A full-process evaluation method for uncertainty of pixel-scale land surface temperature

Abstract

This invention belongs to the field of land surface temperature remote sensing monitoring technology, and specifically discloses a full-process evaluation method for pixel-scale land surface temperature uncertainty. The method includes: measuring data on atmospheric upward radiation, atmospheric downward radiation, and land surface emissivity to obtain point source land surface temperature; estimating the uncertainty and its components of the point source land surface temperature based on error propagation theory; constructing a random forest downscaling model and using this model to downscale the image to obtain the downscaled land surface temperature; quantifying the uncertainty of land surface temperature downscaling using Monte Carlo simulation; constructing a two-point scale transformation model based on the point source land surface temperature and the downscaled land surface temperature; predicting the downscaled land surface temperature using the two-point scale transformation model to obtain the predicted land surface temperature, and aggregating the predicted land surface temperatures to obtain the pixel-scale land surface temperature; constructing a DTC model based on MODIS data; and quantifying the uncertainty of the pixel-scale land surface temperature using Monte Carlo simulation, error propagation theory, and the DTC model.

Description

Technical Field

[0001] This invention relates to the field of land surface temperature remote sensing monitoring technology, specifically to a full-process evaluation method for pixel-scale land surface temperature uncertainty based on Monte Carlo simulation and error propagation theory. Background Technology

[0002] Surface temperature, as a crucial physical quantity between the Earth's surface and atmosphere, is a key parameter in the study of climate systems and surface processes, and is of great significance for understanding surface energy balance, water cycle, ecosystem health, and urban planning. With the rapid development of remote sensing technology, surface temperature retrieval based on satellite observations has become the main means of obtaining large-scale surface temperatures. In addition, surface temperature can also be measured through ground-based meteorological stations, and this method is an important means of verifying surface temperature satellite products. However, the uncertainty of ground-based surface temperature measurements is affected by various factors, including radiometer calibration, surface emissivity measurement, measurement methods, scale conversion methods, and time effects, requiring comprehensive quantification.

[0003] Currently, most studies directly compare ground station observations with remotely sensed land surface temperature products. However, this method does not consider uncertainties caused by various measurement errors during the acquisition of ground observations, nor by the differences in spatial scale between ground measurement data and remote sensing data. Furthermore, acquiring pixel-scale land surface temperature involves instrument calibration, measurement methods, and scale conversion; current research mostly quantifies uncertainties in only some of these steps. Therefore, a comprehensive quantification method for pixel-scale land surface temperature uncertainty is urgently needed to accurately assess the deviation between remote sensing products and actual land surface temperatures, providing a reliable basis for remote sensing product quality assessment. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention discloses a full-process evaluation method for pixel-scale land surface temperature uncertainty. This method analyzes and quantifies the uncertainty of each step in the pixel-scale land surface temperature acquisition process, ultimately achieving an accurate, scientific, and comprehensive evaluation of pixel-scale land surface temperature uncertainty. Specifically, it includes the following steps:

[0005] Data on atmospheric upward radiation, atmospheric downward radiation, and surface emissivity are measured to obtain point source surface temperature.

[0006] Estimate the uncertainty and its components of point source surface temperature based on error propagation theory;

[0007] A random forest downscaling model was constructed based on low-cloud Landsat-8 images, and the model was used to downscale the surface temperature of the images to obtain the downscaled surface temperature.

[0008] The uncertainty of surface temperature downscaling was quantified using Monte Carlo simulation.

[0009] A two-point scale transformation model is constructed based on point source surface temperature and downscaled surface temperature.

[0010] A two-point scaling transformation model is used to predict downscaled land surface temperature to obtain predicted land surface temperature. The predicted land surface temperatures are then aggregated to obtain pixel-scale land surface temperature.

[0011] DTC model built based on MODIS data;

[0012] The uncertainty of pixel-scale surface temperature is quantified using Monte Carlo simulation, error propagation theory, and the DTC model.

[0013] Obtaining point source surface temperature includes:

[0014] Instruments were used at ground stations to measure atmospheric upward and downward radiation.

[0015] The formula for calculating point source surface temperature using Stefan-Boltzmann's law is as follows:

[0016]

[0017] Among them, LST ground It is the point source surface temperature, R g It measures atmospheric upward radiation, R d It is the measured atmospheric downward radiation, and σ is the Stefan Boltzmann constant (5.67 × 10⁻⁶). -8 W / m 2 / K 4 ), ε b It is the surface emissivity;

[0018] Based on error propagation theory, the uncertainty of point source surface temperature and its components are estimated, including:

[0019] The formula for estimating the uncertainty of point source surface temperature is as follows:

[0020]

[0021] Among them, U ground_point It is the uncertainty of point source surface temperature, U g U d U emis These are the uncertainties of upward atmospheric radiation, downward atmospheric radiation, and emissivity, respectively.

