Ground surface temperature learning dense reconstruction method based on integration and fusion of generalized space-time spectrum

Abstract

The application discloses a ground surface temperature learning dense reconstruction method based on generalized space-time spectrum integration fusion, and comprises the following steps: using a generalized space-time spectrum integration fusion network to perform deep learning dense reconstruction on multi-source satellite-mode auxiliary parameters of ground surface temperature; network training comprises the following steps: constructing a multi-source satellite-mode training sample library; constructing a generalized space-time spectrum integration fusion network, which comprises the following steps: initializing modeling of an atmospheric-ground surface radiation transmission process and a ground surface energy balance process, extracting nonlinear residual features of multi-source satellite-mode auxiliary parameters through a multi-dimensional feature collaborative perception module to compensate for initialization deviation of the physical process; performing feature aggregation on optimized ground surface temperature and simulated ground surface temperature through an adaptive feature fusion module to obtain a final prediction result of the ground surface temperature; and training the generalized space-time spectrum integration fusion network based on the multi-source satellite-mode training sample library until convergence. Therefore, the accuracy and reliability of the ground surface temperature reconstruction result are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing technology for remote sensing images, specifically to a method for learning-intensive reconstruction of land surface temperature based on the integrated fusion of generalized spatiotemporal spectrum. Background Technology

[0002] Surface temperature is a key physical parameter characterizing the interaction between the Earth's surface and the atmosphere. Designated as a crucial climate variable in the terrestrial biosphere by global climate observation systems, it is widely used in research fields such as hydrology, ecology, and urban science. Thermal infrared remote sensing inversion is the primary method for obtaining large-scale, long-term, and high-precision surface temperatures. However, limited by factors such as cloud cover and sensor hardware imaging performance constraints, thermal infrared remote sensing data generally suffers from severe spatial gaps. Furthermore, it faces technical bottlenecks where spatiotemporal resolution mutually restricts its application, making it difficult to directly obtain dense, all-weather surface temperature sequences. This limits the ability to characterize the dynamic evolution of the surface thermal field and fails to meet the demands of highly dynamic and refined applications.

[0003] To address the aforementioned issues, dense reconstruction techniques for land surface temperature mainly include methods based on harmonic regression and information fusion. Harmonic regression methods typically use trigonometric functions to construct periodic function models that describe the continuous temporal variation of land surface temperature, such as time series analysis based on Fourier series and diurnal temperature cycle models. These algorithms are relatively simple to compute and computationally efficient, but their modeling effectiveness is highly dependent on continuous spatiotemporal observation conditions, making robust reconstruction difficult in scenarios with long-term data gaps, such as persistent cloud cover. On the other hand, information fusion leverages the complementary advantages of multi-source satellite observations and model simulations across time, space, and spectral dimensions to construct spatiotemporal fusion models based on multi-temporal high- and low-resolution data, or generalized spatial-spectral fusion models based on land surface temperature and other auxiliary parameters, to generate highly dynamic, all-weather land surface temperature data products. However, existing surface temperature intensive reconstruction technologies typically focus only on the fusion of information in two dimensions, and have not yet achieved integrated and collaborative perception of spatiotemporal and spectral multidimensional information. Furthermore, they are mostly based on linear or weakly nonlinear correlation assumptions, making it difficult to fully characterize the complex heterogeneous features of the surface and the rapid changes in surface temperature. Summary of the Invention

[0004] To overcome the shortcomings of existing surface temperature dense reconstruction methods in terms of multi-dimensional information collaborative perception and nonlinear expression capabilities, this invention provides a surface temperature learning-based dense reconstruction method based on the integration of generalized spatiotemporal spectrum. Based on daily-scale multi-source satellite observations and hourly-scale common land surface process model simulations, it can fully integrate spatiotemporal and spectral multi-dimensional complementary information and has strong nonlinear modeling capabilities, thereby improving the accuracy and reliability of surface temperature reconstruction results.

[0005] According to one aspect of the present invention, a method for learning-intensive reconstruction of land surface temperature based on generalized spatiotemporal-spectral integrated fusion is provided, comprising: acquiring multi-source satellite-model auxiliary parameters of a target time phase; fusing the multi-source satellite-model auxiliary parameters of the target time phase using a trained generalized spatiotemporal-spectral integrated fusion network to output an hourly high-dynamic all-weather land surface temperature sequence; wherein, the training of the generalized spatiotemporal-spectral integrated fusion network includes: constructing a spatiotemporally matched multi-source satellite-model training sample library; constructing a generalized spatiotemporal-spectral integrated fusion network for intensive reconstruction of land surface temperature; the fusion network uses the spatiotemporally matched multi-source satellite-model auxiliary parameters as input to construct a structure pair The proposed dual-branch deep learning fusion architecture initializes the atmospheric-surface radiative transfer process and the surface energy balance process, respectively. A multi-dimensional feature collaborative sensing module performs temporal, spatial, and generalized spectral phase feature sensing on the multi-source satellite-model auxiliary parameters, outputting an optimized surface temperature. An adaptive feature fusion module aggregates the optimized surface temperature output from the two branches with the initial thermal state field provided by the land surface process model, and a feature reconstruction module outputs the final predicted surface temperature. Based on the multi-source satellite-model training sample library, a joint loss function that considers physical constraints is used to train the generalized temporal-spatial-spectral integrated fusion network until convergence.

