A physical model and data-driven based all-weather land surface temperature reconstruction method

Abstract

The scheme discloses a kind of all-weather ground surface temperature reconstruction method based on physical model and data driving, the method considers the explainability and strong extrapolation ability of physical model, and data driving method is strong to data adaptability and data feature mining ability, will be combined with flexible data driving modeling to reconstruct LST to generate LST source data using physical modeling, improve accuracy using data driving method, can generate a kind of spatial resolution higher all-weather similar MODIS data, make up the shortcoming that satellite TIR image cannot obtain complete LST due to cloud pollution.The scheme will reconstruct model be established on the basis of the authenticity and science of LST source data, then use multisource remote sensing data and image spatial information to improve LST precision, finally can obtain high-precision all-weather LST image.

Description

Technical Field

[0001] This invention belongs to the field of land surface temperature reconstruction technology, and in particular relates to an all-weather land surface temperature reconstruction method based on physical models and data-driven approaches. Background Technology

[0002] Currently, the most commonly used technology for acquiring land surface temperature (LST) is thermal infrared (TIR) ​​remote sensing (RS). However, this method of acquiring LST imagery via satellite is susceptible to cloud interference, limiting the complete spatiotemporal coverage of LST that satellite TIR sensors can provide. Studies have shown that more than 50% of the data is invalid due to cloud contamination, and this is especially evident during the day, greatly reducing the value of LST remote sensing applications.

[0003] Reconstructing cloud-free, all-weather LST remains a challenging problem, urgently requiring in-depth research. To address this issue, scholars have proposed a series of reconstruction methods, which can be mainly categorized into four types: gap-filling method, passive microwave measurement method, data-driven method, and physical model method.

[0004] The gap-filling method primarily relies on the first law of geography, "proximity and similarity," to derive cloud-contaminated pixel values ​​through substitution or interpolation using spatiotemporally adjacent pixel information. For example, Kriging interpolation and inverse distance weighted interpolation can be used to reconstruct the LST (Low Sky Spectrum). This method is based on spatiotemporally adjacent clear-sky pixels, therefore it can only derive the hypothetical clear-sky LST, not the actual LST under clouds.

[0005] Passive microwave (PMW) measurements can directly acquire all-weather LST (Local ST) information because PMWs can penetrate clouds. However, estimating LST using microwaves has limitations. First, microwave signals vary significantly with surface characteristics. Second, the spatial resolution of LST measured by PMW (0.25° to 0.1°) is much coarser than that of TIR (<1 km), and the accuracy of LST retrieved by PMW (approximately 5-6 K) is worse than that retrieved by TIR (approximately 1-2 K). Therefore, this method is generally used as supplementary information to TIR or other data, rather than as primary information.

[0006] With the development of artificial intelligence technology, data-driven methods have begun to be applied to LST reconstruction research. This method starts with initial data or observations and uses tools such as machine learning to find and establish relationships between internal features, thereby achieving predictive patterns. Data-driven LST reconstruction methods have achieved remarkable success in terms of efficiency and accuracy. However, this method focuses on data mining and neglects physical laws, thus often failing to explain causal relationships. This lack of transparency in the mechanism and the inability to explain causal relationships has led to some controversy and skepticism.

[0007] Physical modeling methods start from physical laws and construct expressions relating observed data to target parameters. Examples include the surface energy balance method and various methods based on it. The advantage of this method is that there is a clear causal relationship between observed data and target parameters, and the model building process is transparent and interpretable. However, physical modeling methods struggle to accurately describe complex and ever-changing real-world situations. The models are relatively complex, and some input variables are difficult to obtain and calculate, introducing uncertainty into the modeling process and requiring improvement in accuracy.

[0008] Among the aforementioned LST reconstruction methods, data-driven and physical model methods are the most widely used and technically mature. While data-driven methods can reconstruct LST, the "black box" problems, such as lack of theoretical support and weak interpretability, remain unresolved. Unlike data-driven methods, physical models are based on knowledge and deduction, containing rich underlying system knowledge, resulting in strong interpretability. However, existing physical models require numerous input parameters, some of which are difficult to obtain, leading to significant uncertainty in parameterization and consequently, considerable uncertainty in the physical model results. Furthermore, many parameters in existing physical models require assumed meteorological conditions as a premise for simulating LST, resulting in lower accuracy in the simulated LST. In addition, although physical models output all-weather data, their spatial detail is insufficient (e.g., insufficient representation of LST differences in different locations, elevation distribution, and the contours of ridges and valleys), requiring further improvement in spatial feature distribution and pixel accuracy compared to remote sensing satellite imagery data. Summary of the Invention