[0022] The formula for estimating the components of the uncertainty in the surface temperature of a point source is as follows:

[0023]

[0024]

[0025] The sensitivity coefficients of each component are estimated by the first-order partial derivative, ΔR g ΔR d , Δε b These are atmospheric upward radiation error, atmospheric downward radiation error, and surface emissivity error, respectively.

[0026] Land surface temperature downscaling is performed using a random forest downscaling model, including:

[0027] Preprocessing of low-cloud Landsat-8 images, such as cloud removal and scaling factor application;

[0028] A random forest downscaling model is trained using preprocessed images. This model can be summarized as follows:

[0029] LST h =LST l +[f l (S h )-f l (S l )]

[0030] Among them, LST h It is high-resolution downscaled surface temperature, LST l It is a low-resolution surface temperature, S h It is the high-resolution scale factor, S l It is the low-resolution scale factor, f l (·) is a functional expression of the statistical relationship between low-resolution surface temperature and low-resolution scale factor;

[0031] The uncertainty of surface temperature downscaling is quantified using the Monte Carlo (MC) simulation method, including:

[0032] (1) For LST l S l S h Add the corresponding Gaussian errors respectively;

[0033] (2) Use the random forest downscaling model to downscale the land surface temperature and obtain the downscaled land surface temperature;

[0034] (3) Repeat (1) and (2) simulation operations 10,000 times, and calculate the standard deviation of the downscaled surface temperature obtained from the 10,000 simulations, which is the uncertainty of the downscaled surface temperature.

[0035] Based on point source surface temperature and downscaled surface temperature, a two-point scaling transformation model is constructed, as shown in the following formula:

[0036]

[0037] Among them, T s h It is the downscaled surface temperature corresponding to the high temperature point, T s l It is the downscaled surface temperature corresponding to the low temperature point. It is the point source surface temperature of the high-temperature point. It is the point source surface temperature at the low temperature point, and a and b are the parameters of the derived two-point method scale transformation model;

[0038] A two-point scaling transformation model is used to predict downscaled land surface temperatures. These predicted land surface temperatures are then aggregated to obtain pixel-scale land surface temperatures, including:

[0039] The predicted land surface temperature is obtained by downscaling, using the following formula:

[0040] T g,i =(1-a)T s,i -b

[0041] Among them, T s,i T represents the downscaled surface temperature of each 30m sub-pixel within a 1km × 1km area. g,i This represents the predicted land surface temperature corresponding to each 30m sub-pixel within a 1km×1km area;

[0042] The average aggregate yields the pixel-scale surface temperature using the following formula:

[0043]

[0044] Among them, T g,true It is the pixel-scale surface temperature, and N represents the number of 30m sub-pixels within a 1km×1km area;

[0045] Based on MODIS data, a DTC model is constructed, including:

[0046] Acquire high-quality daytime and nighttime surface temperature data for MOD11A1 and MYD11A1.

[0047] A DTC model is constructed, using a cosine function to predict surface temperature changes during the day and a hyperbolic function to describe the surface temperature decay process at night. The formulas are as follows:

[0048]

[0049] Where t represents time, T0 is the residual temperature before and after sunrise, and T a It is the temperature amplitude, ω is the half-period width of the cosine term, and t mt represents the moment when the temperature reaches its peak. sr Corresponding to sunrise time, t s The free decay start time is given by δT, which represents the temperature difference between T0 and T (t→∞), and k is the decay constant.

[0050] The uncertainty of pixel-scale surface temperature is quantified using Monte Carlo simulation, error propagation theory, and the DTC model, including:

[0051] For T s h T s l , Add Gaussian errors with mean values ​​equal to their corresponding uncertainties, perform 10,000 simulations, and calculate T. g,true Standard deviation;

[0052] Based on the DTC model, the difference in surface temperature between the satellite transit time and the ground measurement time is calculated as the time uncertainty;

[0053] According to error propagation theory, input T g,true The standard deviation and time uncertainty are used to calculate the uncertainty of surface temperature at the pixel scale.

