. Construct a kernel-driven model with dynamic constraints for the daily temperature cycle (DTC) model to simultaneously achieve spatial downscaling and ideal surface temperature reconstruction;

[0050] Based on the kernel-driven method, and using various static feature variables as auxiliary parameters, combined with the dynamic constraints of the intraday temperature cycle model, a lightweight gradient boosting model with a 4km scale is constructed to achieve regression. Feature variables with a 1km scale are input into the constructed 4km scale lightweight gradient boosting model to achieve prediction, simultaneously realizing spatial downscaling and ideal surface temperature reconstruction. Specifically, step S2 includes:

[0051] S2-1. Constructing a lightweight gradient boosting model at a 4km scale to characterize the relationship between land surface temperature and feature variables for regression: Based on the premise that ideal land surface temperature is closely related to land surface properties and human activities, and that this relationship changes with time and spatial location, a lightweight gradient boosting model is used to construct the relationship between ideal land surface temperature and various feature variables at a 4km scale, expressed as:

[0052]

[0053] in This indicates the FY-4A surface temperature under clear skies, selected using cloud coverage data. The surface attribute information represents the 4km spatial resolution, including normalized vegetation index, normalized water index, surface albedo, land cover type, climate zone, elevation, slope, and aspect. Information on human activity with a spatial resolution of 4km, including nighttime light and population density; This represents a spatiotemporal index, including longitude, latitude, date category variables, date values, time category variables, time sine values, and time cosine values; The regression residuals are derived from weather disturbances that cannot be explained by static variables.

[0054] Based on this, a lightweight gradient boosting model was selected as the driving tool, and five-fold cross-validation was used to improve stability and generalization ability to construct a model representing the relationship between surface temperature and feature variables at a 4km scale. To ensure feature coverage and sufficient samples, for each target day, data from that day and the day before and after it were used for model training. If the data from the day before and after was incomplete, alternative days were searched within two days before and after the target day.

[0055] S2-2. Reconstructing the ideal surface temperature at 1km: Input the feature variables at the 1km scale into the lightweight gradient boosting model at the 4km scale constructed in step S2-1 to achieve prediction, discard the regression residuals, and achieve spatial downscaling.

[0056] Based on the scale invariance assumption of the kernel-driven method, the relationship model between land surface temperature and feature variables at the 4km scale constructed in step S2-1 remains applicable at the 1km scale. The 4km scale model constructed in step S2-1 is then applied to the 1km resolution feature variables of the target day. Specifically, the 1km resolution feature variables of the target day are input into the lightweight gradient boosting model at the 4km scale constructed in step S2-1 to achieve prediction, discarding the regression residuals to achieve spatial downscaling. Since this stage is supported by clear-sky land surface temperature and static variable data, the result represents the ideal land surface temperature. This process can be expressed as:

[0057]

[0058] in, The LST reconstruction results are for ideal surface temperature, with a spatial resolution of 1 km.

[0059] S2-3. Introducing Dynamic Constraints from the Intra-Day Temperature Cycle Model to Reconstruct Results. Although machine learning models have integrated multiple features, their non-parametric and weakly structured constraints may lead to unnatural jumps in results at the intra-day scale. To enhance the physical plausibility and dynamic consistency of the results over time, an improved six-parameter intra-day temperature cycle model (GOT09) is introduced to dynamically fit and constrain the results, thereby improving their physical plausibility and dynamic consistency over time.

[0060]

[0061] in To achieve the final ideal surface temperature reconstruction result, This is a set of intraday temperature cycle parameters.

[0062] S3. Estimate weather interference and sensor system errors through regression and prediction using a spatiotemporal fusion method driven by a hybrid neural network.

[0063] Based on a spatiotemporal fusion method, using low-frequency, high spatial resolution all-weather data, meteorological factors, radiation information, and other characteristic data as auxiliary data, a hybrid neural network model is constructed at a reference time. The characteristic variables at the target time are input into the model to estimate weather interference and sensor system errors. Specifically, step S3 includes:

[0064] S3-1. At the reference time, based on the low temporal resolution, high spatial resolution all-weather surface temperature auxiliary data and the ideal surface temperature constructed in S2, calculate the sensor system error and weather interference between the low temporal resolution, high spatial resolution all-weather surface temperature auxiliary data and the ideal surface temperature constructed in S2. The core assumption of the spatiotemporal fusion method is that there is only a system error caused by sensor differences between non-homogeneous satellite images, expressed by the formula:

[0065]

[0066] in, and The same time LST obtained from different sensors, It is a systematic error.