[0006] Furthermore, the construction of the generalized spatiotemporal-spectral integrated fusion network includes: based on the atmospheric-surface radiative transfer theory, initial modeling of the radiative transfer process is performed using surface emissivity, model-simulated upward longwave radiation, and model-simulated downward longwave radiation to obtain an initial estimate of the surface temperature; a generalized spatiotemporal-spectral multidimensional feature collaborative sensing module for the radiative transfer process is constructed, and spatiotemporal and generalized spatiotemporal spectral compensation features for the initial modeling bias of the radiative transfer theory are jointly extracted using a residual dense network, and then combined with the initial estimate of the surface temperature to obtain an optimized surface temperature; based on the surface energy balance theory, vegetation is used... Initial modeling of the energy balance process is performed using coverage, simulated air temperature, simulated net radiation, and simulated aerodynamic impedance to obtain an initial estimate of the land surface temperature. A generalized spatiotemporal-spectral multidimensional feature collaborative sensing module for the energy balance process is constructed. A residual dense network is used to jointly extract spatiotemporal and generalized spatiotemporal compensation features to address the limitations of energy balance theory. These features are then combined with the initial land surface temperature estimate to obtain an optimized land surface temperature. The adaptive feature fusion module aggregates the simulated land surface temperature and the two optimized land surface temperatures to obtain fused features. These fused features are then reconstructed into the final predicted land surface temperature.

[0007] Furthermore, a residual dense network is used to jointly extract spatiotemporal and generalized spatial spectrum compensation features for the initial modeling bias in radiative transfer theory. These features are then combined with the initial surface temperature estimate to obtain an optimized surface temperature, as shown in the following formula:

[0008] ,

[0009] ,

[0010] in, To optimize the surface temperature corresponding to the radiative transfer process, This represents the initial estimate of the surface temperature during the radiative transfer process. This is a collaborative sensing module for multi-dimensional features of the radiative transfer process. To compensate for the spatiotemporal features extracted for the initial modeling bias in radiative transfer theory, To compensate for the generalized spatial spectrum features extracted for the initial modeling bias in radiative transfer theory, and for and Corresponding feature extractors, To simulate surface temperature in the model, and For reference time phase satellite-model surface temperature data pairs, For digital elevation model data, For vegetation coverage, For surface emissivity, To simulate air temperature in the model, To simulate aerodynamic impedance for the model, To simulate net radiation in the model, Due to water vapor saturation difference, This represents the slope of the saturated water vapor pressure curve.

[0011] Furthermore, a residual dense network is used to jointly extract spatiotemporal and generalized spatial spectrum compensation features to address the limitations of energy balance theory. These features are then combined with the initial estimate of surface temperature to obtain an optimized surface temperature. The corresponding formula is as follows:

[0012] ,

[0013] ,

[0014] in, To optimize the surface temperature corresponding to the energy balance process, To simulate air temperature in the model, air density, The specific heat of air at constant pressure. The dry-bulb and wet-bulb constants are... This is a collaborative sensing module for multi-dimensional features of the energy balance process; To address the limitations of energy balance theory in obtaining spatiotemporal compensation characteristics, To address the limitations of energy balance theory, a generalized spatial spectrum compensation feature was obtained. and They are respectively and Corresponding feature extractors, To simulate upward longwave radiation in the model, The model simulates downlink longwave radiation.

[0015] Furthermore, based on the multi-source satellite-model training sample library, the generalized spatiotemporal-spectral integrated fusion network is trained to convergence using a joint loss function that takes into account physical constraints. This includes: constructing a joint loss function that takes into account physical constraints, including: constructing a numerical consistency loss using the reconstructed final predicted surface temperature and the real satellite surface temperature label under ideal clear sky conditions; constructing physical consistency losses for radiative transfer and energy balance processes respectively using the optimized surface temperature output by the two generalized spatiotemporal-spectral multidimensional feature collaborative sensing modules and the real satellite surface temperature label under ideal clear sky conditions; constructing a joint loss function based on the numerical consistency loss, the physical consistency loss of radiative transfer and energy balance processes; and training the generalized spatiotemporal-spectral integrated fusion network using the joint loss function that takes into account physical constraints based on the multi-source satellite-model training sample library until the generalized spatiotemporal-spectral integrated fusion network converges, thus obtaining the trained generalized spatiotemporal-spectral integrated fusion network.

[0016] Furthermore, the formula corresponding to the joint loss function is:

[0017] ,

[0018] ,

[0019] ,

[0020] ,

[0021] in, Let be the total loss function of the generalized spatiotemporal-spectral integrated fusion network. For numerical consistency loss, and These represent the physical consistency loss in radiative transfer and energy balance processes, respectively. These are the weight coefficients of each sub-loss function. Let i be the number of training samples, and i be the index of the training sample. This represents a real satellite surface temperature label under ideal clear sky conditions. For L2 norm loss, This indicates the final predicted result for surface temperature. The surface temperature is the result of initial modeling of the atmospheric-surface radiative transfer process and optimization by the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module. The surface temperature is initialized by modeling the surface energy balance process and optimized by the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module.

[0022] Furthermore, the construction of the spatiotemporally matched multi-source satellite-model training sample library includes: acquiring satellite surface temperature, satellite auxiliary parameters, meteorological forcing data, and surface state data; preprocessing the satellite auxiliary parameters; sampling the satellite surface temperature in completely cloudless areas within the window range in blocks to extract real satellite surface temperature labels under ideal clear sky conditions; driving a common land surface process model to simulate based on meteorological forcing data and surface state data to obtain model simulation parameters; performing diurnal-scale synthesis processing on satellite surface temperature and model simulated surface temperature to construct a reference temporal satellite-model surface temperature data pair; and constructing a spatiotemporally matched multi-source satellite-model training sample library using real satellite surface temperature labels and multi-source satellite-model auxiliary parameters; the multi-source satellite-model auxiliary parameters include satellite auxiliary parameters, model simulation parameters, and reference temporal satellite-model surface temperature data pairs.

