A method for simulating annual changes in remotely sensed land surface temperature by combining global and local information

Abstract

The present invention relates to a method for simulating annual changes in remotely sensed surface temperature that combines global and local information (ATC_GL model), and simulates annual changes in remotely sensed surface temperature in a target area through the following formula LST(d)=LST_Global(d)+LST_Local(d). The present invention proposes an annual change model of remotely sensed surface temperature that combines global time domain and local time and space information, and couples a series of surface temperature related parameters through a random forest algorithm. The invention also improves the accuracy and applicability of the annual change model of surface temperature, and is expected to be applied to the production of large-scale and high-quality spatiotemporal continuous surface temperature products, and to assist in the analysis of the spatiotemporal pattern evolution characteristics of the urban thermal environment.

Description

Technical Field

[0001] The invention relates to a remote sensing surface temperature annual variation simulation method combining global and local information in the time domain, and belongs to the research field of remote sensing surface temperature annual variation simulation. Background Art

[0002] Land surface temperature is a key parameter for measuring energy exchange between the land surface and the atmosphere, and plays a vital role in various geoscience applications, such as urban thermal environment monitoring, soil moisture and evapotranspiration estimation, and forest fire detection. Thermal infrared remote sensing is the main way to obtain land surface temperature over a large area. However, due to factors such as satellite revisit period and cloud cover, remotely sensed land surface temperature has many omissions in both time and space dimensions, which greatly limits the application of land surface temperature data. Therefore, reconstructing the spatiotemporally continuous land surface temperature through time-series discrete satellite observations of land surface temperature is a hot issue in thermal infrared remote sensing research.

[0003] As a typical global algorithm in the time domain, the annual surface temperature change model can reconstruct the continuous surface temperature of the year through a small number of discrete observed surface temperatures. At the same time, as an important surface temperature reconstruction algorithm, this type of model is the link between the interannual and intraday surface temperature change models and the key to exploring the time scale evolution pattern of surface temperature. It has received close attention from the academic community, such as the production of seamless surface temperature products in time and space, long-term series analysis of surface temperature and urban heat island, and spatiotemporal downscaling of surface temperature. The academic community has carried out a series of explorations on the reconstruction of annual surface temperature. The existing annual surface temperature reconstruction models can be mainly divided into three categories. The first category is the statistical or Fourier series model, which mainly uses pure mathematical statistical methods to interpolate the surface temperature in time, such as Fourier analysis and statistical time series analysis. This type of model can simulate the annual change of surface temperature well, but the model is relatively complex and has low generalization ability. The second category is the physical or quasi-physical model. Starting from the surface energy balance equation, this type of model can reflect the change mechanism of surface thermal properties, and the parameters of the model are of great value in quantifying surface properties. For example, there are three-parameter annual surface temperature variation models and five-parameter annual surface temperature variation models. Compared with statistical or Fourier series models, the model form is simpler, with fewer free parameters, and it is easier to build a model with a small number of data points. In addition, the free parameters in the model can quantify the properties of landmarks. However, due to the simple form, the accuracy of this type of model needs to be improved. The third type is a semi-physical model that combines statistical and physical models. This type of model is based on the physical model and aims to reflect the daily fluctuation of surface temperature by introducing statistical methods or auxiliary data such as surface measured auxiliary data (such as air temperature, NDVI, etc.). For example, the improved annual surface temperature variation model (ATCE model) that considers air temperature and NDVI, the annual surface temperature variation model based on vegetation (PATC model) that considers the influence of vegetation phenology, and the hybrid annual surface temperature variation model (ATCH model) that considers multiple auxiliary parameters and the number of harmonics. Compared with the above two types of models, this type of model improves the simulation accuracy of daily fluctuations in surface temperature, but the above models are all global models in the time domain and do not consider the correlation of surface temperature at the local scale.

[0004] The above research shows that the accuracy of the annual surface temperature variation model can be improved by expanding harmonics and coupling auxiliary parameters related to surface temperature. However, the existing model still has the following problems: (1) It is difficult to reflect the spatial correlation of surface temperature. The annual surface temperature variation model is essentially a global algorithm in the time domain. It only considers the characteristics of surface temperature from a global perspective in the time domain, and does not consider the correlation of surface temperature on a local spatiotemporal scale. This will cause the model accuracy to decrease when there are fewer missing values ​​of surface temperature. (2) It is difficult to reveal the complex relationship between surface temperature and related auxiliary parameters. In addition to solar radiation, surface temperature is also affected by weather conditions, elevation and other factors. However, there is no simple linear regression relationship between them, which makes it difficult for related parameters to better describe the daily fluctuations of surface temperature. Therefore, it is necessary to study new surface temperature reconstruction methods that can consider both global and local spatiotemporal information in the time domain to further improve the accuracy and applicability of the model. Summary of the invention

[0005] The purpose of the present invention is to overcome the neglect of spatiotemporal local information in the existing annual surface temperature variation model and to propose a remote sensing surface temperature annual variation simulation method that combines global and local temporal information.

