A method and system for surface emissivity continuous wave spectrum inversion
By combining random forest and wavelet transform techniques, the surface emissivity is directly inverted from hyperspectral thermal infrared data, solving the problems of insufficient inversion accuracy and dependence on atmospheric correction in existing methods, and achieving high-precision continuous emissivity spectrum acquisition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA
- Filing Date
- 2025-12-31
- Publication Date
- 2026-07-10
AI Technical Summary
Existing emissivity inversion methods lack dynamic change information, have limited inversion accuracy, and rely on atmospheric correction in the separation of surface temperature and emissivity, making it difficult to achieve high-precision continuous spectrum emissivity inversion.
By combining the random forest method and wavelet transform technology, the surface emissivity is directly inverted from hyperspectral thermal infrared data. The optimal channel is selected through information entropy screening, the dimensionality is reduced by wavelet transform, and a machine learning model is constructed for inversion.
It improves the inversion accuracy and shape rationality of surface emissivity spectral lines, realizes high-precision continuous emissivity spectrum acquisition, and reduces dependence on atmospheric correction.
Smart Images
Figure CN121859566B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to thermal infrared remote sensing parameter inversion technology, and more particularly to a method and system for inverting the continuous spectrum of surface emissivity. Background Technology
[0002] Land surface emissivity (LSE) is a key variable reflecting the thermal radiation characteristics of land features, and it holds significant research potential and importance in fields such as soil property studies, vegetation cover change, bedrock mapping and resource exploration, and surface energy estimation. With the development of remote sensing technology, the emergence of hyperspectral thermal infrared remote sensing has made it possible to obtain continuous emissivity spectra. Compared to the wide-channel values obtained by multispectral thermal infrared remote sensing, continuous emissivity spectra can more fully reflect the attributes and characteristics of land features, and have far-reaching application value.
[0003] However, existing emissivity inversion methods are mainly designed for multispectral data, such as classification methods and NDVI thresholding methods. These methods typically obtain emissivity based on land cover type or other band data. Because they require additional information, the results usually lack dynamic variation information and have limited inversion accuracy.
[0004] As the number of thermal infrared channels in a sensor increases, co-separation of surface temperature and emissivity becomes possible. Existing solutions for separating surface temperature and emissivity are typically based on certain assumptions and require precise atmospheric correction, which is itself quite difficult. Even with atmospheric correction, N observations still correspond to N+1 unknowns (one surface temperature and N channel emissivity), making inversion still challenging. Furthermore, these methods focus more on the accuracy of surface temperature inversion, paying less attention to the surface emissivity inversion results.
[0005] In recent years, machine learning and deep learning methods have been widely applied to parameter inversion of hyperspectral thermal infrared data. Machine learning methods have gained widespread use because they can handle highly complex, nonlinear, ill-conditioned inversion problems, bypassing complex computational processes and eliminating the need for precise atmospheric corrections to retrieve land surface temperature from hyperspectral thermal infrared data. However, currently, machine learning methods are mainly applied to the inversion of land surface temperature or atmospheric parameters, with relatively limited application in the field of land surface emissivity inversion. Furthermore, existing research often focuses on emissivity values for a small number of channels, paying insufficient attention to the continuity of emissivity between hyperspectral channels and neglecting the importance of spectral line characteristics. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention provides a method and system for inverting the continuous spectrum of surface emissivity. For hyperspectral thermal infrared data, this scheme combines the random forest method and wavelet transform technology to directly invert the continuous spectrum of surface emissivity from satellite-observed radiance, thereby achieving high-precision acquisition of the continuous spectrum of emissivity and providing effective methods and system support for subsequent industry applications.
[0007] This invention provides a method for inverting the continuous spectrum of surface emissivity, which includes the following steps:
[0008] Step 1: Input the atmospheric profile database and the surface emissivity spectrum database, and combine the characteristics of the hyperspectral thermal infrared sensor (observation geometry and channel convolution function of the hyperspectral thermal infrared sensor, etc.) to perform hyperspectral thermal infrared data simulation based on the radiative transfer model to obtain the simulated dataset of hyperspectral thermal infrared.