[0023] According to one aspect of this invention, a land surface temperature learning-based dense reconstruction system based on generalized spatiotemporal-spectral integrated fusion is provided, implemented using the aforementioned method. The system includes: a target multi-source data acquisition module for acquiring multi-source satellite-model auxiliary parameters for the target time phase; and a land surface temperature deep learning-based dense reconstruction module for fusing the multi-source satellite-model auxiliary parameters of the target time phase using a trained generalized spatiotemporal-spectral integrated fusion network to output an hourly high-dynamic all-weather land surface temperature sequence. The training of the generalized spatiotemporal-spectral integrated fusion network includes: constructing a spatiotemporally matched multi-source satellite-model training sample library; and constructing a generalized spatiotemporal-spectral integrated fusion network for dense land surface temperature reconstruction. The fusion network uses spatiotemporally matched multi-source satellite-model auxiliary parameters as input to construct a structurally symmetrical dual-branch deep learning fusion architecture. The two branches initialize and model the atmospheric-surface radiative transfer process and the surface energy balance process, respectively. Through a multi-dimensional feature collaborative perception module, the multi-source satellite-model auxiliary parameters are collaboratively perceived in three dimensions: time, space, and generalized spectral phase, outputting an optimized surface temperature. An adaptive feature fusion module aggregates the optimized surface temperature output from the two branches with the initial thermal state field provided by the land surface process model, and finally uses a feature reconstruction module to output the final predicted surface temperature. Based on the multi-source satellite-model training sample library, the generalized spatiotemporal-spectral integrated fusion network is trained to convergence using a joint loss function that takes into account physical constraints.

[0024] According to one aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing program instructions that are executed by the processor, the processor invoking the program instructions to execute the surface temperature learning-based dense reconstruction method based on generalized spatiotemporal spectrum fusion.

[0025] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the aforementioned surface temperature learning-intensive reconstruction method based on generalized spatiotemporal spectrum integration.

[0026] The aforementioned technical solution provides a dense reconstruction method for land surface temperature based on generalized spatiotemporal spectral integration. This method comprehensively utilizes the complementary advantages of multi-source satellite observations and land surface process model simulations in the temporal, spatial, and spectral dimensions. It designs a joint loss function combining a multi-dimensional feature collaborative sensing module and physical knowledge constraints, ultimately generating hourly high-dynamic, all-weather land surface temperature sequences. This method initializes the modeling of atmospheric-surface radiative transfer and surface energy balance theory based on multi-source satellite-model auxiliary parameters. It extracts generalized spatiotemporal-spectral multi-dimensional nonlinear residual features through a dual-branch deep learning network to compensate for systematic estimation biases in the aforementioned physical processes. Based on this, a model is further introduced to simulate land surface temperature and characterize the initial thermal state field of the land surface. An adaptive feature fusion module aggregates multi-branch land surface temperature prediction results, improving the accuracy and stability of the reconstruction method. This invention fully considers the multi-dimensional complementary characteristics of multi-source data and the physical mechanisms corresponding to rapid changes in land surface temperature, achieving a comprehensive improvement in the spatiotemporal continuity and physical interpretability of land surface temperature data, providing important data support for finely characterizing the structure and changes of the land surface thermal field.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] (1) For the task of dense reconstruction of surface temperature, we fully explore the complementary advantages of multi-source satellite observation and public land surface process model simulation in the generalized time-space-spectrum dimension, realize the integrated collaborative modeling of multi-dimensional information at the hourly scale, generate a 1-kilometer hourly spatiotemporally continuous dense sequence of surface temperature, and break through the theoretical limitations of the method that only relies on the fusion of spatiotemporal or generalized space-spectrum information.

[0029] (2) Based on deep learning technology, a generalized spatiotemporal-spectral integrated fusion framework is constructed to effectively characterize the complex nonlinear relationship between multi-source satellite-mode auxiliary parameters, and alleviate the shortcomings of linear models such as harmonic regression in describing the high dynamic changes of surface temperature.

[0030] (3) The design of a joint loss function that takes into account physical constraints guides the optimization direction of the deep learning network. It takes into account the numerical and physical consistency between the reconstruction results and the real satellite surface temperature data, effectively reduces the initial modeling bias and uncertainty of the surface physical process, improves the black box learning process of the network, and significantly enhances the spatiotemporal generalization ability of the fusion framework and the physical interpretability of the reconstruction results. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 The overall flowchart of the surface temperature learning-intensive reconstruction method based on generalized spatiotemporal spectrum integration is provided for the embodiments of the present invention.

[0033] Figure 2 The diagram shows the results provided for an embodiment of the present invention. Detailed Implementation

[0034] It should be noted that:

[0035] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0036] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0038] Please refer to the appendix. Figure 1 This invention provides a method for learning-based dense reconstruction of land surface temperature based on the integrated fusion of generalized spatiotemporal spectra, comprising the following steps:

[0039] Step S1: Obtain daily satellite surface temperature and satellite auxiliary parameters, and drive the common land surface process model simulation based on meteorological forcing data and surface state data to obtain model simulation parameters, so as to construct a spatiotemporally matched multi-source satellite-model training sample library. The specific steps are as follows:

[0040] Step S11: Acquire satellite surface temperature, satellite auxiliary parameters, meteorological forcing data, and surface condition data.

[0041] This embodiment acquired daily satellite surface temperature (LST) over a 1-kilometer radius, satellite auxiliary parameters, meteorological forcing data, and surface condition data. The satellite auxiliary parameters included surface emissivity (…). Normalized Difference Vegetation Index (NDVI) and Digital Elevation Model (DEM) data.

[0042] Step S12: Resample the acquired digital elevation model data to the spatial resolution corresponding to the satellite surface temperature.

[0043] In this embodiment, the acquired digital elevation model (DEM) data is resampled to a spatial resolution of 1 km for satellite surface temperature.

[0044] Step S13: Repair the surface emissivity using a time-series reconstruction method based on one-dimensional variational filtering. To address noise and missing values ​​in the normalized difference vegetation index (NDVI) and other indicators, the stability of the dense surface temperature reconstruction algorithm is enhanced. The specific form of the temporal reconstruction method can be expressed as:

[0045]

[0046] in, It is the sequence of results obtained from the solution. This is under ideal conditions. and NDVI sequence, The number of observations within a year, where t is the observation time. This represents the total number of samples throughout the entire time series. These are the weighting coefficients for the corresponding time positions. For the current time t The results of the NDVI repair. This represents the repair result at the next time step t+1. This represents the repair result from the previous time step t-1. For original satellite observations, , The regularization coefficient is . This is the result of the repair on this day the following year. The first item is the data fidelity item, used for constraints. Repair results with NDVI Compared with original satellite observations There is no excessive deviation between them. The second term is a time-series smoothing term, used to suppress high-frequency noise in the observation sequence. The third term is an interannual similarity term, used to characterize the periodicity of vegetation growth and surface emissivity changes on an annual scale.