[0009] The purpose of this invention is to address the aforementioned problems by providing a method for reconstructing all-weather land surface temperature (LST) based on a physical model and data-driven approaches. This solution considers the strong interpretability and extrapolation capabilities of physical models, while data-driven methods possess strong data adaptability and the ability to mine data features. It combines physical modeling with flexible data-driven modeling to reconstruct LST. The physical model generates LST source data, and the data-driven approach improves accuracy, forming a model framework that uses data-driven methods to assist the physical model. The specific solution is as follows:

[0010] A physical model- and data-driven method for reconstructing all-weather land surface temperature includes the following steps:

[0011] S1. Simulate the target day's all-weather LST with the same resolution as the MODIS LST using a physical model;

[0012] S2. Perform ellipsoidal transformation on the target day all-weather LST to obtain the transformed target day all-weather WRF LST; here WRF LST is only an abbreviation for the transformed target day all-weather, and does not mean that the target day all-weather is obtained based on the WRF model, nor does it mean that the physical model must use the WRF model. If other types of physical models are used to implement this scheme, they should also be within the protection scope of this scheme.

[0013] The effective pixels of S3.MODIS LST are used as the dependent variable y, and the pixels of WRF LST corresponding to the dependent variable y in the time space are used as the independent variables x. Each independent variable x and the corresponding dependent variable y are grouped together. Multiple sets of data are divided into training set and validation set. The training set is used to train the data model to obtain the trained target day data model. The validation set is used to validate the trained target day data model.

[0014] S4. Input the transformed target day all-weather WRF LST as the independent variable X into the target day data model. The target day data model outputs the target day all-weather reconstructed LST. The independent variable X contains all the independent variable pixels x of the corresponding image.

[0015] In the above-mentioned all-weather surface temperature reconstruction method based on physical models and data-driven methods, in steps S3 and S4, the spatial neighboring pixels of the target pixel to be trained / predicted are used as spatial characteristic information of the target pixel and input into the data model to be trained or the target daily data model that has already been trained.

[0016] In the above-mentioned all-weather surface temperature reconstruction method based on physical model and data-driven approach, in step S3, the pixels of the dependent variable y corresponding to the time and space of multi-source remote sensing data are simultaneously used as independent variables x.

[0017] In step S4, multi-source remote sensing data are simultaneously input as independent variable X into the target day data model.

[0018] In the aforementioned all-weather surface temperature reconstruction method based on physical models and data-driven methods, the multi-source remote sensing data includes elevation DEM, latitude LAT, slope SLOPE, and normalized difference vegetation index (NDVI).

[0019] In the above-mentioned all-weather land surface temperature reconstruction method based on physical model and data drive, if there are enough effective pixels of MODIS LST for the target day, then step S4 is executed to obtain the all-weather reconstructed LST for the target day.

[0020] If there are not enough valid pixels in the MODIS LST for the target date, step S5 is included after step S4:

[0021] Several trained data models for adjacent times of the target day are obtained as time-adjacent models. The independent variable X of the target day is input into several time-adjacent models respectively. The outputs of all time-adjacent models are summed according to the set weights to obtain the all-weather reconstructed LST of the target day.

[0022] In the above-mentioned all-weather surface temperature reconstruction method based on physical models and data, when the number of effective pixels of MODIS LST on the target day is greater than the set pixel value, it is considered that there are enough effective pixels, and step S4 is not executed, but S5 is executed directly; the set pixel value is determined by those skilled in the art according to the specific circumstances.

[0023] Alternatively, if the correlation coefficient between the target date MODIS LST and the reconstructed LST obtained in step S4 is less than the set value, it is considered that there are not enough effective pixels, and step S5 continues. The correlation coefficient can be calculated using the Pearson correlation coefficient formula. Pixels from the MODIS LST and the reconstructed LST are extracted one-to-one to form two sets of data. The correlation coefficient between these two sets of data can be calculated using the Pearson correlation coefficient formula, which is the correlation coefficient of this scheme. The set value of the coefficient is determined by those skilled in the art based on specific circumstances.

[0024] In the above-mentioned all-weather land surface temperature reconstruction method based on physical models and data-driven approach, when there are not enough effective pixels in the target day's MODISLST, two trained data models from the day before and the day after the target day are obtained as time-adjacent models. The independent variable X of the target day is input into the two time-adjacent models respectively, and the outputs of the two time-adjacent models are summed according to the set weights to obtain the reconstructed LST of the target day.