[0006] The surface temperature has a strong correlation in both the global time domain and the local time and space scales. The annual variation of surface temperature can be decomposed into two components: one is the annual periodic variation caused by changes in solar radiation (i.e., global variation), and the other is the daily fluctuation caused by factors such as weather conditions and surface properties (i.e., local variation). Among them, the annual periodic variation of surface temperature is mainly simulated by the global information of surface temperature, while the daily variation of surface temperature is mainly characterized by the local information of surface temperature. Therefore, the present invention couples the global and local information of surface temperature at the same time, fully considers the simulation of the two annual components of surface temperature, and proposes a remote sensing surface temperature annual variation simulation method (ATC_GL model) that combines global and local information. This method can reconstruct the surface temperature of the target area through formula (1),

[0007] LST(d)=LST_Global(d)+LST_Local(d) (1)

[0008] Where LST(d) represents the reconstructed surface temperature on the dth day relative to the vernal equinox, LST_Global(d) and LST_Local(d) represent the annual cycle change and daily fluctuation of the surface temperature on the dth day relative to the vernal equinox, respectively.

[0009] In the global time domain, the present invention simulates the annual cycle change of the surface temperature through the five-parameter surface temperature annual change model (ACP5 model). Due to the simplicity of the model and the full consideration of the periodicity of solar radiation, the five-parameter surface temperature annual change model can better simulate the annual cycle change of the surface temperature. The simulation of the annual cycle change of the surface temperature can be expressed as:

[0010]

[0011] Where MASK represents the annual average surface temperature, YAST1 and YAST2 represent the two amplitudes of the surface temperature in the year, d represents the number of days relative to the vernal equinox, θ represents the phase shift relative to the vernal equinox, and k1 and k2 are two constants, 2π / 365 and 4π / 365, respectively.

[0012] At the local scale of space and time, the statistical regression relationship between surface temperature and related parameters is established through the random forest algorithm, thereby describing the missing daily fluctuations of surface temperature. Since it is difficult to take into account both the accuracy and applicability of the model with a fixed space-time window, the present invention establishes a dynamic sliding space-time window with the missing pixel as the center, and reconstructs the daily fluctuations of surface temperature in the dynamic sliding space-time window.

[0013] The present invention takes into account the correlation of surface temperature in time and space, so let x and y represent the distance of the horizontal axis and vertical axis in the spatial dimension, respectively, and let d be the distance in the time dimension. Assuming that the entire study is A, it can be expressed as:

[0014]

[0015] where z x , z y and z d They respectively represent the range covered by the study area A in the spatial horizontal axis, spatial vertical axis and time dimension, and X, Y and D represent the maximum values ​​of A in the three dimensions.

[0016] In the study area, any pixel point can be expressed as P = P [x, y, d], where x and y represent the horizontal and vertical coordinates of P in the study area, respectively, and d represents the number of days offset from the vernal equinox. Let a missing pixel of surface temperature be P0, then P0 = P0 [x0, y0, d0]. We establish the initial space-time window with the missing pixel P0 as the center. The horizontal axis offset of the initial space-time window is α0, the vertical axis offset is β0, and the time dimension offset is γ0. Therefore, the initial space-time window B1 can be expressed as:

[0017] B1={p[x,y,z]∈A:x0-a0≤x≤x0+a0,y0-β0≤y≤y0+β0,d0-γ0≤d≤d0+γ0} (4)

[0018] In order to explore the size of the initial spatiotemporal window, the present invention uses a cross-validation method to explore the selection of the optimal initial spatiotemporal window (see Appendix Figure 3 As shown in the figure, by exploring, we set the offsets of our initial spatial horizontal axis (α0) and initial spatial vertical axis (β0) to 4, and the initial time offset (γ0) to 0.

[0019] In addition, the number and distribution of known surface temperatures in the initial space-time window will also affect the accuracy of the model. Therefore, the present invention sets a threshold value ω=10. When the known surface temperature in the initial space-time window is less than ω, the space-time window is expanded at the same time, and the expansion step in the spatial horizontal axis, spatial vertical axis and time dimension is i each time. Therefore, when the known surface temperature in the initial space-time window is less than the threshold value ω, the dynamic space-time window B can be expressed as:

[0020]

[0021] The increment of the expansion step is 1 each time until the number of known surface temperatures in the space-time window is greater than the threshold ω.