[0009] Step 2: Based on the information entropy of each channel in the simulated dataset, iteratively select the channel with the highest current information entropy to be included in the preferred channel combination; when the set iterative convergence condition is met, record the current preferred channel combination to obtain the surface information sensitive feature channel;
[0010] Step 3: Perform binary discrete wavelet transform on the surface emissivity at each channel in the simulated dataset, and save the low-frequency wavelet coefficients obtained from the transform.
[0011] Step 4: Construct a machine learning model of surface emissivity based on the on-board brightness temperature of the feature channel obtained in Step 2 and the low-frequency wavelet coefficients obtained in Step 3 in the simulation dataset. The input of the model is the on-board brightness temperature of the feature channel, and the output is the low-frequency wavelet coefficients of the emissivity continuous spectrum.
[0012] The actual on-board brightness temperature observed by the sensor is fed into the trained machine learning model to obtain the low-frequency wavelet coefficients corresponding to the true surface emissivity spectrum; then, the obtained low-frequency wavelet coefficients are subjected to inverse wavelet transform to obtain the surface emissivity of all channels.
[0013] Furthermore, step 1 includes:
[0014] According to Planck's function Calculate the equivalent blackbody radiance at different surface temperatures and different channels. :
[0015]
[0016] in, This is the surface temperature simulated based on atmospheric temperature disturbances. and These are physical constants. Indicates the channel number of the hyperspectral thermal infrared sensor;
[0017] Based on the radiative transfer model of thermal infrared remote sensing, the on-board radiance of the hyperspectral sensor is calculated:
[0018]
[0019] in, This represents the simulated on-board radiance at channel λ. The simulated surface emissivity at channel λ is obtained from the surface emissivity spectral library; Let λ be the simulated atmospheric transmittance at channel λ. This represents the simulated atmospheric downward radiance at channel λ. The simulated atmospheric upward radiance at channel λ;
[0020] right By performing the inverse Planck transform, the simulated on-board brightness temperature at channel λ is obtained. .
[0021] Furthermore, in step 2, the process of iteratively selecting the channel with the highest current information entropy to be included in the preferred channel combination specifically includes:
[0022] In each iteration, based on the information entropy of the channels, the current set of candidate channels is selected. The channel with the highest information entropy is selected as the preferred channel for this iteration. Among them, the set of candidate channels The initial values are all channels of the hyperspectral thermal infrared sensor;
[0023] Update preferred channel combination With candidate channel set : Where t is the number of iterations. , This refers to the updated preferred channel combination and candidate channel set.
[0024] Furthermore, in step 2, the formula for calculating the channel's information entropy is:
[0025]
[0026] Where m is the channel to be selected. The background field covariance matrix, For observation error covariance, The weight function value corresponding to the candidate channel m can be obtained by calculating the partial derivative of the satellite brightness temperature with respect to the surface emissivity according to the radiative transfer equation.
[0027] Furthermore, in step 2, the iterative convergence condition is that the number of preferred channels reaches the set number of preferred channels, or the gain effect is greater than or equal to the preset precision gain threshold.
[0028] Furthermore, in step 2, the gain effect , and These are the current preferred channel combinations. With candidate channel set The corresponding accuracy of surface emissivity inversion.
[0029] Another aspect of the present invention provides a continuous spectrum inversion system for surface emissivity, including a data simulation module, a channel optimization module, a spectrum dimensionality reduction module, and a surface emissivity inversion module;
[0030] in,
[0031] The data simulation module is used to simulate hyperspectral thermal infrared data based on the input atmospheric profile database and surface emissivity spectrum library, combined with the characteristics of hyperspectral thermal infrared sensors (observation geometry and channel convolution functions of hyperspectral thermal infrared sensors, etc.), according to the radiative transfer model, and output a simulated dataset of hyperspectral thermal infrared data; to support the construction of channel feature optimization and surface emissivity inversion models.
[0032] The channel selection module is used to filter a subset of feature channels suitable for the surface emissivity inversion model. Based on the information entropy of each channel in the simulation dataset, the channel selection module iteratively selects the channel with the highest current information entropy and includes it in the preferred channel combination. When the set iterative convergence condition is met, the current preferred channel combination is used as the feature channel subset and sent to the surface emissivity inversion module.
[0033] The spectral dimensionality reduction module is used to obtain the surface emissivity spectral conversion model. Based on wavelet transform and emissivity spectral library, the spectral conversion model is constructed. The binary discrete wavelet transform is performed on the surface emissivity at each channel in the simulation dataset, and the low-frequency wavelet coefficients obtained by the transformation are sent to the surface emissivity inversion module.
