Large-small model surface temperature and emissivity inversion method, system and device based on AI-Agent nested coordination

Abstract

The invention discloses a large-small model surface temperature and emissivity inversion method, system and device based on AI-Agent nested coordination. Aiming at the multi-band thermal infrared data of the spaceborne radiometer, the invention aims to solve the precision limitation of the traditional temperature-emissivity separation algorithm in the aspects of atmospheric residual error correction, stripe noise, surface heterogeneity and the like. By fusing physical expert knowledge and a deep learning mechanism, an expert-multilayer perceptron hybrid network (EMHN) is constructed, and AI-Agent is introduced to coordinate dynamic routing and knowledge distillation of a large model (expert hybrid framework) and a small model (multilayer perceptron), so that high-precision simultaneous inversion of LST and LSE is realized. According to the specific scheme, physical logic reasoning and simulation data verification are carried out based on a radiation transfer equation, and multiband input configuration is optimized; an AI-Agent system is deployed to drive an EMHN architecture, and nonlinear feature extraction and task allocation are achieved; and an iterative label refining and Adam optimization strategy is adopted to form closed-loop parameter fine adjustment, so that the robustness of the model is improved. According to the method, the inversion precision is remarkably improved, the generalization ability is enhanced, the method is suitable for climate monitoring, disaster assessment and agricultural water resource management under complex terrains, and an efficient and intelligent remote sensing parameter estimation solution is provided.

Description

Technical Field

[0001] This invention belongs to the fields of remote sensing technology and artificial intelligence, specifically involving a method for retrieving land surface temperature and emissivity from thermal infrared remote sensing data using an AI-Agent to coordinate large and small models. This technology is primarily applied in climate monitoring, agricultural management, and environmental assessment. Furthermore, this invention integrates subfields of geophysical parameter inversion, such as radiative transfer model simulation and satellite brightness temperature data processing, while incorporating machine learning applications, including expert hybrid architectures, nonlinear feature extraction using multilayer perceptrons, and iterative optimization of label refinement. According to the International Patent Classification (IPC), this invention can be classified under G06N 3 / 04 (neural network architecture) and G01N 21 / 17 (remote sensing measurements based on infrared radiation) to ensure its standardization and operability in multi-channel remote sensing data-driven inversion. Background Technology

[0002] Land surface temperature (LST) and land surface emissivity (LSE), as core parameters of the land-atmosphere interface energy balance, play a crucial role in global climate change monitoring, agricultural drought assessment, environmental and ecological evaluation, and urban heat island effect analysis. High-precision spatiotemporal distribution data of LST and LSE help reveal the dynamics of surface thermal state, supporting weather forecast model optimization, disaster risk early warning, and sustainable land management decisions. However, traditional LST and LSE inversion methods face multiple challenges, including correction residuals due to atmospheric parameter uncertainties, data quality issues caused by fringe noise and sensor aging, high sensitivity to low-spectral-contrast surfaces, neglect of surface heterogeneity and geometric effects, and insufficient generalization ability in high-humidity or complex terrain areas. These shortcomings are particularly prominent in practical applications, resulting in LST errors of up to 2-3 K, which limits the reliability of refined applications.

[0003] Existing inversion algorithms mainly include physical model-driven methods, such as single-channel algorithms, split-window algorithms, and Temperature-Emissivity Separation (TES) algorithms. These methods exhibit high accuracy under homogeneous surface conditions, but their accuracy is often limited by atmospheric absorption, emissivity heterogeneity, and underlying surface complexity. For example, single-channel algorithms rely on accurate emissivity and atmospheric water vapor estimates, split-window algorithms utilize differences between adjacent channels to eliminate atmospheric effects, and the TES algorithm uses a minimum-maximum difference mechanism to simultaneously separate LST and LSE. However, these traditional methods neglect atmospheric water vapor and aerosol variability, multiple scattering, and shading effects, especially in complex terrains such as the Qinghai-Tibet Plateau and eastern hilly areas of China, where their generalization ability is insufficient and cannot meet the requirements for high accuracy. Traditional surface temperature and emissivity inversion methods rely on physical models (such as radiative transfer equations), but they suffer from low accuracy and weak generalization under heterogeneous surface conditions.

[0004] In recent years, the rapid development of artificial intelligence technology has provided a new paradigm for land surface temperature (LST) and emissivity (LSE) inversion. Automated machine learning (AutoML) and deep learning methods have driven the fusion of physics and data, demonstrating superior robustness and accuracy compared to traditional algorithms in thermal infrared remote sensing parameter estimation. For example, convolutional neural networks are used to extract nonlinear features from multi-spectral data, improving inversion adaptability; expert hybrid models iteratively optimize satellite LST and LSE products; and knowledge distillation strategies jointly optimize inversion using large and small models. Despite these significant advancements, existing AI methods still have certain limitations in the joint inversion of LST and LSE from multi-channel thermal infrared data. In particular, the model generalization ability is limited by the diversity of land cover types, complex terrain conditions, and dataset size, and physical consistency and iterative optimization effects need further improvement. Compared with the existing patent ZL202410758049.7 (AI-based large and small model land surface temperature and emissivity inversion method, system, and device), this invention introduces an AI-Agent nested coordination large and small model, quantitatively improving inversion accuracy by 20% (MAE from 0.05 to 0.04).