[0047] Step S14: Using a sliding window filtering method, the satellite surface temperature (in this embodiment, thermal infrared remote sensing surface temperature data) of completely cloudless areas within the window is sampled in blocks, and surface temperature samples under ideal clear sky conditions are extracted as label data for subsequent training. In this embodiment, the sliding window size is 40×40, and the overlap area of ​​adjacent sliding windows is 20 pixels. It should be noted that the size of the sliding window and the overlap area of ​​adjacent sliding windows can be set according to actual needs and are not limited here.

[0048] Step S15: Drive the common land surface process model to perform simulation based on meteorological forcing data and surface state data to obtain model simulation parameters.

[0049] In this embodiment, hourly model simulation parameters for 1 kilometer were generated. These parameters include the model-simulated surface temperature (…). Model simulation of upward longwave radiation ( Model simulation of downlink longwave radiation ( ), model simulation of air temperature ( Model simulation of net radiation ( ) and model simulation of aerodynamic impedance ( ).

[0050] Step S16: Compare satellite surface temperature (LST) and model-simulated surface temperature (LST). Daily-scale composite processing was performed to construct a reference time-phase satellite-model land surface temperature data pair.

[0051] Step S17: Using the real satellite surface temperature labels and multi-source satellite-model auxiliary parameters obtained in step S14 under ideal clear sky conditions, construct a training sample library for dense reconstruction of surface temperature.

[0052] In step S17, the multi-source satellite-model auxiliary parameters include satellite auxiliary parameters, model simulation parameters, and reference temporal satellite-model land surface temperature data pairs. The training sample library for dense reconstruction of land surface temperature is the spatiotemporally matched multi-source satellite-model training sample library. The multi-source satellite-model training sample library consists of several multi-source satellite-model training samples. Each multi-source satellite-model training sample includes multi-source satellite-model auxiliary parameters and corresponding real satellite land surface temperature labels under ideal clear sky conditions.

[0053] Step S2: Based on the complementary advantages of multi-source satellite-mode auxiliary parameters in the three dimensions of time, space and generalized spectral phase, construct a generalized spatiotemporal-spectral integrated fusion network for dense reconstruction of surface temperature.

[0054] In step S2, the generalized spatiotemporal-spectral integrated fusion network uses spatiotemporally matched multi-source satellite-model auxiliary parameters as input to construct a structurally symmetrical dual-branch deep learning fusion architecture. The two branches respectively initialize and model the atmospheric-surface radiation transfer process and the surface energy balance process. Through a multi-dimensional feature collaborative perception module, the multi-source satellite-model auxiliary parameters are collaboratively perceived in three dimensions: time, space, and generalized spectral phase, and the optimized surface temperature is output. Through an adaptive feature fusion module, the optimized surface temperature output by each branch is aggregated with the initial thermal state field provided by the land surface process model, and finally, the feature reconstruction module outputs the final predicted surface temperature.

[0055] It should be noted that the multi-dimensional feature collaborative perception module actually achieves collaborative perception of multi-dimensional features by jointly extracting features of the multi-source satellite-mode auxiliary parameters in three dimensions: time, space, and generalized spectral phase. It also integrates the complementary advantages of these three dimensions in terms of spatial texture details, temporal continuous evolution, and spectral phase feature correlation. Understandably, "jointly" refers to the fact that this invention can extract features in these three dimensions (e.g., spatiotemporal compensation features and generalized spatial-spectral compensation features) through deep learning. The extracted features in these three dimensions have complementary characteristics in terms of spatial details, temporal continuity, and spectral phase feature correlation.

[0056] Step S21: Based on the atmospheric-surface radiative transfer theory, the surface emissivity, model simulation of upward longwave radiation, and model simulation of downward longwave radiation are used to obtain the initial estimate of the surface temperature during the radiative transfer process.

[0057] It should be noted that the initial estimate of surface temperature in the radiative transfer process is used to describe the surface's response to solar radiation. Specifically, the upward longwave radiation at the surface consists of surface-emitted radiation and reflections of downward longwave radiation. Based on the Stefan-Boltzmann law, the initial expression for surface temperature in the radiative transfer process is as follows:

[0058]

[0059] in, This is an initial estimate of the land surface temperature based on the radiative transfer process. This is the Stefan-Boltzmann constant, with a value of 5.67 × 10⁻⁶. -8 W / (m 2 K 4 ); For surface emissivity, To simulate upward longwave radiation in the model, The model simulates downlink longwave radiation.

[0060] Step S22: Construct a generalized spatiotemporal-spectral multidimensional feature collaborative sensing module for radiative transfer processes. Utilize a residual dense network to jointly extract the spatiotemporal features of multi-temporal surface temperature and the generalized spatiotemporal features of multi-source heterogeneous parameters. This mitigates the systematic bias in radiative transfer theory caused by parameter approximations or simplified process expressions. Specifically, it can be expressed as:

[0061]

[0062] in, To compensate for the spatiotemporal features extracted from the initial modeling bias of radiative transfer theory, To compensate for the generalized spatial spectrum features extracted for the initial modeling bias in radiative transfer theory, and They are respectively and The corresponding feature extractor. To simulate surface temperature in the model, and These are the diurnal composite data of satellite surface temperature and model-simulated surface temperature from step S16, used to provide a reference time phase satellite-model surface temperature data pair. For digital elevation model data, For surface emissivity, To simulate air temperature in the model, To simulate aerodynamic impedance for the model, To simulate net radiation for the model.