[0034] The surface emissivity inversion module combines the sensor entrance pupil brightness temperature corresponding to the preferred channel subset with the low-frequency wavelet coefficients of the surface emissivity to construct a machine learning model of the surface emissivity (preferably a random forest inversion model). The input of this model is the on-board brightness temperature of the feature channels in the feature channel subset, and the output is the low-frequency wavelet coefficients of the continuous emissivity spectrum.
[0035] The surface emissivity inversion module is also used to input the actual on-board brightness temperature observed by the sensor into the trained machine learning model to obtain the low-frequency wavelet coefficients corresponding to the real surface emissivity spectrum, and then perform an inverse wavelet transform on them to obtain the surface emissivity of all channels.
[0036] The technical solution provided by this invention brings at least the following beneficial effects:
[0037] Based on the requirements of emissivity inversion modeling, this invention optimizes the selection of sensitive channels for surface emissivity information, then reduces the dimensionality and complexity of parameters through signal processing methods, and finally combines machine learning to achieve direct inversion from satellite observation data to the continuous surface emissivity spectrum. This invention can effectively improve the inversion accuracy and the rationality of the shape of surface emissivity spectral lines. Attached Figure Description
[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0039] Figure 1 A schematic diagram of the processing flow of a continuous spectrum inversion method for surface emissivity provided in an embodiment of the present invention;
[0040] Figure 2 This is a schematic diagram of the structure of a continuous spectrum inversion system for surface emissivity provided in an embodiment of the present invention. Detailed Implementation
[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.
[0042] This invention focuses on the problem of emissivity spectrum inversion under hyperspectral thermal infrared remote sensing, and proposes a method and system for continuous surface emissivity spectrum inversion. This invention combines a machine learning model (random forest) with signal processing techniques (wavelet transform) to achieve direct inversion of the surface emissivity spectrum. This invention not only fully utilizes the high precision and efficient computation of the random forest algorithm, but also effectively captures the shape characteristics of the emissivity spectrum in the thermal infrared band, thereby alleviating the neglect of emissivity spectrum shape by traditional methods, and avoiding the limitations of traditional methods such as dependence on atmospheric correction accuracy or difficulty in decoupling surface temperature and emissivity values. Therefore, this invention can directly invert high-precision, continuous, and complete surface emissivity spectra from observation data from spaceborne or airborne hyperspectral thermal infrared sensors, which has significant value for subsequent industry applications.
[0043] To directly obtain a smooth and continuous emissivity spectrum from the on-board radiance of a hyperspectral sensor and improve the accuracy of emissivity retrieval, wavelet transform and random forest methods are combined for emissivity spectrum retrieval. In one embodiment, such as... Figure 1 As shown in the figure, the surface emissivity continuous spectrum inversion method provided by this invention includes the following four steps: 1) production of supporting simulation data; 2) channel selection of hyperspectral data; 3) wavelet transform of the emissivity spectrum; 4) construction and application of the random forest model. Specifically, as follows:
[0044] S1: Production of Simulated Data
[0045] To support the construction of the physical inversion model of surface emissivity, supporting simulation data is required. This embodiment uses the HyTES airborne pushbroom imaging spectrometer developed by a jet propulsion laboratory as a reference, and conducts simulation of hyperspectral thermal infrared dataset (spectral range 7.5-12μm, with 256 channels) based on its channel response function and observation geometry.
[0046] First, by combining the TIGR (Thermodynamic Initial Guess Retrieval) atmospheric profile database and the ECOSTRESS surface emissivity spectrum database, typical atmospheric profiles and emissivity spectra for different surface types were selected. The surface temperature was set as a perturbation based on the lower-level air temperature of the atmospheric profile. Then, using the MODTRAN atmospheric radiative transfer model, the hyperspectral thermal infrared atmospheric transmittance and atmospheric uplink and downlink radiation information of HyTES under different land-atmosphere conditions were simulated. Finally, by integrating the channel response functions of the HyTES sensor with the emissivity spectra selected from the spectral library and the simulated atmospheric parameters, the surface and atmospheric data for the corresponding HyTES channels were obtained.