[0025] In the aforementioned all-weather land surface temperature reconstruction method based on physical models and data-driven approaches, when there are a sufficient number of effective pixels for the target day's MODISLST, the calculation method for the data-driven part is as follows:

[0026] f(X)=h(X) T β+Γ(X) (1)

[0027] LST PW-DGRS =f(LST) WRF (2)

[0028] Among them, LST PW-DGRS For the target date, LST will be rebuilt around the clock. WRF DEM, LAT, SLOPE, and NDVI are independent variables, f is the data model, and h(X) is the variable name. T Let Γ(X) be a basis function, β be a parameter vector, and Γ(X) be a Gaussian process function.

[0029] When there are not enough valid pixels in the MODIS LST for the target date, the calculation method for the data-driven part is as follows:

[0030]

[0031]

[0032] Among them, LST PW-DDRS For the all-weather reconstruction of the LST for the target day, f is the time-adjacent model for the target day, N is the number of time-adjacent models, and R is the time-adjacent MODIS LST and its corresponding LST. PW-DDRS The correlation coefficient, where T is the number of days between image dates, and LST. WRF ρ is the target date WRF LST, ρ is the correlation coefficient between the target date MODIS LST and LST WRF, and ω is the time-adjacent model weight.

[0033] In the above-mentioned all-weather land surface temperature reconstruction method based on physical models and data, the physical model is a WRF model with six physical schemes: microphysics, cloud layer, shortwave radiation, longwave radiation, boundary layer, and land surface processes, and the urban canopy physical scheme is added to the WRF model.

[0034] The data model described is a machine regression algorithm, GPR model;

[0035] In step S1, the physical model adopts a three-layer unidirectional nesting, and the spatial resolutions of the three layers of unidirectional nesting are 9KM, 3KM and 1KM respectively, so as to simulate the target day all-weather LST with a resolution of 1KM. The spatial resolution of the MODIS LST image is 1KM.

[0036] The physical model simulates the all-weather LST for the target day based on geographic static field data, historical reanalysis data, and physical schemes;

[0037] The input to the physical model includes study area parameters that specify the nesting method, projection position, and projection method.

[0038] In the above-mentioned all-weather land surface temperature reconstruction method based on physical models and data-driven methods, in step S2, the target day's all-weather LST is ellipsoidally transformed to obtain the transformed target day's all-weather WRF LST using the following method:

[0039] S21. Create a coordinate system reference sphere of a perfect sphere;

[0040] S22. Based on the coordinate system of a sphere, the four boundaries of the physical model's planar projection are calculated using a reference sphere. The four boundaries include the coordinate values ​​of the four corners of the target day's all-weather LST image.

[0041] S23. Convert these four coordinates into geographic coordinate values ​​of a coordinate system reference sphere that is a perfect sphere, and then convert these four geographic coordinate values ​​to a WGS84 coordinate reference sphere;

[0042] S24. Resample the all-weather LST image of the target day according to the affine matrix to obtain the LST data under the WGS84 reference sphere, that is, the converted all-weather WRF LST of the target day.

[0043] The advantages of this invention are:

[0044] 1. By combining physical modeling with flexible data-driven modeling, a model framework of data-driven assisted physical model is formed. A new all-weather LST reconstruction model is proposed. The physical model generates LST source data, and the data-driven method is used to improve accuracy. It can generate a high spatial resolution all-weather MODIS-like data, which makes up for the shortcomings of satellite TIR imagery that cannot obtain complete LST due to cloud contamination.

[0045] 2. The physical model of this scheme requires only a few parameters and is based on historical reanalysis data from real observations, while using geographic static field data as an aid, so the simulated LST has high accuracy.

[0046] 3. When using data-driven methods to improve accuracy, introduce some multi-source remote sensing auxiliary data that may affect LST changes as the driving force of data-driven methods, so as to give full play to the data mining capabilities of data-driven methods and further improve accuracy.

[0047] 4. The reconstruction model of this scheme does not only consider pixel-by-pixel point-to-point fitting, but also takes the spatial information of pixels as a driving force, so that the reconstructed LST has obvious spatial continuity characteristics and can restore the spatial distribution details that the LST should have.

[0048] 5. The commonly used physical method is the surface energy balance method. This method requires a large amount of input parameters, and the input data is difficult to obtain. Although it can acquire all-weather LST, its accuracy needs improvement. Other non-physical methods either use spatiotemporal interpolation or data-driven data fusion, which can only generate pixel values ​​under assumed conditions, not true pixel values. Moreover, when the cloud pollution area is large, the reconstructed LST error will become significant. This scheme's reconstruction model is based on the authenticity and scientific validity of the LST source data. Then, it uses multi-source remote sensing data and image spatial information to improve LST accuracy, thus ultimately obtaining high-precision all-weather LST imagery. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the reconstruction process of the all-weather surface temperature reconstruction method based on physical models and data-driven approaches of this invention.