[0005] To address the shortcomings of existing technologies, this invention proposes a method, system, and device for inverting land surface temperature and emissivity (LST) using a large-small model based on AI-Agent nested coordination. By integrating physical expert knowledge with deep learning mechanisms, it achieves high-precision simultaneous inversion of LST and LSE from multi-band thermal infrared data. This method innovatively integrates an expert-multilayer perceptron hybrid network and iterative optimization strategies, overcoming the limitations of traditional algorithms, improving inversion accuracy and generalization ability, and providing highly reliable data support for climate monitoring and disaster assessment. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses a method, system, and device for inverting land surface temperature (LST) and emissivity (LSE) using a large-small model based on AI-Agent nested coordination. This aims to solve the problem of limited accuracy in existing land surface temperature (LST) and emissivity (LSE) inversion techniques. This invention achieves high-precision simultaneous inversion of LST and LSE for multi-band thermal infrared data by integrating an AI-Agent mechanism with a hybrid architecture of large and small models. Specifically, this invention innovatively integrates physical expert knowledge (such as the radiative transfer equation) and deep learning mechanisms to construct an Expert-MLP Hybrid Network (EMHN) and introduces an iterative optimization strategy to overcome the limitations of traditional temperature-emissivity separation algorithms in terms of atmospheric correction residuals, stripe noise, surface heterogeneity, and insufficient generalization ability. This method, system, and device are suitable for remote sensing applications in complex terrains and diverse land cover types, providing reliable data support for climate change monitoring, disaster assessment, and agricultural water resource management.

[0007] To achieve the above objectives, the technical solution provided by this invention is as follows: First, physical logic reasoning is performed based on the Radiative Transfer Equation (RTE) to establish the theoretical foundation for the model's input-output relationship; second, simulated data is used to verify the impact of multi-band brightness temperature input on inversion accuracy and to optimize the input configuration; third, dynamic task allocation and nonlinear feature extraction are achieved through the EMHN architecture; finally, label refinement and parameter fine-tuning strategies are adopted to form an iterative closed loop, further improving model performance. The core of this solution lies in the coordinating role of the AI-Agent, where the large model is responsible for complex feature extraction and knowledge distillation, while the small model is used for efficient real-time inversion, ensuring that the system operates efficiently in environments with limited computing resources. The technical solution of this invention is described in detail step by step below.

[0008] I. Technical Problems and Solutions

[0009] In existing technologies, LST and LSE inversions primarily rely on physical models, such as single-channel algorithms, split-window algorithms, and TES algorithms. While these methods are effective under ideal conditions, they face challenges such as atmospheric parameter uncertainties, sensor noise, and surface heterogeneity, leading to a significant increase in inversion errors. For example, in high-humidity or complex terrain areas, LST errors can reach 2-3 K, affecting practical applications. This invention addresses these issues by proposing a large-scale model framework based on AI-Agent. The AI-Agent acts as an intelligent agent, dynamically coordinating large models (such as MoE-based expert networks) and small models (such as lightweight MLPs) to achieve a fusion of physics and data-driven approaches. Through iterative optimization, the AI-Agent continuously refines labels and parameters, ensuring that the inversion results conform to the radiative transfer mechanism while improving generalization ability. The core of this approach is to leverage the powerful representational capabilities of large models and the computational efficiency of small models, transferring knowledge through a knowledge distillation mechanism to achieve efficient and high-precision inversion.

[0010] II. Derivation of Parameter Inversion Theory

[0011] The theoretical basis of this invention stems from the radiative transfer equation (RTE), which describes the transmission of surface thermal radiation through the atmosphere. First, the blackbody radiation law is considered as a benchmark for graybody correction. The spectral radiance of an ideal blackbody is expressed as:

[0012]

[0013] Where B(λ,T) represents the blackbody radiance at wavelength λ and temperature T (in W / m²·sr·μm), h represents Planck's constant (6.626 × 10^(-34) J·s), c represents the speed of light (3 × 10^8 m / s), k represents Boltzmann's constant (1.381 × 10^(-23) J / K), λ represents the wavelength (in m), and T represents the temperature (in K). This formula characterizes the blackbody radiation distribution at a specific wavelength under a given temperature, providing a benchmark for actual graybody radiation.

[0014] The actual Earth's surface is an emissive gray body, and the radiation signal is affected by atmospheric absorption and scattering during atmospheric transmission. For the i-th thermal infrared channel, the channel radiance observed by the sensor is expressed as:

[0015]

[0016] in, The i-th channel represents the observed radiance (in W / m²·sr·μm). Let represent the surface emissivity of the i-th channel (dimensionless, range [0,1]), and B(λi,Ts) represent the blackbody radiance at a surface temperature Ts (in K). denoted as the atmospheric transmittance of the i-th channel (dimensionless, range [0,1]). This represents the downward atmospheric radiation of the i-th channel (unit: W / m²·sr·μm). This represents the upward atmospheric radiation in the i-th channel (unit: W / m²·sr·μm). This equation integrates three types of contributions: direct surface emission, reflected downward radiation, and atmospheric path radiation, ensuring physical consistency.

[0017] To simplify the calculation, atmospheric path radiation is further approximated as:

[0018]

[0019] in, Let K be the effective atmospheric temperature (K), which is approximately the near-surface air temperature. This approximation is based on the continuity of the atmospheric profile, and its error is less than 0.2 K under mid-latitude conditions. It simplifies the equation form while preserving the essential characteristics.

[0020] Substituting into the master equation yields the compact form:

[0021]

[0022] This formula reveals four core unknowns: surface temperature Near-surface air temperature Atmospheric transmittance (Each channel is independent, but is simultaneously determined by atmospheric water vapor) and channel emissivity. (Each channel is independent, but all are determined by the surface type). A single channel generates only one equation, but contains four unknowns, leading to underdetermined problems; multi-channel input provides information redundancy and supports parameter separation. This invention is mainly aimed at sensors with more than four thermal infrared brightness temperature bands (such as ASTER / MODIS sensors).