[0022] When the number of known surface temperatures within the space-time window B is greater than the threshold, the present invention establishes a statistical regression relationship between the surface temperature and related auxiliary parameters through a random forest algorithm, and then reconstructs the missing daily fluctuations in surface temperature. In view of the high similarity between the daily fluctuations of surface temperature and air temperature, and the change in surface temperature is also affected by the change in surface cover type and elevation difference, the present invention selects the daily fluctuation of air temperature, the normalized difference vegetation index (NDVI) and the elevation (DEM). It is worth noting that when the surface temperature data is missing, the daily fluctuation of the surface temperature of the pixel on that day is missing. Therefore, the reconstruction of the missing daily fluctuation of surface temperature can be expressed as:

[0023] LST_Local(d)=f RF (SAT_Local(d),NDVI,DEM) (6)

[0024] Among them, the daily fluctuation of air temperature (SAT_Local(d)) can be expressed as:

[0025]

[0026] Where SAT_Local(d) represents the daily fluctuation of air temperature on the dth day relative to the vernal equinox, SAT(d) represents the observed air temperature, SAT_Global(d) represents the annual variation of air temperature simulated by the five-parameter annual surface temperature variation model, and the parameters MASK1, YAST3, YAST4, and θ1 are fitted by the least squares method using more than four observed surface temperatures in each pixel per year.

[0027] It is worth noting that the ATC_GL model proposed in this invention is a surface temperature reconstruction framework, which can introduce more auxiliary parameters in different combinations according to different research needs, which can consider the accuracy and efficiency of the model in different scenarios.

[0028] The present invention improves both the accuracy and applicability of the model, can accurately reconstruct the surface temperature, and is expected to be suitable for the production of high-precision surface temperature products. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The present invention will be further described in conjunction with the accompanying drawings.

[0030] Figure 1 The present invention is a flow chart of the remote sensing surface temperature annual variation simulation method.

[0031] Figure 2 This is a diagram of the reconstruction results and original observation data of the ATC_GL model proposed in the present invention.

[0032] Figure 3 This is a diagram showing the influence of the selection of the initial spatial window on the model accuracy. w1–w8 represent the initial spatial horizontal and vertical axis offsets of 2, 3, 4, 5, 6, 7, 8, and 9 km, respectively. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0034] like Figure 1 Shown is a flow chart of a method for simulating annual changes in remotely sensed surface temperature that combines global and local information according to an embodiment of the present invention.

[0035] The method of the present invention comprises the following steps:

[0036] Step 1: Calculate the annual cycle change of surface temperature by the following formula

[0037] LST_Global(d)=ACP5(d)

[0038] =MASK+YAST1·sin(k1(d+θ))+YAST2·sin(k2(d+θ))

[0039] LST_Local(d) represents the annual cycle change of the surface temperature on the dth day relative to the vernal equinox, MASK represents the annual average surface temperature, YAST1 and YAST2 represent the two amplitudes of the surface temperature in the year, d represents the number of days relative to the vernal equinox, θ represents the phase shift relative to the vernal equinox, and k1 and k2 are two constants. In this step, k1 = 2π / 365, k2 = 4π / 365.

[0040] Step 2: Establish a dynamic sliding space-time window with the missing pixel as the center, and reconstruct the daily fluctuation of surface temperature in the dynamic sliding space-time window. The specific steps are as follows:

[0041] Step 2.1, let x and y represent the horizontal and vertical distances in the spatial dimension respectively, let d be the distance in the time dimension, assuming that the entire study is A, it can be expressed as:

[0042]

[0043] where z x , z y and z d Respectively represent the range covered by the study area A in the spatial horizontal axis, spatial vertical axis and time dimension, X, Y and D represent the maximum values ​​of the study area A in the three dimensions; in the study area A, any pixel point can be expressed as P = P[x, y, d], where x and y represent the horizontal and vertical coordinates of P in the study area, respectively, and d represents the number of days offset from the vernal equinox.