[0050] Figure 2 This is a schematic diagram illustrating the extraction and use of pixel spatial information in the all-weather surface temperature reconstruction method based on physical models and data-driven methods of the present invention.

[0051] Figure 3 This is a comparison chart of air temperature simulated by the WRF physical model in this invention and the temperature at meteorological stations;

[0052] Figure 4a This refers to the WRF output results of the physical model before coordinate transformation, without using the ellipsoid transformation method proposed in this invention.

[0053] Figure 4b This is the WRF output result after coordinate transformation of the physical model in this invention;

[0054] Figure 5 All-weather LST reconstructed from MODIS LST and corresponding date PW-DGRS model under cloud-sparse scenario;

[0055] Figure 6 To reduce the error in reconstructing cloud-contaminated pixels using the PW-DGRS model under low-cloud conditions;

[0056] Figure 7 The error in reconstructing all pixels of the PW-DGRS model under the scenario of low cloud cover;

[0057] Figure 8 All-weather LST reconstructed from MODIS LST and corresponding date PW-DGRS models under multiple cloud scenarios

[0058] Figure 9 Errors in reconstructing cloud-contaminated pixels using the PW-DGRS model under multiple cloud scenarios;

[0059] Figure 10 Reconstruct the MAE, RMSE, and ρ of the all-weather LST in Tianjin in December 2020 using the PW-DGRS model. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0061] This solution for rebuilding the 24 / 7 LST process mainly consists of two parts: (as follows) Figure 1As shown, firstly, a preliminary all-weather LST for the target day is simulated using a physical model with geographic static field data and real historical reanalysis data (FNL data), along with a physical scheme. The preferred physical model for this scheme is the next-generation mesoscale weather research and forecasting model (WRF), which has multiple modular physical schemes and can describe various complex weather phenomena, simulating all-weather LST with high spatiotemporal resolution. Then, the all-weather LST for the target day is ellipsoidally transformed to WGS84 to obtain the transformed all-weather WRF LST for the target day. The transformed all-weather WRF LST corresponds one-to-one with MODIS LST pixels. Finally, some multi-source remote sensing data (elevation DEM, slope SLOPE, latitude LAT, and normalized difference vegetation index NDVI) that affect LST changes are introduced as auxiliary data for LST reconstruction. The effective pixels of MODIS LST (i.e., pixels not contaminated by clouds) are used as the dependent variable y. The pixels corresponding to the dependent variable y in time and space in WRF LST and multi-source remote sensing data are used as independent variables x. Each independent variable x and its corresponding dependent variable y form a pair; that is, the number of xy combinations equals the number of effective pixels. Multiple data sets are divided into training and validation sets. The training set is used to train the data model for the target day, and the validation set is used to validate the trained target day data model. Specifically, MODIS LST and WRF LST data from the same time and space on the target day are used for training and validation. The training and validation sets are divided in a certain ratio, generally 7:3. Due to unavoidable errors, a certain amount of temporal and spatial error is allowed within the same time and space, provided it does not affect the training data model. The preferred data model is the Gaussian Regression (GPR) model, hereinafter referred to as the GPR model. GPR has recently proven to be a sophisticated image pixel gap-filling processor, capable of filling gaps in optical images, making it very suitable for processing satellite images into all-weather product data. Furthermore, as... Figure 2 As shown, the spatial neighboring pixels of the target pixel to be trained / predicted are used as spatial characteristic information of the target pixel and input into the data model to be trained or the already trained target day data model to obtain highly continuous reconstructed pixels. After training and validating the data model using the training set and validation set, the target day data model is obtained, which is referred to here as the target day data model. Then, the independent variable X of the target day (the transformed all-weather WRF LST and multi-source remote sensing data of the target day output by the physical model, and the independent variable X contains all the independent variable pixels x of the corresponding image data) is input into the target day data model. Then, the target day data model outputs an all-weather reconstructed LST similar to MODIS LST based on the input independent variable X, achieving the effect of improving the accuracy of WRF LST using a data-driven method.

[0062] When put into use, since the MODIS LST of the day is used as a type of data in the training set, this data may be completely cloud-contaminated or have too few effective pixels to train the model. Therefore, it is necessary to divide it into two cases: when the MODIS LST has enough effective pixels (i.e., the cloud cover on the target day is small and there are enough pixels that are not cloud-contaminated), the GPR model trained on the data of that day is used directly to obtain the LST; otherwise, the GPR model with adjacent time is used to obtain the reconstructed LST.