[0054] This invention discloses a comprehensive evaluation method for pixel-scale land surface temperature uncertainty. Based on point-source land surface temperature and satellite image data, this method designs a random forest downscaling model and a two-point scale transformation model. Through error propagation theory and Monte Carlo simulation, it constructs a complete uncertainty transmission chain for land surface temperature from point sources to the pixel scale. Unlike existing technologies that only focus on the quantification of single or partial uncertainties, this invention incorporates the synergistic effects of multiple influencing factors (atmospheric parameter measurement errors, surface emissivity errors, downscaling model errors, scale transformation errors, time inconsistencies, etc.) during the acquisition of pixel-scale land surface temperature into a unified framework, achieving comprehensive quantification of pixel-scale land surface temperature uncertainty. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 A flowchart of a full-process evaluation method for pixel-scale surface temperature uncertainty provided by the present invention;

[0057] Figure 2This is a diagram showing the predicted results of point source surface temperature and its uncertainty based on measurement data from the BON site of the SURFRAD observation network according to the present invention.

[0058] Figure 3 This is a diagram showing the results of the random forest downscaling model and the land surface temperature downscaling uncertainty constructed based on Landsat-8 satellite data corresponding to the BON site of the SURFRAD observation network. Subplots a, b, and c respectively show the results at a distance of 1 km from the BON site. 2 The study area includes the original surface temperature at 30m resolution, the downscaled surface temperature at 30m resolution, and the uncertainty of surface temperature downscaling.

[0059] Figure 4 This is a diagram showing the results of the DTC model constructed based on MODIS data corresponding to the BON sites of the SURFRAD observation network in this invention. Detailed Implementation

[0060] To make the technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention:

[0061] like Figure 1 The full-process evaluation method for pixel-scale surface temperature uncertainty shown includes the following steps:

[0062] First, instruments are used at ground stations to measure atmospheric upward radiation, atmospheric downward radiation, and surface emissivity data. Then, the point source surface temperature is calculated using the Stefan-Boltzmann law, as shown in the following formula:

[0063]

[0064] Among them, LST ground It is the point source surface temperature, R g It measures atmospheric upward radiation (W / m²). 2 ), R d It is the measured atmospheric downdraft radiation (W / m²) 2 ), where σ is the Stefan Boltzmann constant (5.67 × 10⁻⁶). -8 W / m 2 / K 4 ), ε b It is the surface emissivity.

[0065] Based on error propagation theory, the uncertainty of point source surface temperature is estimated using the following formula:

[0066]

[0067] Among them, U g U d Uemis These are the uncertainties of upward atmospheric radiation, downward atmospheric radiation, and surface emissivity, respectively. The components of the uncertainty in point source surface temperature, i.e., U... g U d U emis It can be represented as:

[0068]

[0069] The sensitivity coefficients of each component are estimated by the first-order partial derivative, ΔR g ΔR d , Δε b These are the errors in atmospheric upward radiation, atmospheric downward radiation, and surface emissivity, respectively.

[0070] Point source surface temperature and its uncertainty quantification results are as follows Figure 2 As shown, atmospheric up-range radiation, atmospheric down-range radiation, and surface emissivity data from the SURFRAD network BON site were used, and ΔR was set according to the measurement instrument manual of the BON site. g ΔR d , Δε b ±5W / m 2 ±5W / m 2 ±0.005. The results indicate that the uncertainty of the point source surface temperature, i.e. Figure 2 The total uncertainty in the equation fluctuates between 0.8K and 1.0K. The ε of the BON site... b The surface emissivity from ASTER GED products is calculated using a linear relationship, as shown in the following formula:

[0071] ε b =0.197+0.025ε 10 +0.057ε 11 +0.237ε 12 +0.333ε 13 +0.146ε 14

[0072] Where, ε 10 To ε 14 This refers to the surface emissivity of channels 10 through 14 in the ASTER GED product.

[0073] Acquire low-cloud Landsat-8 images of the site, including B1-B7, B10, and QA_PIXEL, and decloudify B1-B7 and B10 based on QA_PIXEL. Train a random forest downscaling model using the declouded B1-B7 and B10 images. This model can be summarized as follows:

[0074] LST h=LST l +[f l (S h )-f l (S l )]

[0075] Among them, LST h It is high-resolution downscaled surface temperature, LST l It is a low-resolution surface temperature, S h It is the high-resolution scale factor, S l It is the low-resolution scale factor, f l (·) is a functional expression of the statistical relationship between low-resolution surface temperature and low-resolution scale factor.