[0067] Based on this, weather disturbances are introduced, and at the reference time, the low temporal resolution, high spatial resolution all-weather surface temperature auxiliary data and the ideal surface temperature constructed by S2 have the following relationship:

[0068]

[0069] in Indicates reference time. This provides low-frequency, high spatial resolution, all-weather land surface temperature data. The ideal surface temperature constructed in step S2, It is a systematic error. This is due to weather disturbances. Based on this, the sensor system error and weather interference between the two reference times can be calculated.

[0070] S3-2. At the reference time, construct a hybrid neural network model representing weather disturbances, sensor system errors, and characteristic variables to achieve regression. Weather disturbances are driven by meteorological factors, surface radiation, etc., and these disturbance effects and sensor system errors exhibit relative stability under similar surface properties and human activity backgrounds. A convolutional neural network (CNN) is used... Figure 2 The CNN module in the deep learning model (Transformer) + self-attention mechanism. Figure 2 The Transformer module in the dataset uses a hybrid neural network to construct the relationship between feature parameters and sensor system errors and weather interference.

[0071]

[0072] in, Meteorological factors include cloud cover, fog monitoring, 2m dew point temperature, 2m air temperature, soil temperature, snow cover, total evaporation, 10m U / V component wind, surface air pressure, and precipitation. It is radiation information, including solar radiation incident on the ground and long-wave radiation rising from the Earth's surface; It represents land surface attribute information, including normalized vegetation index, normalized water index, surface albedo, land cover type, climate zone, elevation, slope, and aspect; To supplement surface properties with parameters that represent soil type; Information indicating human activity, including nighttime light and population density; This represents a spatiotemporal index, including longitude, latitude, date category variables, date values, time category variables, time sine values, and time cosine values.

[0073] S3-3. Estimate weather interference and sensor system errors at the target time. Based on the principle of spatiotemporal fusion, the disturbance estimation model established based on the reference time will also be applicable at other nearby times. Therefore, by inputting the characteristic variables of the target time, the weather interference and sensor system errors at the target time can be extrapolated. That is, based on the principle of spatiotemporal fusion, the weather interference and sensor system errors at the target time can be extrapolated. The characteristic variables at each moment are input into the model to estimate sensor system errors and weather interference at other times of the day, in order to achieve prediction:

[0074] ,

[0075] in, Indicates other times of the day;

[0076] S4. Superimpose weather interference and sensor system errors onto the ideal surface temperature to reconstruct the all-weather surface temperature (1km, hourly). Specifically, this includes:

[0077] By superimposing the ideal surface temperature constructed by S2 with the systematic errors and weather disturbances estimated by S3, an all-weather surface temperature reconstruction can be achieved, which can be expressed as:

[0078]

[0079] in This is a reconstructed surface temperature result for all weather conditions, with a spatial resolution of 1 km and a temporal resolution of 1 hour. The ideal surface temperature constructed in step S2, It is a systematic error. Due to weather disturbances, This indicates other times of the day. This invention integrates spatial downscaling and all-weather reconstruction methods to achieve hourly, 1 km-level all-weather surface temperature reconstruction based on the FY-4A geostationary satellite.

[0080] Figure 3Figures A through E show the overall accuracy of the 2020 reconstruction results at the five Heihe stations. Ground observation validation shows that the method of this invention has high accuracy, with RMSE of 2.42–4.08 K, MAE of 1.9–3.08 K, and R... 2 With a value of 0.93-0.96, it can meet the needs of practical applications.

[0081] Figure 4 Figures A through E show the reconstruction results of the test area on November 6, 2020. The reconstruction results are highly consistent with the spatial distribution and temporal variation pattern of surface temperature in FY-4A, supplementing the spatial information caused by cloud pollution, and significantly enhancing spatial information. Because the method of this invention incorporates a deep learning model and is interpretable, it not only has high accuracy but also good transferability and generalization ability. Therefore, it can achieve high-accuracy reconstruction results in any region and time, making it highly practical.