[0063] This embodiment combines WRF and GPR to obtain the reconstruction model of this scheme, which is referred to here as the PW-DGRS model (Physicalmodel WRF and Data-driven GPR using multisource RS data coupled model, PW-DGRS). An example of the Beijing area is used for illustration:

[0064] The complex surface characteristics of the Beijing area lead to strong heterogeneity in the spatial distribution of surface temperature, making it an ideal region for testing the effectiveness of this solution. For the testing period, this embodiment selected February 21 to March 21, 2022. Since the MODIS imagery during this period includes both days with high cloud contamination and days with low cloud contamination, it is well-suited for testing the performance of the PW-DGRS model under different scenarios. The specific implementation steps are as follows:

[0065] A. Configuration of the WRF physical model:

[0066] A1. Setting parameters for the study area

[0067] In this embodiment, since the spatial resolution of the MODIS LST image is 1 km, the WRF model uses a three-layer unidirectional nested approach to enclose the study area, with spatial resolutions of 9 km, 3 km, and 1 km to simulate the target day's all-weather LST with a spatial resolution of 1 km. The dimensions are 124×124, 160×160, and 205×205, respectively. Of course, the WRF model does not necessarily need to be in the aforementioned form when put into use; it only needs to be able to simulate the all-weather LST with the same spatial resolution as the MODIS LST image. Since Beijing is located in the mid-latitude region, this embodiment uses the Lambert projection, with the projection center set at the center of Beijing, at latitude 40.21°N and longitude 116.39°E. The latitudinal intersections of the projection cone and the sphere are 60° and 30°, respectively.

[0068] A2. Setting up the physical scheme

[0069] The physical scheme settings of the WRF model have a significant impact on the accuracy of the results. Based on relevant land surface temperature research, this embodiment selects a physical scheme commonly used in land surface temperature models for China. Additionally, to obtain a more accurate WRF LST, this embodiment also uses the urban canopy (UCM) model as one of the physical schemes. Combining the WRF with the UCM model provides a more detailed description of surface radiation, energy transfer, and other processes. The physical scheme settings and their corresponding codes are shown in Table 1. When running WRF, the corresponding codes are entered into the program, and WRF will execute the corresponding scheme based on the codes. The WRF model has a time resolution of 30 minutes. Since the MODIS satellite passes over China at 13:30 local time, the results at 13:30 local time each day are selected as the WRF LST for this embodiment to ensure consistency with the MODIS LST time.

[0070] Table 1 Physical scheme settings for the WRF physical model

[0071]

[0072]

[0073] A3. Configuration Performance Verification

[0074] To test the performance of the WRF model in temperature simulation, the determined physical scheme was used to simulate the air temperature in the study area from February 21 to March 21, 2022 (corresponding to 52 to 80 days in 2022). The simulation results are compared with the air temperatures at six meteorological stations as follows: Figure 3 As shown in the figure, the comparison results show that the simulated temperature is quite close to the temperature at the stations. The mean absolute error (MAE) for the six stations is 1.63K, 1.79K, 1.67K, 1.88K, 1.89K, and 1.99K, respectively, all less than 2K. The root mean square error (RMSE) is 1.88K, 1.98K, 2.26K, 2.21K, 2.29K, and 2.36K, respectively, approximately 2K, indicating that the errors between the two datasets are very small and the dispersion of errors is also small. The correlation coefficients (ρ) are 0.96, 0.95, 0.93, 0.92, 0.91, and 0.91, respectively, all greater than 0.9, indicating that the two datasets are highly correlated, meaning that their trends are basically the same. Therefore, the WRF physical scheme set up in this paper has high performance in temperature simulation.

[0075] B. Converting a WRF-generated sphere to a WGS84 ellipsoid:

[0076] If the WRF simulation results in NC format, i.e., the target day's all-weather LST, are transformed into coordinates using ordinary NC format, such as through software like ArcGIS, QGIS, and GDAL's built-in tools, significant coordinate deviations will occur. Figure 4a As shown, this is because the results of WRF simulations are based on a perfect sphere, not the WGS84 ellipsoid commonly used in remote sensing. Ordinary NC format files provide the WGS84 coordinates of a certain point, and the coordinates of other points are calculated by equal intervals, thus calculating the boundaries of the image. However, for the results of WRF simulations, it is difficult to calculate the correct boundaries using the aforementioned method. To address this, this solution proposes an ellipsoid transformation method: First, a WRF reference sphere is created, and a spherical coordinate system reference sphere with a radius of 6370 km is created using the GDAL library. Second, the boundaries of the Lambert plane projection of the WRF (coordinates of the four corners of the image) are calculated based on the spherical coordinate system reference sphere. Next, these four coordinates are converted to geographic coordinates of the spherical coordinate system reference sphere, and then these coordinates are transformed onto the WGS84 coordinate reference sphere. Finally, the LST source image is resampled using the GDAL library based on the affine matrix, thus obtaining LST data under the WGS84 reference sphere. The affine matrix is ​​pre-compiled by technicians based on the boundaries of WGS84 and the image resolution of 0.01°. The transformed result is as follows: Figure 4b As shown, the pixel positions without this ellipsoidal transformation method differ significantly from their true positions (vector lines). Figure 4a The two small boxes in the middle are the most obvious.