[0063] Furthermore, Vegetation cover is calculated using the Normalized Difference Vegetation Index (NDVI), and the calculation method is as follows:

[0064]

[0065] in, and The NDVI thresholds are set for bare soil areas and areas with complete vegetation cover, for example, 0.2 and 0.86 respectively.

[0066] Furthermore, Due to water vapor saturation difference, The slope of the saturated vapor pressure curve can be calculated using air temperature and the Penman-Monteith equation, specifically in the following form:

[0067]

[0068] in, This is the average of the day's highest and lowest temperatures. To calculate intermediate variables, for Input variables, and These are the day's highest and lowest temperatures, respectively.

[0069] Step S23: Based on the initial estimate of the surface temperature from the radiative transfer process, and combining the spatiotemporal characteristics of the multi-temporal surface temperature with the generalized spatial-spectral characteristics of the multi-source heterogeneous parameters, the surface temperature optimized by the atmospheric-surface radiative transfer process initialization modeling and the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module is obtained, which can be specifically expressed as:

[0070]

[0071] in, The surface temperature is the result of initial modeling of the atmospheric-surface radiative transfer process and optimization by the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module. This is an initial estimate of the land surface temperature based on the radiative transfer process. This is a collaborative sensing module for multi-dimensional features of the radiative transfer process. and These are the spatiotemporal and generalized spatial spectrum compensation features extracted to address the initialization modeling bias in radiative transfer theory.

[0072] Step S24: Based on the theory of surface energy balance, the initial model of the energy balance process is performed using vegetation cover, model simulated air temperature, model simulated net radiation, and model simulated aerodynamic impedance to obtain the initial surface temperature of the energy balance process.

[0073] In step S24, the surface energy balance theory, through comprehensive modeling of processes such as the hydrological cycle and vegetation transpiration, constructs a conservation relationship for the interconversion of water and heat fluxes at the Earth's surface, which is a key mechanism for understanding the dynamic changes in surface temperature. This theory posits that net surface radiation can be expressed as the sum of sensible surface heat flux, latent surface heat flux, and soil heat flux. By parameterizing this process using multi-source satellite-model auxiliary parameters, the initial surface temperature based on the surface energy balance theory can be expressed as:

[0074]

[0075] in, Initialize the surface temperature for the surface energy balance theory. To simulate air temperature in the model, To simulate aerodynamic impedance for the model, To simulate net radiation in the model, The density of air is taken as 1.29 kg / m³. 3 , The specific heat at constant pressure of air is 1.005 kJ / (kg). K), and The values ​​are the ratios of soil heat flux to net radiation under vegetation and bare soil cover, respectively, which are 0.05 and 0.4. The wet-bulb and dry-bulb constants can be calculated from atmospheric pressure, and their specific form is:

[0076]

[0077] in, Atm, the pressure is taken as 1.01325 × 10⁻⁶. 5 Pa.

[0078] Step S25: Construct a generalized spatiotemporal-spectral multidimensional feature collaborative sensing module for energy balance processes. Utilize a residual dense network to extract joint features of multi-source satellite-mode auxiliary parameters to mitigate initial estimation biases in the energy balance process. The specific process can be represented as follows:

[0079]

[0080] in, To address the limitations of energy balance theory in obtaining spatiotemporal compensation characteristics, To address the limitations of energy balance theory, a generalized spatial spectrum compensation feature was obtained. and They are respectively and Corresponding feature extractors, To simulate upward longwave radiation in the model, To simulate the downward long-wave radiation of the model.

[0081] Step S26: Based on the initial surface temperature of the energy balance process, and combined with the joint characteristics of multi-source satellite-model auxiliary parameters, the surface temperature optimized by the surface energy balance process initialization modeling and the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module is obtained, which can be specifically expressed as:

[0082]

[0083] in, The surface temperature is the result of initial modeling of the surface energy balance process and optimization by the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module. To simulate air temperature in the model, This is a multi-dimensional feature collaborative sensing module for the energy balance process. It should be noted that the first two terms in formula (10) are equivalent to the initial estimates of the surface temperature. There are three initial estimates of the surface temperature in formula (7), but the last term has a large uncertainty, so it is directly replaced by the deep learning network that follows. The surface temperature is initialized by modeling the surface energy balance process and optimized by the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module.

[0084] Step S27: Construct an adaptive feature fusion module and a feature reconstruction module. The adaptive feature fusion module aggregates the optimized land surface temperature and the initial thermal state field of the land surface simulated by the common land surface process model (i.e., the model simulated land surface temperature) to obtain fused features. The feature reconstruction module reconstructs the fused features to output the final prediction result of the land surface temperature.

[0085] In step S27, the three types of surface temperature data (surface temperature optimized by the multi-dimensional feature collaborative sensing module constructed for the radiative transfer process in step S23, surface temperature optimized by the multi-dimensional feature collaborative sensing module constructed for the energy balance process in step S26, and the initial thermal state field of the surface simulated by the common land surface process model) are aggregated by the adaptive feature fusion module, and the fused features are converted into surface temperature output by the feature reconstruction module, specifically in the following form:

[0086]

[0087] in, The final predicted surface temperature output after adaptive fusion and feature reconstruction is shown below. To simulate surface temperature in the model, For adaptive feature fusion networks, This is the feature reconstruction module.

[0088] Step S3: Using the multi-source satellite-mode training sample library established in Step S1, train the generalized spatiotemporal-spectral integrated fusion network constructed in Step S2 until convergence using a joint loss function that takes into account physical constraints.

[0089] Step S31: Construct a joint loss function that takes into account physical constraints, including: constructing a numerical consistency loss using the final predicted result of the reconstructed land surface temperature and the real satellite land surface temperature label under ideal clear sky conditions; constructing physical consistency losses for radiative transfer and energy balance processes respectively using the optimized land surface temperature output by two generalized spatiotemporal-spectral multidimensional feature collaborative sensing modules and the real satellite land surface temperature label under ideal clear sky conditions; and constructing a joint loss function based on the numerical consistency loss, the physical consistency loss of radiative transfer and energy balance processes.