[0023] In physical logic reasoning, the applicability of this model is reflected in the following: First, the model input (brightness temperature vector BT=[BT1,BT2,BT3,BT4,BT5], unit K) and output (Ts, ε=[ε1,ε2,ε3,ε4,ε5]) are directly defined; second, the sensitivity is calculated through the Jacobian matrix.

[0024]

[0025] in, Represents the elements of the Jacobian matrix (units depend on the parameter). This represents unknown parameters (including Ts, Ta, τi, εi).

[0026] This matrix quantifies the contribution of each band to the parameters and is used to evaluate the inversion performance of the band combination, using mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient (PCC) as indicators.

[0027] III. EMHN Model Architecture

[0028] The core of this invention is the AI-Agent-based EMHN architecture, which employs a Mixture of Experts (MoE) framework. This architecture uses a Multilayer Perceptron (MLP) as an expert sub-network to achieve modular processing of ASTER brightness temperature data. The AI-Agent coordinates the large model (MoE expert network for complex feature extraction) and the small model (lightweight MLP for efficient prediction), improving overall efficiency through knowledge distillation to transfer parameters. In the input processing stage, the brightness temperature vector is standardized.

[0029]

[0030] in, This represents the standardized brightness temperature value of the i-th band (dimensionless). This represents the original brightness temperature value (in K). This represents the mean value of the i-th band (in K). This represents the standard deviation of the i-th band (in K). This step balances the scale and alleviates the gradient problem.

[0031] The gating network extracts features from the standardized input, and the output of the l-th layer is:

[0032] )

[0033] in, and The first Layer weight matrix and bias vector, The activation function is (e.g., ReLU). The final gated output is converted into a probability vector by the SoftMax function:

[0034]

[0035] in, Denotes the weight of the k-th expert (dimensionless, ∑ =1), This indicates a gated linear output, where KKK is the number of experts (typically 4-8). This step enables dynamic routing.

[0036] Each MLP expert generates a local output; for the k-th expert:

[0037]

[0038] in, Output for experts (predicting LST or LSE). , , , These are the parameters for the hidden and output layers. This structure captures non-linear mappings. It integrates expert outputs.

[0039]

[0040] in, This indicates the final prediction (LST unit K or LSE dimensionless). This represents the output of the k-th expert. To ensure load balancing, an auxiliary loss is introduced:

[0041]

[0042] in, This represents the load balancing loss (dimensionless). This represents the hyperparameter (typically 0.01). The overall loss function is:

[0043] The overall loss function is:

[0044]

[0045] in, Let N represent the total loss, y represent the number of samples, β represent the true label, and γ represent the balance coefficients (typically 0.1). This represents the physical regularization loss. The architecture achieves knowledge distillation from large models to small models through AI-Agent coordination, ensuring deployment on resource-constrained devices. "def ai_agent_coordinate(big_model, small_models): # Parameter migration big_model_params = big_model.get_params() for small insmall_models: small.set_initial_params(big_model_params) # Fine-tuning small.train(subregion_data)".

[0046] IV. Tag Refinement and Parameter Fine-Tuning Strategies

[0047] This invention introduces an iterative optimization mechanism, with AI-Agent driving label refinement. The Adam optimizer is used to update parameters, and the first-order moment is estimated as follows:

[0048]

[0049] in, Represents a first-order moment vector. This represents the momentum coefficient (typically 0.9). This represents the current gradient.

[0050] The second moment estimate is:

[0051]

[0052] in, Represents a second-order moment vector. This represents the coefficient (typically 0.999). The parameter update formula is:

[0053]

[0054] in, Let η represent the parameter vector, and η represent the learning rate (initially 0.001). , ϵ represents the correction moment and ϵ represents the stability constant (10^(-8)).

[0055]

[0056] in, The adjusted value. For the first The original LST / LSE values ​​from the next iteration (from the initial product dataset collected). These are the values ​​retrieved from the model. This is a relaxation factor (typically 0.5) used to control the adjustment range. This step directly optimizes the acquired data, ensuring physical consistency with the radiative transfer equation.

[0057] V. Stepwise Inversion of Emittance and System Implementation

[0058] To achieve the above objectives, the present invention also provides a system comprising a data acquisition module, an AI-Agent coordination module, an EMHN processing module, and an output module; and a device, such as a computer-readable medium, storing instructions to implement the above method.

[0059] Obtain the inversion dataset of geophysical parameters;

[0060] Based on the inversion dataset, an AI-Agent system is deployed. This system dynamically searches and optimizes the neural network architecture through autonomous perception and decision-making mechanisms, and introduces a pre-trained large model as a guiding model.

[0061] By utilizing AI-Agent coordinated knowledge transfer technology, the complex knowledge of the guiding model is transferred to a simplified model, forming an adaptive agent model;

[0062] The adaptive agent model and the guidance model are integrated through the AI-Agent interaction optimization framework to construct a dynamic joint agent model, which is then used to perform real-time inversion of the target geophysical parameters.

[0063] The process of dynamically searching and optimizing the neural network architecture for optimal deployment of AI-Agent agent systems includes:

[0064] Inversion datasets of geophysical parameters obtained through satellite acquisition;

[0065] The inversion dataset is standardized and preprocessed to obtain processed data;

[0066] The processed data is input into the AI-Agent system, and the agent search space is defined;

[0067] Utilize the autonomous learning mechanism of AI-Agent to initialize architecture exploration;

[0068] The architecture search is guided by a reinforcement learning algorithm, and the network mutation is controlled by a proxy state transition function to traverse a predefined search space.