[0076] The specific steps for downscaling land surface temperature using a random forest downscaling model include:

[0077] (1) Calculate 14 relevant spectral indices based on B1-B7 at 30m resolution, and the 30m high-resolution scale factor (S) h This refers to B1-B7 and 14 related spectral indices;

[0078] (2) S h Average aggregation yields a low-resolution scale factor (S) of 1 km. l );

[0079] (3) Average the 30m resolution B10 to generate a 1km low resolution LST. l ;

[0080] (4) Set the minimum sample size for the random forest model to 10, use 5-fold cross-validation to automatically determine the number of trees, and base it on LST. l and S l Train the random forest model to obtain f l (·);

[0081] (5) S l and S h Input the pre-trained random forest model into each input and obtain f. l (S l ) and f l (S h );

[0082] (6) Calculate the difference f l (S h )-f l (S l Add to LST l The surface temperature downscaled to 30m resolution, i.e., LST, was obtained. h .

[0083] The uncertainty of surface temperature downscaling is quantified using Monte Carlo simulations. The specific steps are as follows:

[0084] S1, Assuming the original surface temperature (LST) o ) and S h It is a normal distribution, for LST o and S h LST is obtained by adding the corresponding Gaussian error respectively. o,MC and S h,MC LST was obtained by average polymerization. l,MC and S l,MC .

[0085] S2, based on LST l,MC S h,MC and S l,MC The surface temperature (LST) was downscaled using a trained random forest downscaling model to obtain the LST. h,MC .

[0086] S3. Repeat steps S1 and S2 10,000 times to obtain 10,000 sets of LST. h,MC .

[0087] S4. Calculate 10,000 sets of LST. h,MC The standard deviation is the uncertainty of surface temperature downscaling using the random forest downscaling model.

[0088] This invention constructs and trains a random forest downscaling model based on Landsat-8 image data from the BON site of the SURFRAD observation network. While downscaling surface temperature, it simultaneously quantifies the uncertainty of the downscaling process. Experimental results are as follows: Figure 3 As shown, where Figure 3 a, 3b, and 3c respectively demonstrate the 1km radius of the BON station. 2 The study area includes the original surface temperature at 30m resolution, the downscaled surface temperature at 30m resolution, and the surface temperature downscaling uncertainty. Based on the S1 step for quantifying the surface temperature downscaling uncertainty, specifically, the LST of the Landsat-8 image of the BON site is analyzed. o and S h Gaussian errors with a mean of 1.0K and 5% were added respectively.

[0089] Acquire four MODIS images of the site from one day, including high-quality daytime and nighttime land surface temperature data (QC=0) for MOD11A1 and MYD11A1. Construct a DTC model, using a cosine function to predict land surface temperature changes during the day and a hyperbolic function to describe the decay process of land surface temperature at night, as shown in the following formula:

[0090]

[0091] Where t represents time, T0 is the residual temperature before and after sunrise, and T a It is the temperature amplitude, ω is the half-period width of the cosine term, and t m t represents the moment when the temperature reaches its peak. sr Corresponding to sunrise time, t s δT represents the free decay start time, δT represents the temperature difference between T0 and T(t→∞), and k is the decay constant.

[0092] like Figure 4 As shown, this invention constructs a DTC model based on high-quality surface temperature data from the MOD11A1 and MYD11A1 sites of the SURFRAD observation network BON station. The DTC model has five free parameters (T0, T...). a δT,t m and t s MODIS can only provide a maximum of four observations per day at a given location, so a fixed t s =t ss -1,t ss This is the time of sunset. Of the remaining free parameters, T0 is set to the minimum of the four MODIS surface temperatures, and T... a Set as the difference between the maximum and minimum values ​​of the four MODIS surface temperatures, t m The time was set to one hour after noon, and δT was set to 0.5. Furthermore, the Levenberg-Marquardt (LM) method was used to fit the DTC model.

[0093] Based on point source surface temperature and downscaled surface temperature, a two-point scaling transformation model is constructed, as shown in the following formula:

[0094]

[0095] Among them, T s h It is the downscaled surface temperature corresponding to the high temperature point, T s l It is the downscaled surface temperature corresponding to the low temperature point. It is the point source surface temperature of the high-temperature point. is the point source surface temperature at the low temperature point, and a and b are the parameters of the derived two-point scale transformation model.

[0096] The predicted land surface temperature is obtained by using a two-point scaling transformation model to predict downscaled land surface temperature, as shown in the following formula:

[0097] T g,i =(1-a)T s,i -b

[0098] Among them, Ts,i T represents the downscaled surface temperature of each 30m sub-pixel within a 1km × 1km area. g,i This represents the predicted land surface temperature corresponding to each 30m sub-pixel within a 1km×1km area.

[0099] The predicted surface temperature is averaged and aggregated to obtain the pixel-scale surface temperature, as shown in the following formula:

[0100]

[0101] Among them, T g,true It is the surface temperature at the pixel scale, and N represents the number of 30m sub-pixels within a 1km×1km area.