[0090] It should be noted that step S31 aims to guide the generalized spatiotemporal-spectral integrated fusion network to stable convergence under the condition of multi-dimensional feature synergy by adding physical knowledge constraints during training, thereby improving the generalization ability of the method and the physical interpretability of the reconstruction results. The joint loss function can be expressed as:

[0091]

[0092] in, Let the joint loss function of the network be... For numerical consistency loss, and These represent the physical consistency loss in radiative transfer and energy balance processes, respectively. These are weighting coefficients used to adjust the relative contribution of each sub-loss function in the network. In this embodiment, the weighting coefficients... All values ​​are 1. It should be noted that the weighting coefficients... Other values ​​can also be set according to actual needs; no restrictions are imposed here.

[0093] Specifically, the numerical consistency loss is defined as:

[0094]

[0095] in, Let i be the number of training samples, and i be the index of the training sample. This represents a real satellite surface temperature label under ideal clear sky conditions. For L2 norm loss, This indicates the final predicted result for Earth's surface temperature.

[0096] The generalized spatiotemporal-spectral multidimensional feature collaborative sensing physical consistency loss of the radiative transfer process is defined as:

[0097]

[0098] in, The surface temperature is the result of initial modeling of the atmospheric-surface radiation transfer process and optimization by the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module.

[0099] The generalized spatiotemporal-spectral multidimensional feature collaborative sensing physical consistency loss of the energy balance process is defined as:

[0100]

[0101] in, The surface temperature is initialized by modeling the surface energy balance process and optimized by the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module.

[0102] Step S32: Train the generalized spatiotemporal-spectral integrated fusion network built in step S2 based on the multi-source satellite-mode training sample library constructed in step S1. At the same time, use the joint loss function designed in step S31 that takes into account physical constraints to constrain the training process of the generalized spatiotemporal-spectral integrated fusion network until the network converges and the optimal deep learning model is obtained, that is, the trained generalized spatiotemporal-spectral integrated fusion network.

[0103] Step S4: Using the generalized spatiotemporal-spectral integrated fusion network trained in Step S3, the multi-source satellite-mode auxiliary parameters of the target time phase are fused and processed to output an hourly high dynamic all-weather land surface temperature sequence.

[0104] In step S4, the multi-source satellite-model auxiliary parameters for the target time phase include satellite auxiliary parameters, model simulation parameters, and reference time phase satellite-model land surface temperature data pairs. In this embodiment, the multi-source satellite-model auxiliary parameters at the 1 km hourly level for the target time phase are jointly input into the generalized spatiotemporal-spectral integrated fusion network trained in step S32, and the resulting 1 km hourly resolution all-weather land surface temperature sequence is tested and generated (please refer to the appendix). Figure 2 This allows for the intensive reconstruction of satellite surface temperature data.

[0105] Based on the same technical concept as the aforementioned embodiments, this invention also provides a surface temperature learning-based dense reconstruction system based on generalized spatiotemporal-spectral integrated fusion, implemented using the aforementioned surface temperature learning-based dense reconstruction method, comprising: a target multi-source data acquisition module for acquiring multi-source satellite-model auxiliary parameters of the target time phase; and a surface temperature deep learning-based dense reconstruction module for fusing the multi-source satellite-model auxiliary parameters of the target time phase using a trained generalized spatiotemporal-spectral integrated fusion network to output an hourly high-dynamic all-weather surface temperature sequence; wherein, the training of the generalized spatiotemporal-spectral integrated fusion network includes: constructing a spatiotemporally matched multi-source satellite-model training sample library; and constructing a generalized spatiotemporal-spectral integrated fusion network for dense surface temperature reconstruction. A fusion network is constructed using spatiotemporally matched multi-source satellite-model auxiliary parameters as input, forming a symmetrical dual-branch deep learning fusion architecture. The two branches initialize and model the atmospheric-surface radiative transfer process and the surface energy balance process, respectively. A multi-dimensional feature collaborative perception module jointly extracts complementary information from the multi-source satellite-model auxiliary parameters across time, space, and generalized spectral phase dimensions, outputting an optimized surface temperature. An adaptive feature fusion module aggregates the optimized surface temperature output from the two branches with the initial thermal state field provided by the land surface process model, and a feature reconstruction module outputs the final predicted surface temperature. Based on the multi-source satellite-model training sample library, the generalized spatiotemporal-spectral integrated fusion network is trained to convergence using a joint loss function that considers physical constraints.

[0106] Based on the same technical concept as the foregoing embodiments, this embodiment of the invention also provides an electronic device, including a memory and a processor. The memory stores program instructions that are executed by the processor. The processor calls the program instructions to execute the surface temperature learning dense reconstruction method based on generalized spatiotemporal spectrum integration.

[0107] Based on the same technical concept as the foregoing embodiments, this embodiment of the invention also provides a non-transitory computer-readable storage medium that stores computer instructions that cause the computer to execute the surface temperature learning-intensive reconstruction method based on generalized spatiotemporal spectrum integration.

[0108] In summary, the above embodiments provide a dense reconstruction method for land surface temperature based on generalized spatiotemporal spectral integration. This method leverages the complementary advantages of multi-source satellite observations and land surface process model simulations in the temporal, spatial, and spectral dimensions. It designs a joint loss function combining a multi-dimensional feature collaborative sensing module and physical knowledge constraints, ultimately generating hourly high-dynamic, all-weather land surface temperature sequences. This method initializes the modeling of atmospheric-surface radiative transfer and surface energy balance theory based on multi-source satellite-model auxiliary parameters. A dual-branch deep learning network extracts generalized spatiotemporal-spectral multi-dimensional nonlinear residual features to compensate for systematic estimation biases in the aforementioned physical processes. Based on this, a model is further introduced to simulate land surface temperature and characterize the initial thermal state field of the land surface. An adaptive feature fusion network aggregates multi-branch land surface temperature prediction results, improving the accuracy and stability of the reconstruction method. This invention fully considers the multi-dimensional complementary characteristics of multi-source data and the physical mechanisms corresponding to rapid changes in land surface temperature, achieving a comprehensive improvement in the spatiotemporal continuity and physical interpretability of land surface temperature data, providing important data support for finely characterizing the structure and changes of the land surface thermal field.

