AI size model surface temperature and emissivity inversion method, system and equipment

Abstract

The present invention discloses an AI large and small model surface temperature and emissivity inversion method, system and equipment, the inversion method includes: obtaining an inversion data set of geophysical parameters; based on the data drive of the inversion data set, using the automatic machine learning tool AutoKeras to obtain the optimal neural network architecture for model adaptation and train the model, based on the optimal neural network architecture, introducing the trained large model as the teacher model; using the knowledge distillation technology, the deep knowledge of the teacher model is passed to the small model to form a student model; the student model is interactively integrated into the teacher model to form a joint optimization model, and the joint optimization model is used to invert the geophysical parameters to be measured. The present invention can effectively reduce the model development cycle and computing resource consumption through automated model architecture selection and hyperparameter adjustment, and provides an efficient and flexible solution for remote sensing parameter inversion.

Description

Technical Field

[0001] The present invention belongs to the technical field of geophysical parameter inversion, and in particular relates to an AI size model surface temperature and emissivity inversion method, system and equipment. Background Art

[0002] In the field of geophysical parameter inversion, land surface temperature (LST) and emissivity (LSE) are two key parameters, which are of great significance for climate change research, environmental monitoring and assessment, etc. Traditional land surface temperature inversion techniques include single window, split window and multi-band algorithms combining day and night observations. These methods can provide reliable results under specific environmental conditions, but they also have some limitations.

[0003] Among them, the single-window algorithm relies on prior knowledge of surface type and atmospheric parameters, and usually requires additional observations or is based on assumptions. The split-window algorithm uses a pre-estimation of the surface emissivity and atmospheric water vapor content to eliminate most of the atmospheric effects, but this relies on accurate estimates of these parameters. The multi-band algorithm assumes that the emissivity remains constant during the day and night, which is suitable for dynamically changing surface environments, but requires dynamic adjustment.

[0004] In order to overcome the limitations of traditional remote sensing inversion algorithms, more and more studies have begun to explore the use of artificial intelligence technologies such as deep learning to accurately invert remote sensing parameters such as surface temperature and emissivity. For example, deep learning frameworks such as convolutional neural networks (CNNs) are used to derive surface temperature from remote sensing data, demonstrating the potential to overcome data synchronization challenges. Although these artificial intelligence algorithms have achieved significant technological breakthroughs based on traditional methods, they often fail to fully consider the diversity of surface cover, seasonal changes, and the particularity of regional characteristics, which still restricts their generalization ability and application accuracy.

[0005] In summary, although traditional algorithms and deep learning methods have made significant progress in the field of surface temperature and emissivity inversion, they are still limited by the lack of in-depth understanding of the inherent complexity of the Earth system and the multidimensional nonlinear characteristics of atmospheric interactions, especially when it comes to data in complex terrain areas. Although the data has been standardized and preprocessed, its accuracy is still insufficient. These algorithm models are essentially simplifications and abstractions of the real world, and their outputs can often only approximate rather than fully capture the real characteristics. Despite the use of advanced technology, the model is difficult to fully reveal the subtle changes in the real world, especially when the data distribution is complex. This limitation has become a major obstacle to improving the inversion accuracy. Finally, for data-driven remote sensing inversion, the traditional exhaustive method relies on a computationally intensive iterative process, which limits the scalability and complexity of the model structure. Summary of the invention

[0006] To solve the above technical problems, the present invention proposes an AI large and small model surface temperature and emissivity inversion method, system and equipment, which integrates physical methods and statistical methods as the theoretical basis for remote sensing parameter inversion, and combines the latest artificial intelligence technologies such as automatic machine learning, large models and knowledge distillation to enhance the theoretical and practical application performance of the model to solve the problems existing in the above-mentioned prior art.

[0007] To achieve the above object, the present invention provides an AI size model surface temperature and emissivity inversion method, comprising:

[0008] Obtain inversion datasets of geophysical parameters;

[0009] Based on the data-driven nature of the inversion dataset, the automatic machine learning tool AutoKeras is used to obtain the optimal neural network architecture for model adaptation and train the model. Based on the optimal neural network architecture, the trained large model is introduced as a teacher model.

[0010] Using knowledge distillation technology, the deep knowledge of the teacher model is transferred to the small model to form a student model;

[0011] The student model is interactively integrated into the teacher model to form a joint optimization model, and the joint optimization model is used to invert the geophysical parameters to be measured.