[0077] C. Reconstruction and accuracy analysis of all-weather LST imagery

[0078] C1, Yun Shao's situation

[0079] After obtaining the WRF LST through the physical model WRF, the next step is to improve accuracy. When the MODIS LST is less cloud-polluted, a large number of pixels can be used to train the GPR model. In this scenario, the PW-DGRS model randomly divides the independent variable x and the dependent variable y into a training set and a validation set in a 7:3 ratio. The dependent variable y consists of the effective pixels of the MODIS LST, and the independent variable x includes the LST WRF, elevation DEM, latitude LAT, slope SLOPE, and normalized difference vegetation index NDVI, corresponding to the pixels of the dependent variable y. The PW-DGRS GPR model is obtained through training and validation.

[0080] f(X)=h(X) T β+Γ(X) (1)

[0081] LST PW-DGRS =f(LST) WRF(2)

[0082] Among them, LST PW-DGRS For the reconstruction of the all-weather LST, LST WRF DEM, LAT, SLOPE, and NDVI represent independent variables, f is the GPR model of PW-DGRS, and h(X) is the independent variable. T Let β be the basis function, Γ(X) be the parameter vector, and Γ(X) be the Gaussian process function.

[0083] Since neither the validation set pixels nor the cloud-contaminated pixels participated in the training of the GPR model, the accuracy of the validation set pixels can be considered as the accuracy of the cloud-contaminated pixels. In scenarios with few clouds, the accuracy of cloud-contaminated pixels can be evaluated based on the validation set pixels.

[0084] In this embodiment, the dates with less cloud contamination of the MODIS LST during the test period were selected to test the effect of PW-DDRS reconstruction of the LST, namely February 22, February 25, March 3, March 7, March 15 and March 20. Figure 5 The data shows the all-weather LST reconstructed from MODIS LST and the corresponding PW-DGRS model for the same date. MODIS LST exhibits significant gaps due to cloud pollution, while the corresponding PW-DGRS LST shows high spatial continuity, with clear-sky pixels displaying similar spatial patterns. Furthermore, PW-DGRS LST can recreate the spatial distribution details of LST well, such as the urban heat island effect, differences in LST for various land features, and differences in elevation distribution.

[0085] The error in reconstructing cloud-contaminated pixels using the PW-DGRS model is as follows: Figure 6 As shown, MAE ranges from 0.71K to 1.18K with a mean of 0.90K, RMSE ranges from 1.00K to 1.63K with a mean of 1.28K, and ρ ranges from 0.89 to 0.98 with a mean of 0.94.

[0086] The error of all pixels between PW-DGRS LST and MODIS LST is as follows: Figure 7 As shown, the MAE ranges from 0.39K to 0.66K, with a mean of 0.50K; the RMSE ranges from 0.63K to 1.01K, with a mean of 0.80K; and the ρ ranges from 0.96 to 0.99, with a mean of 0.98. Overall, under low-cloud conditions, the PW-DGRS model demonstrates high accuracy in reconstructing all-weather LSTs, with MAE less than 1K, RMSE less than 2K, and ρ greater than 0.9.

[0087] C2, Cloudy Scene

[0088] When the MODIS LST is heavily cloud-contaminated and lacks sufficient pixels to train the GPR model, the reconstructed LST will not be highly correlated with the MODIS LST. Here, a correlation coefficient of less than 0.8 between the reconstructed LST and the MODIS LST is considered non-highly correlated, i.e., a scenario with abundant clouds. When reconstructing the LST for all weather conditions, if the MODIS LST for the target day is highly correlated with the reconstructed LST for all weather conditions, it is considered a scenario with few clouds. The reconstructed all-weather LST obtained using the C1 method is the final result. If the MODIS LST for the target day is not highly correlated with the reconstructed LST for all weather conditions, it is considered a scenario with abundant clouds. In this case, the all-weather LST is obtained using a time-adjacent GPR model and the independent variable of the target date. The specific calculation formula is as follows.