[0069] Based on the search space, the optimal neural network architecture is obtained through autonomous search within a preset agent iteration cycle.

[0070] Preferably, the agent search space includes: the number of agent levels, the size of each agent unit, the decision function, and the mutation probability;

[0071] The proxy search space is specifically described as follows:

[0072]

[0073] in, It is the number of agency levels. It is the first Level of agency unit size, It is the first Level decision function, It is the first The mutation probability of level.

[0074] The preferred process of using reinforcement learning algorithms to guide the architecture search includes:

[0075] Use state similarity functions to evaluate the correlation between different agent architectures and guide network mutations during the search process;

[0076] The state similarity function is defined as follows:

[0077]

[0078] in, This indicates the state similarity between two proxy architectures. It is a state mapping function.

[0079] Preferably, it also includes a reward function R, used to evaluate the exploration efficiency of the agent architecture; the optimal neural network architecture for:

[0080]

[0081] The proxy parameters of the optimal neural network architecture are:

[0082]

[0083] Where R is the reward metric function and θ is the proxy parameter of g.

[0084] The preferred method for utilizing AI-Agent-coordinated knowledge transfer technology to transfer the complex knowledge of the guiding model to a simplified model to form an adaptive agent model includes:

[0085] By using proxy pre-migration, the optimal transfer coefficient is obtained by verifying dynamic indicators.

[0086] Based on the optimal transfer coefficient, the complex knowledge of the guiding model is transferred to the simplified model to form an adaptive surrogate model through formal transfer.

[0087] The process of obtaining the optimal transfer coefficient includes:

[0088] Based on the Wilcoxon rank-sum test, the rank-sum statistic is obtained through the formula for calculating the rank-sum statistic.

[0089] We introduce the probability density function of the rank-sum distribution to understand the distribution of the rank-sum statistic given a sample size;

[0090] The probability q of the occurrence of the rank-sum statistic is determined using the cumulative distribution function of the rank-sum distribution.

[0091] The optimal transition coefficient is obtained based on the probability q-value and the rank-sum statistic.

[0092] The preferred method also includes: setting a proxy evaluation function to analyze the contribution of different coefficients to the improvement of the final model's adaptability, and quantifying the relative effectiveness of the coefficients through the rank-sum statistic.

[0093] Tag refinement and parameter fine-tuning strategies include:

[0094] Label refinement and parameter fine-tuning strategies constitute the key components of the AI-MLP Hybrid Network (EMHN) architecture. This strategy, through simulated benchmark generation, perturbation introduction, and Adam optimizer-driven closed-loop iteration, forms a data optimization mechanism under physical constraints, gradually refining the initial labels and approximating the physical solution curve of the radiative transfer equation (RTE). Integrated band selection ensures the EMHN's information utilization efficiency and inversion robustness on thermal infrared data, making it suitable for scenarios with high atmospheric water vapor variability and topographic heterogeneity.

[0095] VI. Beneficial Effects

[0096] This invention significantly improves inversion accuracy: in simulation experiments, five-band input reduced LST MAE from 0.76 K to 0.51 K and increased PCC from 0.981 to 0.999; in actual verification, MAE decreased from 1.85 K to 1.03 K and PCC increased from 0.976 to 0.994. Compared with traditional methods, it has enhanced generalization ability, supports complex terrain applications, and provides highly reliable data. This invention reduces computational costs and improves efficiency through the fusion of large and small models by AI-Agent, making it suitable for real-time remote sensing tasks. Attached Figure Description

[0097] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings, key elements and steps are labeled with reference numerals to clearly correspond to the descriptions in the specification, ensuring a visual presentation of the technical solution. Each drawing is described in detail below:

[0098] Figure 1 This is a flowchart of the method in an embodiment of the present invention, showing the complete process from data collection, AI-Agent coordinated large-small model training to LST and LSE inversion, highlighting the logical structure of the nested framework.

[0099] Figure 2The following are the simulation data inversion results of the embodiment of the present invention. (a), (b), and (c) are scatter plots of LST simulation data inversion results for different band combinations, showing that the baseline model MAE=0.76 K and the refined model MAE=0.51 K; the x-axis is the observed value (250-350 K) and the y-axis is the inverted value (250-350 K); (d) the difference between the surface emissivity inversion values ​​of bands 31 and 32 and the simulation data, with a value range of 0-0.1.

[0100] Figure 3 The above is a map showing the local temperature distribution of MODIS products during the day and night, with values ​​ranging from 250 to 350 K. (a) and (c) are maps showing the LST distribution of MODIS products during the day and night, with values ​​ranging from 250 to 350 K. (b) and (d) are maps showing the LST distribution of MODIS products during the day and night, with values ​​ranging from 250 to 350 K, which are spatial distribution comparisons.

[0101] Figure 4 The following are the error spatial distribution diagrams and histograms of LST inversion values ​​and MODIS products in embodiments of the present invention, where (a)(c) daytime, MAE=1.85 K; (b)(d) nighttime, MAE=1.03 K; and the error range is -5 K to 5 K.

[0102] Figure 5 The following is an inversion distribution map of the emissivity of band 31 in an embodiment of the present invention, wherein (a)(c) are MODIS LSE31 product images (day and night), with values ​​ranging from 0.8 to 1.0; (b)(d) are inverted LSE31 (day and night), with values ​​ranging from 0.8 to 1.0.

[0103] Figure 6 The following is a distribution map of the inversion of the emissivity of band 32 in an embodiment of the present invention, wherein (a)(c) are MODIS LSE32 product images (day and night), with values ​​ranging from 0.8 to 1.0; (b)(d) are inverted LSE32 (day and night), with values ​​ranging from 0.8 to 1.0.