[0102] based on Figure 2 Point source surface temperature and its uncertainty and Figure 3 The downscaled surface temperature and its uncertainty were determined. The uncertainty of pixel-scale surface temperature was quantified using Monte Carlo simulation, error propagation theory, and the DTC model. The specific steps are as follows:

[0103] S1, against T s h T s l , Add Gaussian errors with mean equal to their corresponding uncertainties to obtain T. sh_MC T s l _MC、

[0104] S2. Construct a two-point scaling model and calculate a_MC and b_MC;

[0105] S3. Predict T using a_MC and b_MC s,i , get T g,i _MC;

[0106] S4, against T g,i _MC average aggregation yields LST true,MC ;

[0107] S5. Repeat steps S1-S4 above 10,000 times to calculate all LSTs. true,MC Standard deviation;

[0108] S6. Based on the constructed DTC model, calculate the difference in surface temperature between the satellite transit time and the ground measurement time as the time uncertainty;

[0109] S7. According to the error propagation theory, based on all LSTs true,MC The standard deviation and time uncertainty are used to calculate the uncertainty of ground surface temperature measurement at the pixel scale.

[0110] Based on all the above procedures, using ground measurement data from the SURFRAD observation network BON site and its corresponding Landsat-8 and MODIS satellite data, the uncertainty of the pixel-scale ground measurement surface temperature, predicted by Monte Carlo simulation and error propagation theory, is 1.423 K.

[0111] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for full-process evaluation of pixel-scale land surface temperature uncertainty, characterized in that include: Data on atmospheric upward radiation, atmospheric downward radiation, and surface emissivity are measured to obtain point source surface temperature. Estimate the uncertainty and its components of point source surface temperature based on error propagation theory; A random forest downscaling model was constructed, and the model was used to downscale the surface temperature of the image to obtain the downscaled surface temperature. The uncertainty of surface temperature downscaling was quantified using Monte Carlo simulation. A two-point scale transformation model is constructed based on point source surface temperature and downscaled surface temperature. A two-point scaling transformation model is used to predict downscaled land surface temperature to obtain predicted land surface temperature. The predicted land surface temperatures are then aggregated to obtain pixel-scale land surface temperature. DTC model built based on MODIS data; The uncertainty of pixel-scale surface temperature is quantified using Monte Carlo simulation, error propagation theory, and DTC model. The following formula is used to obtain the predicted land surface temperature by using a two-point scale transformation model to predict downscaled land surface temperature, and then aggregating the predicted land surface temperatures to obtain the pixel-scale land surface temperature: wherein, represents the down-scaled land surface temperature of each sub-pixel within a certain range, represents the predicted land surface temperature corresponding to each sub-pixel within a certain range; The predicted surface temperature is averaged and aggregated to obtain the pixel-scale surface temperature, as shown in the following formula: in, It is the surface temperature at the pixel scale, and N represents the number of sub-pixels within a certain range.

2. The method of claim 1, wherein: The following algorithm is used when constructing a two-point scale transformation model based on point source surface temperature and downscaled surface temperature: in, It is the downscaled surface temperature corresponding to the high temperature point. It is the downscaled surface temperature corresponding to the low temperature point. It is the point source surface temperature of the high-temperature point. is the point source surface temperature at the low temperature point, and a and b are the parameters of the derived two-point scale transformation model.

3. The method of claim 1, wherein: When estimating the uncertainty and its components of point source surface temperature based on error propagation theory: the point source surface temperature is calculated based on the Stefan Boltzmann law, the sensitivity coefficients of each error component are estimated through the first-order partial derivative, the uncertainty of each component is predicted, and then the uncertainty of point source surface temperature is synthesized through error propagation theory.

4. The method of claim 1, wherein: When constructing a random forest downscaling model and using it to downscale images for land surface temperature: satellite images are preprocessed, and the model is trained using n-fold cross-validation with a minimum sample size of m; Gaussian errors are added to the original land surface temperature values ​​and scaling factors, and k random forest downscaling is performed through Monte Carlo simulations. The uncertainty of the random forest downscaling model is quantified by the standard deviation of the downscaled land surface temperature.

5. The method of claim 2, wherein: A two-point scale conversion model is constructed based on point source surface temperature and downscaled surface temperature: the predicted surface temperature is obtained by solving for parameters a and b, and the predicted surface temperature is integrated to generate pixel-scale surface temperature using a mean aggregation strategy. Monte Carlo simulation is introduced in the prediction and aggregation process, and the conversion parameters are repeated k times after adding Gaussian error. The uncertainty contribution of the scale conversion process is quantified by standard deviation.

Images