[0044] Step 2.2, let the missing pixel of a surface temperature be P0, then P0 = P0[x0, y0, d0]; take the missing pixel P0 as the center, establish the initial space-time window, the spatial horizontal axis offset of the initial space-time window is α0, the spatial vertical axis offset is β0, and the time dimension offset is γ0. The initial space-time window B1 is expressed as:

[0045] B1={p[x,y,z]∈A:x0-a0≤x≤x0+a0,y0-β0≤y≤y0+β0,d0-γ0≤d≤d0+γ0}

[0046] Step 2.3, set the initial spatial horizontal axis offset α0, the initial spatial vertical axis offset β0 and the initial time offset γ0. When the number of known surface temperatures in the initial space-time window is less than the threshold value ω, the space-time window is expanded with a step of i (in this embodiment, the expansion step i=1) until the number of known surface temperatures in the space-time window is not less than the threshold value ω. The space-time window B is given by the following formula:

[0047]

[0048] In order to explore the size of the initial spatiotemporal window, the present invention uses a cross-validation method to explore the selection of the optimal initial spatiotemporal window (see Appendix Figure 3 As shown in the figure, the offsets of our initial spatial horizontal axis (α0) and initial spatial vertical axis (β0) are both set to 4, and the initial time offset (γ0) is set to 0. In addition, the number and distribution of known surface temperatures in the initial spatiotemporal window will also affect the accuracy of the model. Therefore, the present invention sets a threshold value ω=10, and when the known surface temperature in the initial spatiotemporal window is less than ω, the spatiotemporal window is expanded at the same time.

[0049] Step 2.4: When the number of known surface temperatures in the spatial window B is greater than the threshold, the statistical regression relationship between the surface temperature and related auxiliary parameters is established through the random forest algorithm to reconstruct the missing daily fluctuations of the surface temperature. The related auxiliary parameters include the daily fluctuation of air temperature SAT_Local(d), the Normalized Difference Vegetation Index (NDVI) and the Digital Elevation Model (DEM). The related auxiliary parameters need to be resampled to the same spatial resolution as the surface temperature.

[0050] In view of the high similarity between the daily fluctuations of surface temperature and air temperature, and the change of surface temperature is also affected by the change of surface cover type and elevation difference, the present invention selects the daily fluctuation of air temperature, normalized vegetation index and elevation. It is worth noting that when the surface temperature data is missing, the daily fluctuation of surface temperature of the pixel on that day is missing. Therefore, the reconstruction of the missing daily fluctuation of surface temperature is expressed as:

[0051] LST_Local(d)=f RF (SAT_Local(d),NDVI,DEM)

[0052] The daily fluctuation of air temperature SAT_Local(d) is calculated by the following formula:

[0053] SAT_Local(d)=SAT(d)-SAT_Global(d)

[0054] =SAT(d)-{MASK1+YAST3·sin(k1(d+θ1))+YAST4·sin(k2(d+θ1))}

[0055] Where SAT_Local(d) represents the daily fluctuation of air temperature on the dth day relative to the vernal equinox, SAT(d) represents the observed air temperature on the dth day relative to the vernal equinox, SAT_Global(d) represents the annual variation of air temperature on the dth day relative to the vernal equinox simulated by the five-parameter annual surface temperature variation model, and the parameters MASK1, YAST3, YAST4, and θ1 are obtained by fitting the least squares method of more than four observed surface temperatures in each pixel in the year.

[0056] Step 4: The present invention simulates the annual cycle change of the surface temperature through the five-parameter surface temperature annual change model (ACP5 model). Due to the simplicity of the model and the full consideration of the periodicity of solar radiation, the five-parameter surface temperature annual change model can better simulate the annual cycle change of the surface temperature.

[0057] The surface temperature of the target area is reconstructed by the following formula.

[0058] LST(d)=LST_Global(d)+LST_Local(d)

[0059] Where LST(d) represents the reconstructed surface temperature on the dth day relative to the vernal equinox, LST_Global(d) and LST_Local(d) represent the annual cycle change and daily fluctuation of the surface temperature on the dth day relative to the vernal equinox, respectively.

[0060] After the surface temperature of the target area is reconstructed by the present invention, a random pixel point is selected to display the reconstruction result of the present invention and the original surface temperature map (such as Figure 2 shown).

[0061] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present invention.