[0047] Based on the surface temperature and the selected emissivity spectral lines, firstly according to the Planck function B...λ , calculate the equivalent blackbody radiance at the surface temperature and at channel λ , and the specific calculation formula is:
[0048] (1)
[0049] In the formula, and are physical constants, = 1.191*108 W⋅μm-4⋅Sr-1⋅ m-2, [[ID=2k]]= 1.439*104 μm⋅K. λ represents the channel of the hyperspectral thermal infrared sensor, is the surface temperature simulated according to the air temperature disturbance at the bottom layer of the atmosphere.
[0050] Furthermore, combined with the radiation transfer model of thermal infrared remote sensing, calculate the on-orbit radiance of the hyperspectral sensor:
[0051] (2)
[0052] In the formula, is the simulated on-orbit radiance at channel λ, is the simulated surface emissivity at channel λ, which is selected from the ECOSTRESS database. is the simulated atmospheric transmittance at channel λ, represents the simulated downward atmospheric radiance at channel λ, represents the simulated upward atmospheric radiance at channel λ.
[0053] Perform Planck inverse transformation on the simulated on-orbit radiance of the channel to obtain the simulated on-orbit brightness temperature at channel λ :
[0054] (3)
[0055] In the formula, is the inverse function of the Planck function.
[0056] Furthermore, complete the simulation of the channel dataset for the hyperspectral thermal infrared sensing.
[0057] S2: Channel selection of hyperspectral data
[0058] Based on the simulation dataset obtained in step S1, and considering the core requirements of surface emissivity inversion modeling, a feature channel optimization process is conducted by comprehensively weighing key influencing factors such as effective channel information content, information redundancy, and instrument observation noise. The channel optimization employs a step-by-step iterative method based on information content, iteratively selecting the channel with the highest current information content for inclusion in the optimal combination. Simultaneously, background field information is dynamically updated based on the selected channels, laying the foundation for the next channel selection, thereby effectively reducing the correlation between selected channels.
[0059] Suppose we need to select n channels from the original hyperspectral channel set for model construction, and adopt an iterative approach to channel selection. In the selection process, we first calculate the information entropy index A of each candidate channel based on the simulation dataset. m To assess its potential for independent information contribution.
[0060] (4)
[0061] Where m is the channel to be selected, S a Let σ be the background field covariance matrix. m For the observation error covariance, k m This represents the weight function value corresponding to the candidate channel m.
[0062] Then, an iterative selection loop begins. In each iteration, the information content index of each candidate channel is calculated based on the current background field covariance matrix, and the channel with the highest information content index is selected for inclusion in the optimal combination.
[0063] (5)
[0064] The selected channels for this iteration are then updated, and the selected channel combination S is updated accordingly. t With candidate channel U t gather:
[0065] (6)
[0066] In the formula, t is the number of iterations.
[0067] Based on the newly selected channels, the effect of adding new channels on the inversion accuracy is evaluated in real time during each iteration.
[0068] (7)
[0069] In the formula, and Channel combination S t+1 and S t The corresponding surface emissivity inversion accuracy (which can be quantified through inversion error).
[0070] When Acc meets the accuracy requirements or reaches the set number of preferred channels, the iteration terminates, and the optimal channel scheme suitable for surface emissivity spectrum inversion is finally determined.
[0071] (8)
[0072] In the formula, n is the set number of preferred channels, and α is the preset precision gain threshold.
[0073] S3: Wavelet transform of emissivity spectrum
[0074] Wavelet transform is a commonly used data preprocessing algorithm in hyperspectral data inversion research. It decomposes a function into a linear combination of wavelet functions, thus representing a finite signal through the superposition of waves of different frequencies. When the wavelet representation possesses orthogonality, no redundant information is generated during the decomposition process, allowing the maximum amount of information to be expressed with the least amount of data.
[0075] The binary discrete wavelet transform is a typical form of discrete wavelet transform. Based on this method, wavelet transforms can be performed on the surface emissivity spectrum.
[0076] (9)
[0077] In the formula, DWT is the binary discrete wavelet transform function, and the low-frequency wavelet coefficients of the function can be obtained by fitting the surface emissivity spectrum in the simulation dataset. j represents the number of wavelet transform layers. This represents the low-frequency wavelet coefficients calculated from the surface emissivity spectrum in the simulation data after the j-th wavelet transform. These represent the high-frequency wavelet coefficients in the horizontal, vertical, and diagonal directions, respectively, after the j-th wavelet transform.