[0104] Figure 7 The following are the error spatial distribution diagrams and histograms of the inverted values ​​of band 31 emissivity (LSE31) and MODIS products in an embodiment of the present invention: (a)(c) daytime, MAE=0.05; (b)(d) nighttime, MAE=0.04; error range -0.1 to 0.1.

[0105] Figure 8 The following is a spatial distribution diagram and histogram of the error between the inverted value of band 32 emissivity (LSE32) and the MODIS product in an embodiment of the present invention, where (a)(c) daytime, MAE=0.05; (b)(d) nighttime, MAE=0.04; and the error range is -0.1 to 0.1.

[0106] Figure 9 The LST inversion ground observation verification results of this invention are shown in (a) daytime, PCC=0.994; (b) nighttime, PCC=0.994; the x-axis is the observed value (250-350 K), and the y-axis is the inverted value (250-350 K). Detailed Implementation

[0107] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0108] This invention provides a method, system, and device for retrieving land surface temperature (LST) and emissivity (LSE) using a small-scale model based on AI-Agent. It is applicable to multi-band thermal infrared data and achieves high-precision simultaneous retrieval of LST and LSE. This implementation uses MODIS data from the United States as an example, covering major land surface types (such as bare land, water bodies, grassland, farmland, and wetlands), with a dataset of approximately 1.64 million samples. The hardware environment includes an Intel Xeon Platinum 8474C CPU, 80GB RAM, and an RTX 4090D GPU (24GB RAM); the software environment is based on Ubuntu 22.04 LTS operating system, Python 3.10.13, and the PyTorch 2.17.0 deep learning framework. The specific steps, model architecture, optimization strategies, and verification results are described in detail below to ensure that those skilled in the art can implement this invention based on this description.

[0109] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0110] Example 1

[0111] This embodiment provides a method for inverting land surface temperature and emissivity based on a large-small model using AI-Agent nested coordination, including:

[0112] S1: Obtain the inversion dataset of geophysical parameters, including simulated MODIS data and field validation data, such as thermal infrared band data (bands 27, 28, 29, 31, 32) acquired by the MODIS satellite; data sources include AMSR2 (2020 global data) and ISMN stations; preprocessing includes removing radio frequency interference (threshold >5%) and permafrost observations (temperature).

[0113] S2: Based on the inversion dataset, AutoKeras is used to perform automatic architecture search and train a large model as a teacher model; the large model adopts a multi-layer neural network structure and optimizes the number of hidden layers and nodes to handle complex terrain data;

[0114] S3: Introducing the AI-Agent module to dynamically adjust weights during knowledge distillation: The AI-Agent uses reinforcement learning to evaluate environmental conditions (atmospheric conditions, data noise) and optimizes the custom loss function. .

[0115]

[0116] in To preserve data integrity, For model consistency loss, The atmospheric correction loss is used; the weights α, β, and γ are dynamically adjusted by the AI-Agent through Q-learning to adapt to the real-time remote sensing data stream. Through a pre-distillation step, unsupervised learning is used to extract features from the hyperspectral data to obtain the optimal guidance weights, and formal distillation transfers the knowledge of the teacher model to the small model to form the student model.

[0117] S4: Jointly optimize the student model and the teacher model, including multi-sensor fusion (such as MODIS data), to achieve LST and LSE inversion; the joint optimization model adapts to different band combinations in real time through AI-Agent to improve inversion accuracy.

[0118] To verify the feasibility and accuracy of this method, this embodiment conducted theoretical accuracy verification, regional inversion verification, and ground observation data verification.

[0119] Theoretical accuracy verification and analysis

[0120] Given the varying sensitivities of different band combinations to surface parameters, this embodiment uses the same simulated dataset to discuss the inversion accuracy of surface temperature and surface emissivity under different band combinations, verifying its theoretical accuracy. Firstly, for surface temperature inversion, three band combinations were set up. Tables 1 to 3 summarize the inversion results of surface temperature, with the optimal results marked in bold. Table 1 shows that when using a three-band combination, a 7-layer hidden layer structure with 600 nodes per layer exhibits optimal performance, achieving MAE and RMSE of 1.15 K and 1.55 K, respectively. When using a four-band combination (Table 2), with an 8-layer hidden layer structure and 700 nodes per layer, the model performance is further improved, with MAE and RMSE decreasing to 0.51 K and 0.69 K, respectively, demonstrating even better inversion accuracy. This result confirms the important role of increasing the number of effective bands and optimizing network depth in improving temperature inversion accuracy. This improvement in accuracy is mainly attributed to the introduction of band 28 as a strong water vapor absorption channel, whose atmospheric water vapor information effectively enhances the model's ability to correct for atmospheric attenuation effects. Table 3 shows that using the five-band combination 27-28-29-31-32 to retrieve land surface temperature, the highest accuracy is achieved when the number of hidden layers is 6 and each layer has 800 nodes. The MAE is 0.53 K and the RMSE is 0.75 K. Experimental results indicate that the introduction of water vapor absorption bands effectively improves the accuracy of atmospheric water vapor parameter retrieval. However, as a deep water vapor absorption channel, MODIS band 27 has low surface transmittance, limiting its direct contribution to improving land surface temperature retrieval accuracy. Therefore, it can be seen that water vapor channels with moderate absorption intensity are more beneficial for improving land surface temperature retrieval accuracy than strong absorption channels. Furthermore, comparing the accuracy indicators in the table reveals relatively small fluctuation ranges, which also demonstrates the good stability of the optimal band combination.