Claims

1. A method for simulating annual changes in remotely sensed surface temperature by combining global and local information, comprising the following steps: Step 1: Calculate the annual cycle change of surface temperature by the following formula LST_Global(d)=ACP5(d) =MASK+YAST1·sin(k1(d+θ))+YAST2·sin(k2(d+θ)) LST_Global(d) represents the annual periodic variation of the surface temperature on the dth day relative to the vernal equinox, MASK represents the annual average surface temperature, YAST1 and YAST2 represent the two amplitudes of the surface temperature in the year, d represents the number of days relative to the vernal equinox, θ represents the phase shift relative to the vernal equinox, and k1 and k2 are two constants; Step 2: Establish a dynamic sliding space-time window with the missing pixel as the center, and reconstruct the daily fluctuation of surface temperature in the dynamic sliding space-time window. The specific steps are as follows: Step 2.1, let x and y represent the horizontal and vertical distances in the spatial dimension, respectively, let d be the distance in the time dimension, and the study area A is expressed as: where z x , z y and z d Respectively represent the range covered by the study area A in the spatial horizontal axis, spatial vertical axis and time dimension, X, Y and D represent the maximum values ​​of the study area A in the three dimensions; in the study area A, any pixel point can be expressed as P = P[x, y, d], where x and y represent the horizontal and vertical coordinates of P in the study area, respectively, and d represents the number of days offset from the vernal equinox. Step 2.2, let the missing pixel of a surface temperature be P0, then P0 = P0[x0, y0, d0]; take the missing pixel P0 as the center, establish the initial space-time window, the spatial horizontal axis offset of the initial space-time window is α0, the spatial vertical axis offset is β0, and the time dimension offset is γ0. The initial space-time window B1 is expressed as: B1={p[x,y,z]∈A:x0-a0≤x≤x0+a0,y0-β0≤y≤y0+β0,d0-γ0≤d≤d0+γ0} Step 2.3, set the initial spatial horizontal axis offset α0, the initial spatial vertical axis offset β0 and the initial time offset γ0. When the number of known surface temperatures in the initial space-time window is less than the threshold ω, expand the space-time window with a step of i until the number of known surface temperatures in the space-time window is not less than the threshold ω. The space-time window B is given by the following formula: Step 2.4: When the number of known surface temperatures in the space-time window B is greater than the threshold, the statistical regression relationship between the surface temperature and related auxiliary parameters is established through the random forest algorithm to reconstruct the missing daily fluctuation of the surface temperature. The related auxiliary parameters include the daily fluctuation of air temperature SAT_Local(d), the normalized difference vegetation index NDVI and the digital elevation model DEM. The reconstructed daily fluctuation of the missing surface temperature is expressed as: LST_Local(d)=f RF (SAT_Local(d),NDVI,DEM) The daily fluctuation of air temperature SAT_Local(d) is calculated by the following formula: SAT_Local(d)=SAT(d)-SAT_Global(d) =SAT(d)-{MASK1+YAST3·sin(k1(d+θ1))+YAST4·sin(k2(d+θ1))} Where SAT_Local(d) represents the daily fluctuation of the air temperature on the dth day relative to the vernal equinox, SAT(d) represents the observed air temperature on the dth day relative to the vernal equinox, SAT_Global(d) represents the annual variation of the air temperature on the dth day relative to the vernal equinox simulated by the five-parameter surface temperature annual variation model, parameters MASK1, YAST3, YAST4, θ1 are obtained by fitting more than four observed surface temperatures in each pixel in the year, k1 and k2 are two constants; Step 4: Reconstruct the surface temperature of the target area using the following formula: LST(d)=LST_Global(d)+LST_Local(d) Where LST(d) represents the reconstructed surface temperature on the dth day relative to the vernal equinox, LST_Global(d) and LST_Local(d) represent the annual cycle change and daily fluctuation of the surface temperature on the dth day relative to the vernal equinox, respectively.

2. The method for simulating annual changes in remotely sensed surface temperature by combining global and local information according to claim 1, characterized in that: In step 1, k1=2π / 365, k2=4π / 365.

3. The method for simulating annual changes in remotely sensed surface temperature by combining global and local information according to claim 1, characterized in that: In step 2.3, the optimal initial spatiotemporal window size is set using the cross-validation method.

4. The method for simulating annual changes in remotely sensed surface temperature by combining global and local information according to claim 3, characterized in that: In step 2.3, the initial spatial horizontal axis offset α0 and the initial spatial vertical axis offset β0 are both set to 4, and the initial time offset γ0 is set to 0.

5. The method for simulating annual changes in remotely sensed surface temperature by combining global and local information according to claim 1, characterized in that: In step 2.3, the enlargement step i=1.

6. The method for simulating annual changes in remotely sensed surface temperature by combining global and local information according to claim 1, characterized in that: In step 2.3, the threshold ω=10.

7. The method for simulating annual changes in remotely sensed surface temperature by combining global and local information according to claim 1, characterized in that: In step 2.4, the relevant auxiliary parameters need to be resampled to the same spatial resolution as the surface temperature.

8. The method for simulating annual changes in remotely sensed surface temperature by combining global and local information according to claim 1, characterized in that: In step 2.4, the parameters MASK1, YAST3, YAST4, and θ1 are obtained by least squares fitting of more than four observed surface temperatures in each pixel year.

Images