[0078] Using formula (9), the emissivity spectrum of the N channels of the hyperspectral thermal infrared sensor can be converted into j low-frequency wavelet coefficients and high-frequency wavelet coefficients. By discarding the high-frequency wavelet coefficients and retaining the low-frequency wavelet coefficients, the number of unknowns can be effectively reduced while preserving the surface emissivity spectrum characteristics, thus simplifying the structure of the machine learning model.
[0079] S4: Construction and Application of Random Forest Model
[0080] Following the wavelet transform in step S3, a random forest (RF) machine learning method is applied to invert the continuous emissivity spectrum. Based on the simulated dataset obtained in S1, the RF model can be fitted. Here, the simulated dataset is divided into 70% and 30% portions, with 70% used for training the machine learning model and adjusting hyperparameters, and the remaining 30% used for estimating the machine learning algorithm error. The RF model's input consists of the on-board brightness temperatures of n optimized channels, and its output is the low-frequency wavelet coefficients of the continuous emissivity spectrum.
[0081] (10)
[0082] In the formula, denoted as the on-board brightness temperature of n preferred channels in the simulated dataset; f represents the trained RF model.
[0083] By inputting the actual on-board brightness temperature observed by the sensor into the trained model, the low-frequency wavelet coefficients corresponding to the true surface emissivity spectrum can be obtained.
[0084] (11)
[0085] In the formula, These represent the low-frequency wavelet coefficients corresponding to the true surface emissivity spectrum. This represents the on-board brightness temperature of the n preferred channels actually observed by the sensor.
[0086] By performing an inverse wavelet transform on the low-frequency wavelet coefficients, the surface emissivity of all channels can be obtained.
[0087] (12)
[0088] In the formula, Let N be the surface emissivity spectrum of all N channels.
[0089] In one embodiment, the present invention also provides a surface emissivity continuous spectrum inversion system, such as... Figure 2 As shown, it includes four modules: data simulation module, channel optimization module, spectral dimensionality reduction module, and surface emissivity inversion module. Figure 2 The following is a system module connection diagram for the embodiment, and the functions of each module are described as follows:
[0090] The data simulation module is used to simulate hyperspectral thermal infrared datasets to support channel feature optimization and the construction of surface emissivity inversion models.
[0091] The channel selection module is used to select the most sensitive channels for surface emissivity information based on a step-by-step iterative method using information entropy, given a subset of feature channels suitable for the surface emissivity inversion model and a simulation dataset.
[0092] The spectral dimensionality reduction module is used to obtain the surface emissivity spectral conversion model. Based on wavelet transform technology and emissivity spectral library, the spectral conversion model is constructed to output the low-frequency wavelet coefficients of the surface emissivity.
[0093] The surface emissivity inversion module combines the sensor entrance pupil brightness temperature corresponding to the optimized channel subset and the low-frequency wavelet coefficients of the surface emissivity to construct a random forest inversion model for surface emissivity. Combined with the spectral conversion model, it realizes the estimation of the continuous spectrum of surface emissivity.
[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
[0095] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.
Claims
1. A method for inverting the continuous spectrum of surface emissivity, characterized in that, Includes the following steps: Step 1: Input the atmospheric profile database and the surface emissivity spectrum database, combine the characteristics of the hyperspectral thermal infrared sensor, and perform hyperspectral thermal infrared data simulation based on the radiative transfer model to obtain the simulated dataset of hyperspectral thermal infrared. Step 2: Based on the information entropy of each channel in the simulated dataset, iteratively select the channel with the highest current information entropy to be included in the preferred channel combination; when the set iterative convergence condition is met, record the current preferred channel combination to obtain the surface information sensitive feature channel; Step 3: Perform binary discrete wavelet transform on the surface emissivity at each channel in the simulated dataset, and save the low-frequency wavelet coefficients obtained from the transform. Step 4: Construct a machine learning model of surface emissivity based on the on-board brightness temperature of the feature channel obtained in Step 2 and the low-frequency wavelet coefficients obtained in Step 3 in the simulation dataset. The input of the model is the on-board brightness temperature of the feature channel, and the output is the low-frequency wavelet coefficients of the emissivity continuous spectrum. The actual on-board brightness temperature observed by the sensor is fed into the trained machine learning model to obtain the low-frequency wavelet coefficients corresponding to the true surface emissivity spectrum; then, the obtained low-frequency wavelet coefficients are subjected to inverse wavelet transform to obtain the surface emissivity of all channels.