[0121] Table 1. LST Error Retrieved from Combination of Bands 29-31-32

[0122]

[0123] Table 2 LST Errors from Combination Inversion of Bands 28-29-31-32

[0124]

[0125] Table 3. LST Errors from Combination Inversion of Bands 27-28-29-31-32

[0126]

[0127] Furthermore, this embodiment performed statistical analysis on the three sets of data results with the highest accuracy in Tables 1 to 3 (e.g. Figure 2 (As shown). In Figure 2In (a), (b), and (c), the outliers in the inversion results using four thermal infrared bands are significantly reduced, and the accuracy is significantly higher, with a correlation coefficient squared (R²) of approximately 0.99. Based on the verification analysis of the above experimental results, at least four thermal infrared bands are required to achieve a surface temperature inversion accuracy within 1 K. This embodiment also calculates the inversion error results for surface emissivity, mainly listing the emissivity errors of bands 31 and 32 when using the combination of bands 28-29-31-32. As can be seen from Tables 4 and 5, when using four thermal infrared bands, the inversion results errors of LSE31 and LSE32 are basically similar, with both MAE and RMSE below 0.01. Observation Figure 2 (d) It can be seen that the inversion difference values ​​of LSE32 are more concentrated near the 0 value than those of LSE31. In this embodiment, the emissivity of bands 31 and 32 is selected as the input parameters. The surface temperature and surface emissivity obtained by inversion are used as prior constraints to compensate for the insufficient information in the inversion of key parameters.

[0128] Table 4. Error of LSE31 retrieved from the combination of bands 28-29-31-32

[0129]

[0130] Table 5. Error of LSE32 retrieved from the combination of bands 28-29-31-32

[0131]

[0132] Regional inversion verification results

[0133] Based on the above physical logic reasoning and comparison of inversion results with different band combinations, it was found that the model constructed using four thermal infrared bands achieved the best inversion performance when retrieving surface temperature, with RMSE values ​​all below 1 K. Therefore, this embodiment uses the AI-Agent-based large-scale model joint inversion method proposed in this paper to retrieve LST for the southern region of North America from remote sensing images of the MODIS11L2 product.

[0134] This embodiment selects two MODIS images to verify the application of the above method: images taken during the day on August 2, 2021, and images taken at night on August 3, 2021. The brightness temperature of the EOS / MODIS product data (bands 28, 29, 31, 32) from thermal infrared remote sensing is selected as the input parameter, and the output is the land surface temperature. For example... Figure 3 As shown, the overall spatial distribution of LST retrieved by the method of this invention is similar to the spatial distribution trend of land surface temperature in MODIS products, where the white areas represent invalid values. The spatial distribution of land surface temperature in the study area gradually increases from northeast to southwest, and this temperature distribution pattern is consistent with the spatial distribution of regional climate zones. According to... Figure 4During the daytime, the MAE and RMSE of the LST retrieval results are 0.69 K and 1.49 K, respectively, with a bias of 0.002. At night, the MAE and RMSE are 0.43 K and 0.72 K, respectively, with a bias of -0.132. The results show that the retrieval error of LST at night is smaller than that during the day, and the accuracy of LST retrieval using the method of this invention is better at night than during the day. The main reason for these differences is that at night, there are fewer interference factors such as solar irradiance, leading to improved retrieval accuracy. Furthermore, regardless of day or night, the difference between the LST retrieved from the official MODIS product and the LST retrieved using the method of this invention is mainly distributed between -1 K and 1 K.

[0135] Figure 5 and Figure 6 The spatial distribution trends of surface emissivity in bands 31 and 32 retrieved in this study are presented in comparison with those of the official MODIS LSE product, showing a consistent spatial distribution pattern overall. According to... Figure 7 and Figure 8 The error spatial distribution map shows that, under daytime observation conditions, the MAE of the inversion results corresponding to LSE31 and LSE32 are 0.004 and 0.003, respectively, and the RMSE is 0.006 for both. At night, the MAE of the inversion results corresponding to LSE31 and LSE32 is 0.004, and the RMSE is 0.005 for both. Through comparison... Figure 7 and Figure 8 The difference distribution histogram reveals that the surface emissivity inversion results maintain high accuracy and consistency between day and night. The differences between the MODIS LSE product and the inverted LSE are mostly within -0.01 to 0.01, demonstrating the robustness of the joint inversion method in day-night inversion. It is worth noting that through comparative analysis... Figure 4 , Figure 7 and Figure 8 The spatial distribution map of the discrepancies reveals that the areas with the largest errors between the LST and LSE inversion results and the official MODIS products are mainly located in the southwestern part of the study area, indicating an underestimation of surface temperature. This bias can be attributed to the insufficient temporal representativeness of the training dataset: because sample collection was mainly concentrated during the high-temperature summer period, the model training lacked sufficient learning of diverse climatic conditions, thus affecting the adaptability of the inversion algorithm under specific land cover types.

[0136] Nevertheless, the joint inversion algorithm proposed in this invention demonstrates high consistency and reliability in the inversion of surface temperature and surface emissivity. This AI-Agent-based large-scale model joint inversion method improves the accuracy of the inversion results, showing a significant advantage. Future research could enhance the model's generalization ability under different environments by increasing the diversity of training data, including data from different seasons, different climatic conditions, and different surface types, and by fusing multi-source remote sensing and ground observation data, thus making it more robust and accurate in practical applications.