[0089]

[0090]

[0091] Among them, LST PW-DDRS To reconstruct the LST, f is the GPR model adjacent to the target date, referred to here as the time-adjacent model, and N is the number of time-adjacent models, which is 2 in this embodiment, i.e., taking the time-adjacent models of the two days before and after the target date. R is the time-adjacent MODIS LST and the corresponding LST. PW-DDRS The correlation coefficient is T, where T is the number of days between the image dates. For example, if the image is on the first day after the target date, then T = 1; if it is on the second day after the target date, then T = 2, and so on. LST WRF ρ is the target date WRF LST, ρ is the correlation coefficient between the target date MODIS LST and LST WRF, and ω is the weight of temporally adjacent images.

[0092] In this scenario, since the target day MODIS LST did not participate in the training of the GPR model, the resulting LST is evaluated based on the target day MODIS LST pixels and the reconstructed all-weather LST pixels. All valid pixels of the MODIS LST can be used to evaluate the pixel accuracy of the entire image.

[0093] In this embodiment, the PW-DGRS model was tested under cloud-heavy conditions. The test dates selected were February 21, February 28, March 9, March 10, March 12, and March 13, when the LST was more contaminated by clouds. Figure 8The image shows the MODIS LST and the all-weather LST reconstructed by the PW-DRRS model for the corresponding dates. The MODIS LST is almost completely cloud-contaminated, making it impossible to represent the surface features of the study area. In contrast, the all-weather LST reconstructed by the PW-DRRS model has higher spatial continuity and can restore the spatial distribution details that the LST should have. The errors between the PW-DRRS LST and the MODIS LST are shown below. Figure 9 As shown, the MAE ranges from 0.97K to 1.72, all less than 2K; the RMSE ranges from 1.13K to 1.94K, all less than 2K; and the ρ ranges from 0.79 to 0.97, with almost all values ​​greater than 0.8. Overall, the PW-DGRS model can reconstruct high-precision all-weather LST even in cloudy scenarios.

[0094] D. Validation of the generalization performance of the PW-DGRS model

[0095] To prevent randomness in the test area and test period from causing PW-DGRS to reconstruct a high-precision LST, the PW-DGRS model from our Anna project was used again to reconstruct the LST of Tianjin in December 2020 (corresponding to 337 to 336 days in 2022), and the changes in error evaluation indicators were statistically analyzed, such as... Figure 10 As shown, the performance of the PW-DRRS model was verified by changing the study area and time period. The average values ​​of MAE and RMSE were 0.61K and 0.79K, respectively, both smaller than the error values ​​corresponding to Beijing, while ρ was as high as 0.9, indicating that the PW-DRRS model has high reliability.

[0096] This embodiment combines physical modeling with flexible data-driven modeling to form a data-driven, physical-assisted model framework. It proposes a novel all-weather LST reconstruction model (which, taking WRF combined with GPR as an example, can be called the PW-DGRS model), generating all-weather MODIS-like data with a spatial resolution of 1km, thus overcoming the limitation of satellite TIR imagery being unable to obtain complete LST data due to cloud contamination. Firstly, the physical model WRF is used based on historical reanalysis data from real observations, making the source data of LST closer to the true value. Secondly, the data obtained through physical laws is more scientific. Furthermore, the LST simulated by PW-DGRS has a spatial resolution of 1km and a time point of 13:30, perfectly corresponding to the spatial resolution and time point of MODIS LST imagery. While using data-driven methods to improve accuracy, multi-source remote sensing data that influence LST changes are also introduced as auxiliary driving forces to fully leverage the data mining capabilities of data-driven methods. In addition, the PW-DGRS model does not only consider pixel-by-pixel point-to-point fitting but also uses pixel spatial information as a driving force, thus giving the reconstructed LST obvious spatial continuity characteristics. Furthermore, the PW-DGRS model is based on the authenticity and scientific validity of LST source data. Then, multi-source remote sensing data and image spatial information are used to improve LST accuracy. Experiments have shown that high-precision all-weather LST images can be obtained in the end.

[0097] The specific embodiments described in this example are merely illustrative of the spirit of the invention. Those skilled in the art can make various modifications or additions to the described embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for reconstructing all-weather land surface temperature based on a physical model and data-driven approach, characterized in that, Includes the following steps: S1. Simulate the target day's all-weather LST with the same resolution as the MODIS LST using a physical model; S2. Perform ellipsoidal transformation on the target day all-weather LST to obtain the transformed target day all-weather WRF LST; The effective pixels of S3.MODIS LST are used as dependent variable y, the pixels in WRF LST that correspond to the spatiotemporal of dependent variable y and the pixels in multi-source remote sensing data that correspond to the spatiotemporal of dependent variable y are used as independent variables x. Each independent variable x and its corresponding dependent variable y are grouped together. Multiple sets of data are divided into training set and validation set. The training set is used to train the data model to obtain the trained target day data model. The validation set is used to validate the trained target day data model. S4. Input the transformed target day all-weather WRF LST and multi-source remote sensing data as independent variable X into the target day data model, and the target day data model outputs the target day all-weather reconstructed LST; In steps S3 and S4, the spatial neighboring pixels of the target pixel to be trained / predicted in the independent variables are used as the spatial characteristic information of the target pixel and input into the data model to be trained or the target daily data model that has been trained. The multi-source remote sensing data includes elevation DEM, latitude LAT, slope SLOPE, and normalized difference vegetation index (NDVI).