[0109] 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 or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A dense reconstruction method for land surface temperature based on generalized spatiotemporal spectrum integration, characterized in that, include: Acquire multi-source satellite-mode-aided parameters for the target temporal phase; The trained generalized spatiotemporal-spectral integrated fusion network is used to fuse the multi-source satellite-model auxiliary parameters of the target time phase, outputting an hourly high-dynamic all-weather land surface temperature sequence; wherein, the training of the generalized spatiotemporal-spectral integrated fusion network includes: Construct a spatiotemporally matched multi-source satellite-pattern training sample library; A generalized spatiotemporal-spectral integrated fusion network for dense reconstruction of land surface temperature is constructed. The fusion network takes spatiotemporally matched multi-source satellite-model auxiliary parameters as input and constructs a structurally symmetrical dual-branch deep learning fusion architecture. The two branches initialize and model the atmospheric-surface radiation transfer process and the land surface energy balance process, respectively. Through a multi-dimensional feature collaborative perception module, the multi-source satellite-model auxiliary parameters are collaboratively perceived in three dimensions: time, space, and generalized spectral phase, and the optimized land surface temperature is output. Through an adaptive feature fusion module, the optimized land surface temperature output by the two branches is aggregated with the initial thermal state field provided by the land surface process model. Finally, the feature reconstruction module outputs the final predicted result of land surface temperature. Based on the multi-source satellite-mode training sample library, the generalized spatiotemporal-spectral integrated fusion network is trained to convergence using a joint loss function that takes into account physical constraints. The construction of the generalized spatiotemporal-spectral integrated fusion network includes: Based on the atmospheric-surface radiative transfer theory, the radiative transfer process is initialized by using surface emissivity, model simulation of upward longwave radiation, and model simulation of downward longwave radiation, and the initial estimated value of surface temperature is obtained. A generalized spatiotemporal-spectral multidimensional feature collaborative sensing module for radiative transfer processes is constructed. A residual dense network is used to jointly extract spatiotemporal and generalized spatiotemporal compensation features for the initial modeling bias of radiative transfer theory. Then, combined with the initial estimate of surface temperature, the surface temperature is optimized. Based on the theory of surface energy balance, the initial modeling of the energy balance process is carried out using vegetation cover, model simulated air temperature, model simulated net radiation, and model simulated aerodynamic impedance to obtain the initial estimate of surface temperature. A generalized spatiotemporal-spectral multidimensional feature collaborative sensing module for energy balance processes is constructed. A residual dense network is used to jointly extract spatiotemporal and generalized spatiotemporal compensation features that address the limitations of energy balance theory. Then, combined with the initial estimate of surface temperature, the surface temperature is optimized. The adaptive feature fusion module aggregates the simulated surface temperature and two optimized surface temperatures to obtain fused features, which are then reconstructed into the final predicted surface temperature result.

2. The surface temperature learning-based dense reconstruction method based on generalized spatiotemporal spectrum fusion as described in claim 1, characterized in that, A residual dense network is used to jointly extract spatiotemporal and generalized spatial spectrum compensation features for the initial modeling bias in radiative transfer theory. These features are then combined with the initial surface temperature estimate to obtain an optimized surface temperature, as shown in the following formula: ， ， in, To optimize the surface temperature corresponding to the radiative transfer process, This represents the initial estimate of the surface temperature during the radiative transfer process. This is a collaborative sensing module for multi-dimensional features of the radiative transfer process. To compensate for the spatiotemporal features extracted from the initial modeling bias of radiative transfer theory, To compensate for the generalized spatial spectrum features extracted for the initial modeling bias in radiative transfer theory, and for and Corresponding feature extractors, To simulate surface temperature in the model, and For reference time phase satellite-model surface temperature data pairs, For digital elevation model data, For vegetation coverage, For surface emissivity, To simulate air temperature in the model, To simulate aerodynamic impedance for the model, To simulate net radiation in the model, Due to water vapor saturation difference, This represents the slope of the saturated water vapor pressure curve.

3. The surface temperature learning-based dense reconstruction method based on generalized spatiotemporal spectrum fusion as described in claim 2, characterized in that, By jointly extracting spatiotemporal and generalized spatial spectrum compensation features to address the limitations of energy balance theory using residual dense networks, and then combining these features with the initial estimate of land surface temperature, an optimized land surface temperature is obtained. The corresponding formula is as follows: ， ， in, To optimize the surface temperature corresponding to the energy balance process, To simulate air temperature in the model, air density, The specific heat of air at constant pressure. The dry-bulb and wet-bulb constants are... This is a collaborative sensing module for multi-dimensional features of the energy balance process; To address the limitations of energy balance theory in obtaining spatiotemporal compensation characteristics, To address the limitations of energy balance theory, a generalized spatial spectrum compensation feature was obtained. and They are respectively and Corresponding feature extractors, To simulate upward longwave radiation in the model, The model simulates downlink longwave radiation.