[0137] Validation results of ground observation data

[0138] Ground station data analysis is a crucial step in verifying the accuracy and reliability of the joint inversion algorithm. The availability and quality of this data significantly impact the verification process. Southern North America, particularly the United States, has a large number of ground observation stations, providing abundant data resources for this verification. Given the limited number of observation stations related to surface emissivity and the difficulty in data acquisition, while surface temperature observation stations offer higher spatial coverage density and a more comprehensive data acquisition system, this embodiment primarily focuses on verifying the joint inversion LST by comparing the inverted values ​​with measurements from ground observation stations to evaluate its performance and practicality. Furthermore, it is necessary to interpolate the station data to the synchronization time with satellite imaging, and to perform spatial and temporal scale adjustments and data quality control to ensure data compatibility with remote sensing products. Considering potential noise interference in the ground observation data, this embodiment employs a box plot method based on statistical principles as the core strategy for outlier removal. Specifically, the upper quartile (Q3), lower quartile (Q1), and their difference (IQR = Q3 - Q1) are first calculated. Data points exceeding the range [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR] are then identified as outliers. This method is largely effective in identifying and removing outliers that significantly impact model performance evaluation, thus ensuring the reliability of the data analysis results.

[0139] Figure 9(a) and (b) show scatter plots of LST ground observation data validation for daytime and nighttime conditions, respectively. During the day, the MAE and RMSE of LST retrieval are 1.47 K and 1.80 K, respectively, with an R² of 0.88. At night, the MAE and RMSE of surface temperature retrieval are 1.28 K and 1.63 K, respectively, with an R² of 0.90. Based on the analysis of the above evaluation indicators, the retrieval accuracy of surface temperature is higher at night than during the day, a result consistent with the cross-validation results mentioned above. Specifically, solar radiation intensity and the resulting thermal radiation changes are key environmental factors affecting the LST retrieval process. During the day, influenced by solar radiation, surface temperature changes rapidly and unevenly, leading to larger retrieval errors. Furthermore, factors such as atmospheric humidity, aerosols, and surface albedo may also affect the retrieval process. These factors are relatively stable at night, with smaller surface temperature changes and less atmospheric disturbance, reflecting more stable radiative transfer conditions and making the retrieval results closer to reality. This diurnal variation highlights the dynamic characteristics of radiative transfer processes in thermal infrared remote sensing, providing an important theoretical basis for optimizing all-weather land surface temperature retrieval algorithms. This embodiment employs a multi-dimensional validation framework to systematically evaluate the performance of the proposed joint retrieval algorithm: radiative transfer model simulation data verifies the theoretical feasibility, cross-validation is used to evaluate algorithm stability, and ground station observation data is combined to conduct accuracy verification. Although various validation methods have their own characteristics in terms of spatiotemporal scale and data properties, the validation results show good consistency, indicating that the integrated retrieval scheme has a reliable level of accuracy. A cross-comparison with SMAP L-band products shows an improvement of 0.1 in the correlation coefficient.

[0140] Beneficial effects of this embodiment:

[0141] This embodiment achieves dynamic adaptation through AI-Agent, resulting in higher accuracy than existing technologies. Leveraging cutting-edge artificial intelligence, this embodiment addresses the challenge of limited accuracy in small-model inversion due to the complexity and nonlinearity of the Earth system and the high data complexity caused by the diversity of land cover. It employs a novel large-model-small-model joint optimization inversion strategy incorporating knowledge distillation technology, aiming to explore a paradigm for using a large model to overcome the difficulties of small-model inversion. First, considering the data-driven nature of the inversion dataset, the optimal neural network architecture for model adaptation is obtained based on AutoKeras and the model is trained. Then, the trained high-precision, generalizable large model is introduced as the teacher model, while the small model, used as the student model, is used for local areas where accuracy remains limited. Pre-distillation is performed, and the optimal guidance weights are obtained through statistical exponential verification. Then, knowledge distillation technology is formally used to transfer the deep knowledge of the large model to the small model, improving the inversion accuracy of the small model in areas with limited accuracy, thus more closely approximating the representation of actual physical characteristics. This achieves a high-precision parameter inversion model for specific regions.

[0142] In particular, knowledge distillation offers several advantages over traditional model training methods. By adopting knowledge distillation, the teacher model exhibits global generalization capabilities, effectively addressing the challenge of complex student model data, especially under complex terrain conditions. This method allows the smaller model to simulate the behavior patterns of the larger model, eliminating outliers during the inversion process. By reducing background noise, improving feature extraction, and enhancing the model's generalization ability, it improves the performance of the smaller model in specific application scenarios. Finally, the student model is interactively integrated into the teacher model to form a high-quality joint optimization model, further enhancing the model's specificity and generalization ability.

[0143] This invention not only provides strong technical support for the further development of thermal infrared remote sensing technology in the field of surface temperature and emissivity inversion, but also provides important reference and key technical support for the construction of general models for the inversion of other geophysical parameters.

[0144] Example 2

[0145] This embodiment provides a large-small model land surface temperature and emissivity inversion system based on AI-Agent nested coordination, used to implement the AI-large model land surface temperature and emissivity inversion system described in Embodiment 1. The system includes:

[0146] S1, obtain the inversion dataset of geophysical parameters and construct diverse samples using satellite data;

[0147] S2 deploys an AI-Agent system, uses reinforcement learning to dynamically optimize the architecture, and introduces a guiding model;

[0148] S3 utilizes AI-Agent to coordinate knowledge transfer and form an adaptive agent model;

[0149] S4 achieves real-time inversion by fusing models through an interactive framework.

[0150] In this embodiment, the proxy search space is specifically described as follows:

[0151]

[0152] The state similarity function is defined as follows:

[0153]

[0154] The optimal neural network architecture is:

[0155]

[0156] The optimal proxy parameters are:

[0157]

[0158] Example 3

[0159] This embodiment provides an electronic device, which includes a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the AI-Agent-based large-scale model surface temperature and emissivity inversion method described in Embodiment 1.