2. The all-weather surface temperature reconstruction method based on physical model and data-driven approach according to claim 1, characterized in that, If there are enough valid pixels in the MODIS LST for the target day, then perform step S4 to obtain the all-weather reconstructed LST for the target day. If there are not enough valid pixels in the MODIS LST for the target date, the final all-weather reconstructed LST for the target date is determined through step S5: Several trained data models for adjacent times of the target day are obtained as time-adjacent models. The independent variable X of the target day is input into several time-adjacent models respectively. The outputs of all time-adjacent models are summed according to the set weights to obtain the all-weather reconstructed LST of the target day.

3. The all-weather surface temperature reconstruction method based on physical models and data-driven methods according to claim 2, characterized in that, When the number of valid pixels in MODIS LST on the target day exceeds the set value of pixels, it is considered that there are enough valid pixels. Step S4 is executed, but S5 is not executed. Alternatively, if the correlation coefficient between the target date MODIS LST and the reconstructed LST obtained in step S4 is less than the set value, it is considered that there are not enough effective pixels, and step S5 is then executed.

4. The all-weather surface temperature reconstruction method based on physical model and data-driven approach according to claim 3, characterized in that, When there are not enough effective pixels in the MODIS LST of the target day, two data models trained on the day before and the day after the target day are obtained as time-adjacent models. The independent variable X of the target day is input into the two time-adjacent models respectively. The outputs of the two time-adjacent models are summed according to the set weights to obtain the reconstructed LST of the target day.

5. The all-weather surface temperature reconstruction method based on physical model and data-driven approach according to claim 4, characterized in that, When there are enough valid pixels in the MODIS LST for the target date, the calculation method for the data-driven part is as follows: （1） （2） Among them, LST PW-DGRS For the target date, LST will be rebuilt around the clock. WRF DEM, LAT, SLOPE, and NDVI are independent variables, f is the data model, and h(X) is the independent variable. T Let Γ(X) be a basis function, β be a parameter vector, and Γ(X) be a Gaussian process function. When there are not enough valid pixels for the target date's MODIS LST, the calculation method for the data-driven part is as follows: （3） （4） Among them, LST PW-DDRS For the all-weather reconstruction of the LST for the target day, f is the time-adjacent model for the target day, N is the number of time-adjacent models, and R is the time-adjacent MODIS LST and its corresponding LST. PW-DDRS The correlation coefficient, where T is the number of days between image dates, and LST. WRF ρ is the target date WRF LST, ρ is the correlation coefficient between the target date MODIS LST and LST WRF, and ω is the weight of the time-adjacent model.

6. The all-weather surface temperature reconstruction method based on physical model and data-driven approach according to claim 1, characterized in that, The physical model is a WRF model with six physical schemes: microphysics, cloud layer, shortwave radiation, longwave radiation, boundary layer, and land surface processes, and the urban canopy physical scheme is also included in the WRF model. The data model described is a machine regression algorithm, GPR model; In step S1, the physical model adopts a three-layer unidirectional nesting, and the spatial resolutions of the three layers of unidirectional nesting are 9KM, 3KM and 1KM respectively, while the spatial resolution of the MODIS LST image is 1KM. The physical model simulates the all-weather LST for the target day based on geographic static field data, historical reanalysis data, and physical schemes; The input to the physical model includes study area parameters that specify the nesting method, projection position, and projection method.

7. The all-weather surface temperature reconstruction method based on physical model and data-driven approach according to claim 1, characterized in that, In step S2, the target day all-weather LST is transformed into an ellipsoid using the following method to obtain the transformed target day all-weather WRF LST: S21. Create a coordinate system reference sphere of a perfect sphere; S22. Based on the coordinate system of a sphere, the four boundaries of the physical model's planar projection are calculated using a reference sphere. The four boundaries include the coordinate values ​​of the four corners of the target day's all-weather LST image. S23. Convert these four coordinates into geographic coordinate values ​​of a coordinate system reference sphere that is a perfect sphere, and then convert these four geographic coordinate values ​​to a WGS84 coordinate reference sphere; S24. Resample the all-weather LST image of the target day according to the affine matrix to obtain the LST data under the WGS84 reference sphere, that is, the converted all-weather WRF LST of the target day.

Images