4. The surface temperature learning-based dense reconstruction method based on generalized spatiotemporal spectrum fusion as described in claim 1, characterized in that, Based on the multi-source satellite-mode training sample library, the generalized spatiotemporal-spectral integrated fusion network is trained to convergence using a joint loss function that takes into account physical constraints, including: The joint loss function that takes into account physical constraints is constructed as follows: numerical consistency loss is constructed using the final predicted result of the reconstructed land surface temperature and the real satellite land surface temperature label under ideal clear sky conditions; physical consistency loss corresponding to the radiative transfer and energy balance processes is constructed using the optimized land surface temperature output by two generalized spatiotemporal-spectral multidimensional feature collaborative sensing modules and the real satellite land surface temperature label under ideal clear sky conditions; and a joint loss function is constructed based on the numerical consistency loss, the radiative transfer and the physical consistency loss of the energy balance processes. Based on the multi-source satellite-mode training sample library, the generalized spatiotemporal-spectral integrated fusion network is trained using a joint loss function that takes into account physical constraints until the generalized spatiotemporal-spectral integrated fusion network converges, thus obtaining the trained generalized spatiotemporal-spectral integrated fusion network.

5. The surface temperature learning-based dense reconstruction method based on generalized spatiotemporal spectrum fusion as described in claim 4, characterized in that, The formula corresponding to the joint loss function is: ， ， ， ， in, Let be the total loss function of the generalized spatiotemporal-spectral integrated fusion network. For numerical consistency loss, and These represent the physical consistency loss in radiative transfer and energy balance processes, respectively. These are the weight coefficients of each sub-loss function. Let i be the number of training samples, and i be the index of the training sample. This represents a real satellite surface temperature label under ideal clear sky conditions. For L2 norm loss, This indicates the final predicted result for surface temperature. The surface temperature is the result of initial modeling of the atmospheric-surface radiative transfer process and optimization by the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module. The surface temperature is initialized by modeling the surface energy balance process and optimized by the generalized spatiotemporal-spectral multidimensional feature collaborative sensing module.

6. The method for learning-based dense reconstruction of land surface temperature based on generalized spatiotemporal spectrum fusion as described in claim 1, characterized in that, The construction of the spatiotemporal matching multi-source satellite-pattern training sample library includes: Acquire satellite surface temperature, satellite auxiliary parameters, meteorological forcing data, and surface condition data; The satellite auxiliary parameters are preprocessed; The satellite surface temperature of the completely cloudless area within the window is sampled in blocks to extract the real satellite surface temperature label under ideal clear sky conditions; Simulations of common land surface processes are performed using meteorological forcing data and land surface state data to obtain model simulation parameters; Diurnal-scale composite processing is performed on satellite surface temperature and model-simulated surface temperature to construct a reference temporal satellite-model surface temperature data pair; Using real satellite surface temperature labels and multi-source satellite-model auxiliary parameters, a spatiotemporally matched multi-source satellite-model training sample library is constructed; the multi-source satellite-model auxiliary parameters include satellite auxiliary parameters, model simulation parameters, and reference temporal satellite-model surface temperature data pairs.

7. A dense reconstruction system for land surface temperature learning based on the integrated fusion of generalized spatiotemporal spectrum, characterized in that, The method employs the land surface temperature learning-based dense reconstruction method based on generalized spatiotemporal spectrum fusion as described in any one of claims 1 to 6, comprising: The target multi-source data acquisition module is used to acquire multi-source satellite-mode auxiliary parameters of the target in time phase; The surface temperature deep learning-based dense reconstruction module is used to fuse multi-source satellite-model auxiliary parameters of the target time phase using a trained generalized spatiotemporal-spectral integrated fusion network, outputting an hourly high-dynamic all-weather surface temperature sequence. The training of the generalized spatiotemporal-spectral integrated fusion network includes: constructing a spatiotemporally matched multi-source satellite-model training sample library; constructing a generalized spatiotemporal-spectral integrated fusion network for dense surface temperature reconstruction; the fusion network uses spatiotemporally matched multi-source satellite-model auxiliary parameters as input, constructing a structurally symmetrical two-branch deep learning fusion architecture, with the two branches respectively processing atmospheric parameters. - Initial modeling of the surface radiative transfer process and the surface energy balance process is performed. A multi-dimensional feature collaborative sensing module is used to perform time, space, and generalized spectral phase feature collaborative sensing of the multi-source satellite-model auxiliary parameters, outputting an optimized surface temperature. An adaptive feature fusion module aggregates the optimized surface temperature output from the two branches with the initial thermal state field provided by the land surface process model. Finally, a feature reconstruction module outputs the final predicted surface temperature. Based on the multi-source satellite-model training sample library, a joint loss function that considers physical constraints is used to train the generalized time-space-spectral integrated fusion network until convergence. The construction of the generalized spatiotemporal-spectral integrated fusion network includes: based on the atmospheric-surface radiative transfer theory, initial modeling of the radiative transfer process is performed using surface emissivity, model-simulated upward longwave radiation, and model-simulated downward longwave radiation to obtain an initial estimate of surface temperature; a generalized spatiotemporal-spectral multidimensional feature collaborative sensing module for the radiative transfer process is constructed, and spatiotemporal and generalized spatiotemporal spectral compensation features for the initial modeling bias of the radiative transfer theory are jointly extracted using a residual dense network, and then combined with the initial estimate of surface temperature to obtain an optimized surface temperature; based on the surface energy balance theory, vegetation cover is used to... Initial modeling of the energy balance process is performed using land cover, simulated air temperature, simulated net radiation, and simulated aerodynamic impedance to obtain an initial estimate of land surface temperature. A generalized spatiotemporal-spectral multidimensional feature collaborative sensing module for the energy balance process is constructed. A residual dense network is used to jointly extract spatiotemporal and generalized spatiotemporal compensation features to address the limitations of energy balance theory. These features are then combined with the initial land surface temperature estimate to obtain an optimized land surface temperature. The adaptive feature fusion module aggregates the simulated land surface temperature and the two optimized land surface temperatures to obtain fused features. These fused features are then reconstructed into the final predicted land surface temperature.

8. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to execute the surface temperature learning-intensive reconstruction method based on generalized spatiotemporal spectrum fusion as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the surface temperature learning-intensive reconstruction method based on generalized spatiotemporal spectrum fusion as described in any one of claims 1 to 6.

Images