2. The method as described in claim 1, characterized in that, Step 1 includes: According to Planck's function Calculate the equivalent blackbody radiance at different surface temperatures and different channels. : ； in, This is the surface temperature simulated based on atmospheric temperature disturbances. and These are physical constants. Indicates the channel number of the hyperspectral thermal infrared sensor; Based on the radiative transfer model of thermal infrared remote sensing, the on-board radiance of the hyperspectral sensor is calculated: ； in, This represents the simulated on-board radiance at channel λ. The simulated surface emissivity at channel λ is obtained from the surface emissivity spectral library; Let λ be the simulated atmospheric transmittance at channel λ. This represents the simulated atmospheric downward radiance at channel λ. The simulated atmospheric upward radiance at channel λ; right By performing the inverse Planck transform, the simulated on-board brightness temperature at channel λ is obtained. .
3. The method as described in claim 1, characterized in that, In step 2, the process of iteratively selecting the channel with the highest current information entropy to be included in the preferred channel combination specifically includes: In each iteration, based on the information entropy of the channels, the current set of candidate channels is selected. The channel with the highest information entropy is selected as the preferred channel for this iteration. Among them, the set of candidate channels The initial values are all channels of the hyperspectral thermal infrared sensor; Update preferred channel combination With candidate channel set : Where t is the number of iterations. , This refers to the updated preferred channel combination and candidate channel set.
4. The method as described in claim 3, characterized in that, In step 2, the formula for calculating the channel's information entropy is: ； Where m is the channel to be selected. The background field covariance matrix, For observation error covariance, This represents the weight function value corresponding to the channel m to be selected.
5. The method as described in claim 3, characterized in that, In step 2, the iterative convergence condition is that the number of preferred channels reaches the set number of preferred channels, or the gain effect is greater than or equal to the preset precision gain threshold.
6. The method as described in claim 5, characterized in that, In step 2, the gain effect , and These are the current preferred channel combinations. With candidate channel set The corresponding accuracy of surface emissivity inversion.
7. The method as described in claim 1, characterized in that, The machine learning model is based on random forest.
8. A continuous spectrum inversion system for surface emissivity, comprising a data simulation module, a channel optimization module, a spectrum dimensionality reduction module, and a surface emissivity inversion module; in, The data simulation module is used to simulate hyperspectral thermal infrared data based on the input atmospheric profile database and surface emissivity spectrum library, combined with the characteristics of hyperspectral thermal infrared sensors, and according to the radiative transfer model, and output the simulated dataset of hyperspectral thermal infrared data. The channel selection module is used to filter a subset of feature channels suitable for the surface emissivity inversion model. Based on the information entropy of each channel in the simulation dataset, the channel selection module iteratively selects the channel with the highest current information entropy and includes it in the preferred channel combination. When the set iterative convergence condition is met, the current preferred channel combination is used as the feature channel subset and sent to the surface emissivity inversion module. The spectral dimensionality reduction module is used to obtain the surface emissivity spectral conversion model. Based on wavelet transform and emissivity spectral library, the spectral conversion model is constructed. The binary discrete wavelet transform is performed on the surface emissivity at each channel in the simulation dataset, and the low-frequency wavelet coefficients obtained by the transformation are sent to the surface emissivity inversion module. The surface emissivity inversion module combines the on-board brightness temperature of the sensor corresponding to the feature channel subset with the low-frequency wavelet coefficients of the surface emissivity to construct a machine learning model of the surface emissivity. The input of this model is the on-board brightness temperature of the feature channel in the feature channel subset, and the output is the low-frequency wavelet coefficients of the continuous emissivity spectrum. The surface emissivity inversion module is also used to input the actual on-board brightness temperature observed by the sensor into the trained machine learning model to obtain the low-frequency wavelet coefficients corresponding to the real surface emissivity spectrum, and then perform an inverse wavelet transform on them to obtain the surface emissivity of all channels.