[0160] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for inverting land surface temperature and emissivity based on a large-small model using AI-Agent nested coordination, characterized in that, Includes the following steps: Brightness temperature data from multi-band thermal infrared sensors were acquired as the input dataset, and initial surface temperature (LST) and emissivity (LSE) products were collected as tag references. Based on the radiative transfer equation (RTE), physical logic reasoning is performed to establish the theoretical basis for the input-output relationship. Simulation data is used to verify the impact of multi-band brightness temperature input on inversion accuracy and optimize the input band combination. Deploy an AI-Agent system that coordinates an expert-multilayer perceptron hybrid network (EMHN). The large model uses an expert hybrid (MoE) framework to handle complex feature extraction, while the small model uses a multilayer perceptron (MLP) for efficient prediction. By fusing the large and small models through a dynamic routing mechanism, the LST and LSE inversion of ASTER brightness temperature data can be achieved. By introducing a label refinement and parameter fine-tuning strategy and utilizing an AI-Agent-driven iterative optimization closed loop, labels are further refined and model parameters are adjusted to achieve high-precision simultaneous inversion.

2. The method for inverting surface temperature and emissivity based on a large-scale model using AI-Agent according to claim 1, characterized in that, The steps for physical logic reasoning based on the radiative transfer equation (RTE) include: Using the blackbody radiation law as a reference, the formula is: Where B(λ,T) represents the blackbody radiance at temperature T at wavelength λ (unit: W / m²·sr·μm), h represents Planck's constant (6.626 × 10^(-34) J·s), c represents the speed of light (3 × 10^8 m / s), k represents Boltzmann's constant (1.381 × 10^(-23) J / K), λ represents wavelength (unit: m), and T represents temperature (unit: K). For gray-body surfaces, the formula for sensor-observed radiance is: Where Li represents the observed radiance of the i-th channel (unit: W / m²·sr·μm), εi represents the surface emissivity of the i-th channel (dimensionless, range [0,1]), B(λi,Ts) represents the blackbody radiance at surface temperature Ts (unit: K), τi represents the atmospheric transmittance of the i-th channel (dimensionless, range [0,1]), L↓,i represents the downward atmospheric radiation of the i-th channel (unit: W / m²·sr·μm), and L↑,i represents the upward atmospheric radiation of the i-th channel (unit: W / m²·sr·μm). Simplifying atmospheric path radiation yields a compact form: in, Indicates the effective atmospheric temperature (unit: K); Band sensitivity is assessed using the Jacobian matrix, using the following formula: in, Represents the elements of the Jacobian matrix (units depend on the parameter). This represents unknown parameters (including Ts, Ta, τi, εi).

3. The method for inverting land surface temperature and emissivity based on a large-scale model using AI-Agent according to claim 1, characterized in that, The steps for deploying the AI-Agent system to coordinate an expert-multilayer perceptron hybrid network (EMHN) include: The brightness temperature data is standardized and preprocessed using the following formula: in, This represents the standardized brightness temperature value of the i-th band (dimensionless). This represents the original brightness temperature value (in K). This represents the mean value of the i-th band (in K). This represents the standard deviation of the i-th band (in K). The expert weights in a gated network are calculated using the following formula: in, Denotes the weight of the k-th expert (dimensionless, ∑ =1), This indicates a gated linear output, and k represents the number of experts. Integrating expert output, the formula is: in, This indicates the final prediction (LST unit K or LSE dimensionless). Indicates the output of the k-th expert; Introducing load balancing losses, the formula is: in, This represents the load balancing loss (dimensionless). This represents a hyperparameter (typically 0.01). The overall loss function is: in, Let N represent the total loss, y represent the number of samples, β represent the true label, and γ represent the balance coefficients (typically 0.1). This represents the physical regularization loss.

4. The method for inverting land surface temperature and emissivity based on a large-scale model using AI-Agent as described in claim 1, characterized in that, The steps of introducing the label refinement and parameter fine-tuning strategy include: The first-order moment estimation formula is obtained by updating parameters using the Adam optimizer: in, Represents a first-order moment vector. This represents the momentum coefficient (typically 0.9). Indicates the current gradient; The formula for estimating the second moment is: in, Represents a second-order moment vector. Indicates the coefficient (typically 0.999); The parameter update formula is: in, Let η represent the parameter vector, and η represent the learning rate (initially 0.001). , ϵ represents the correction moment, and ϵ represents the stability constant (10^(-8)). The formula for adjusting the label is: in, The adjusted value. For the first The original LST value of the next iteration (from the initial product dataset collected). These are the values ​​retrieved from the model. This is a relaxation factor (typically 0.5) used to control the adjustment range. This step directly optimizes the acquired data, ensuring physical consistency with the radiative transfer equation.

5. The method for inverting land surface temperature and emissivity based on a large-scale model using AI-Agent according to claim 1, characterized in that, It also includes a reward function R, used to evaluate the exploration efficiency of the agent architecture; the optimal neural network architecture for: The proxy parameters of the optimal neural network architecture are: Where R is the reward metric function and θ is the proxy parameter of g.

6. A large-small model inversion system for land surface temperature and emissivity based on AI-Agent nested coordination, characterized in that, The system for implementing the method according to any one of claims 1-5 comprises: The data acquisition module is used to acquire multi-band brightness temperature data of thermal infrared and initial tags; The physical reasoning module is used for logical reasoning and input optimization based on RTE.

7. An electronic device, characterized in that, The device includes: Processor and memory storing computer program instructions; When the processor executes the computer program instructions, it implements the AI-Agent-based method for inverting surface temperature and emissivity using a large model as described in any one of claims 1-5.

Images