[0082] The above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and are not intended to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or improve the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in the present invention; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for reconstructing land surface temperature hourly and all-weather based on geostationary satellites, characterized in that, Includes the following steps: S1. Collect surface temperature data and auxiliary data of multi-source characteristic variables, and construct a sample dataset; S2. Construct a kernel-driven model with dynamic constraints for the intraday temperature cycle model, and simultaneously achieve spatial downscaling and ideal surface temperature reconstruction; S3. Estimate weather interference and sensor system errors by using a spatiotemporal fusion method driven by a hybrid neural network for regression and prediction; S4. The weather interference and the sensor system error are superimposed on the ideal surface temperature to reconstruct the all-weather surface temperature. Step S2 includes: S2-1: Constructing a lightweight gradient boosting model at a 4km scale to characterize the relationship between land surface temperature and feature variables for regression: The relationship between ideal land surface temperature and each feature variable at a 4km scale is constructed using a quantized gradient boosting model, expressed as: in This indicates the surface temperature of the Fengyun-4A geostationary satellite (FY-4A) under clear skies, selected using cloud coverage data. The surface attribute information represents the 4km spatial resolution, including normalized vegetation index, normalized water index, surface albedo, land cover type, climate zone, elevation, slope, and aspect. Information on human activity with a spatial resolution of 4km, including nighttime light and population density; This represents a spatiotemporal index, including longitude, latitude, date category variables, date values, time category variables, time sine values, and time cosine values; The regression residuals are derived from weather disturbances that cannot be explained by static variables. S2-2. Reconstructing the Ideal 1km Land Surface Temperature: Input the feature variables at the 1km scale into the lightweight gradient boosting model at the 4km scale constructed in step S2-1 to achieve prediction. Discard the regression residuals to achieve spatial downscaling. The result at this stage is the ideal land surface temperature. This process is represented as follows: in, The LST reconstruction results are for ideal surface temperature, with a spatial resolution of 1 km. S2-3. An improved six-parameter intraday temperature cycle model is introduced to dynamically fit the results, thereby enhancing the physical rationality and dynamic consistency of the results over time: in For the final ideal LST reconstruction results, This is a set of intraday temperature cycle parameters. Step S3 includes: S3-1. At the reference time, calculate the sensor system error and weather interference between the low temporal resolution, high spatial resolution all-weather surface temperature auxiliary data and the ideal surface temperature constructed in step S2. The formula is expressed as: in, and The images show the surface temperature at the same time, obtained from different sensors. It is a systematic error; Introducing weather disturbances, at the reference time, the low temporal resolution, high spatial resolution all-weather auxiliary surface temperature data and the ideal surface temperature constructed in step S2 have the following relationship: in Indicates reference time. This provides low-frequency, high spatial resolution, all-weather land surface temperature data. The ideal surface temperature constructed in step S2, It is a systematic error. Due to weather disturbances; S3-2. At the reference time, construct the hybrid neural network model representing weather interference, sensor system error, and feature variables to achieve regression: Construct the relationship between feature parameters and sensor system error and weather interference using a hybrid neural network combining a convolutional neural network (CNN) and a self-attention mechanism deep learning model. in, Meteorological factors include cloud cover, fog monitoring, 2m dew point temperature, 2m air temperature, soil temperature, snow cover, total evaporation, 10m U / V component wind, surface air pressure, and precipitation. It is radiation information, including incident solar radiation on the ground and long-wave radiation rising from the Earth's surface; It represents land surface attribute information, including normalized difference vegetation index, normalized difference water index, surface albedo, land cover type, climate zone, elevation, slope, and aspect; To supplement surface properties with parameters that represent soil type; Information indicating human activity, including nighttime light and population density; This represents a spatiotemporal index, including longitude, latitude, date category variables, date values, time category variables, time sine values, and time cosine values; S3-3. Based on the principle of spatiotemporal fusion, Input the characteristic variables of each time point into the model to estimate sensor system errors and weather interference at other times of the day: ， in, Indicates other times of the day.

2. The method for reconstructing land surface temperature hourly and all-weather based on geostationary satellites according to claim 1, characterized in that, in, In step S1, the surface temperature data includes surface temperature data from the Fengyun-4A geostationary satellite and the TRIMS LST dataset of 1km all-weather surface temperature data for mainland China and surrounding areas; the auxiliary data of multi-source feature variables includes meteorological factor data, surface radiation data, surface attribute data, and human activity data.

3. The method for reconstructing land surface temperature hourly and all-weather based on geostationary satellites according to claim 1, characterized in that, In step S1, the collected surface temperature data and the auxiliary data of the multi-source feature variables are unified to a spatial resolution of 4km and 1km using an appropriate resampling method; time registration between non-same-source data is achieved based on linear interpolation; and pixel-level latitude and longitude and timestamps are extracted from the data to construct a spatiotemporal index and build the sample dataset.

4. The method for reconstructing land surface temperature hourly and all-weather based on geostationary satellites according to claim 1, characterized in that, In step S3, based on the spatiotemporal fusion method, a hybrid neural network model is constructed at the reference time using low-frequency, high spatial resolution all-weather data, meteorological factors, and radiation information as aids. The target time feature variables are input into the hybrid neural network model to estimate the weather interference and the sensor system error.

5. The method for reconstructing land surface temperature hourly and all-weather based on geostationary satellites according to claim 1, characterized in that, Step S4 includes: superimposing the ideal surface temperature constructed in step S2 with the system error and weather interference estimated in step S3 to achieve all-weather surface temperature reconstruction, expressed as: in, This is a reconstructed surface temperature result for all weather conditions, with a spatial resolution of 1 km and a temporal resolution of 1 hour. The ideal surface temperature constructed in step S2, It is a systematic error. Due to weather disturbances, Indicates other times of the day.

Images