[0012] Preferably, the process of using the automatic machine learning tool AutoKeras to obtain the optimal neural network architecture for model adaptation includes:

[0013] The inversion data set of geophysical parameters is obtained through Aster satellite acquisition;

[0014] Preprocessing the inversion data set to obtain preprocessed data;

[0015] Input the preprocessed data into the AutoKeras system and define the search space of the neural structure;

[0016] Initialize neural architecture search using AutoKeras automatic machine learning technology;

[0017] A Bayesian optimization algorithm is used to guide the neural architecture search, and a Gaussian process is used to control network deformation to explore a predefined search space;

[0018] Based on the search space, the optimal neural network architecture is automatically searched within a predetermined number of iterations.

[0019] Preferably, the search space includes: the number of layers of the neural network, the number of units in each layer, the activation function and the dropout probability;

[0020] The search space is specifically described as follows:

[0021]

[0022] Where n is the number of neural network layers, u i is the number of units in the i-th layer, φ i is the activation function of the i-th layer, p i is the Dropout probability of the i-th layer.

[0023] Preferably, the process of using the Bayesian optimization algorithm to guide the neural structure search includes:

[0024] Use the edit distance kernel function to evaluate the similarity between different neural network architectures and guide network deformation during the search process;

[0025] The neural network kernel function of the edit distance is defined as follows:

[0026]

[0027] Among them, d(f a ,f b ) represents the edit distance between two neural networks, and ρ is the mapping function.

[0028] Preferably, a cost function cost is also included to evaluate the search efficiency of the neural network architecture;

[0029] The optimal neural network architecture is:

[0030]

[0031] The learning parameters of the optimal neural network architecture are:

[0032]

[0033] Among them, Cost is the evaluation metric function and θ is the learning parameter of f.

[0034] Preferably, the process of transferring the deep knowledge of the teacher model to the small model to form the student model by using the knowledge distillation technology includes:

[0035] Through pre-distillation, the best guidance weights are obtained by relying on statistical index verification;

[0036] Based on the optimal guidance weights, the deep knowledge of the teacher model is transferred to the small model to form a student model through formal distillation.

[0037] Preferably, the process of obtaining the optimal guidance weight includes:

[0038] Based on the independent sample t-test method, the t statistic is obtained through the t statistic calculation formula;

[0039] The probability density function of t distribution is introduced to understand the distribution of the t statistic under given degrees of freedom.

[0040] Use the cumulative distribution function of the t distribution to determine the probability p value of the t statistic;

[0041] According to the probability p-value and t-statistic, the best guidance weight is obtained.

[0042] Preferably, it also includes: setting an evaluation function for analyzing the contribution of different weights to the improvement of the final model accuracy, and quantifying the relative utility of the weights through t statistics.

[0043] The present invention also provides an AI size model surface temperature and emissivity inversion system, which is used to implement the AI ​​size model surface temperature and emissivity inversion method, and the system comprises:

[0044] A data acquisition module, used to obtain an inversion data set of geophysical parameters;

[0045] A large model building module is used to obtain the optimal neural network architecture for model adaptation and train the model based on the data drive of the inversion data set using the automatic machine learning tool AutoKeras, and introduce the trained large model as a teacher model based on the optimal neural network architecture;

[0046] A knowledge transfer module, used to transfer the deep knowledge of the teacher model to the small model to form a student model by using knowledge distillation technology;

[0047] The joint optimization module is used to interactively integrate the student model into the teacher model to form a joint optimization model, and use the joint optimization model to invert the geophysical parameters to be measured.

[0048] The present invention also provides an electronic device, comprising: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the AI ​​size model surface temperature and emissivity inversion method is implemented.

[0049] Compared with the prior art, the present invention has the following advantages and technical effects:

[0050] The present invention provides an AI large and small model surface temperature and emissivity inversion method, comprising: firstly obtaining an inversion data set of geophysical parameters; based on the data driveability of the inversion data set, using the automatic machine learning tool AutoKeras to obtain the optimal neural network architecture for model adaptation and train the model, and based on the optimal neural network architecture, introducing the trained large model as a teacher model; secondly using the knowledge distillation technology, transferring the deep knowledge of the teacher model to the small model to form a student model; finally, interactively integrating the student model into the teacher model to form a joint optimization model, and using the joint optimization model to invert the geophysical parameters to be measured.

[0051] The present invention selects the AutoKeras automated neural network architecture search tool to reduce manual intervention and optimize the model structure. Through its automated model architecture selection and hyperparameter adjustment, the model development cycle and computing resource consumption can be effectively shortened, providing an efficient and flexible solution for remote sensing parameter inversion.

[0052] By adopting the strategy of knowledge distillation, the teacher model of the present invention demonstrates the ability to generalize globally, which can effectively solve the challenge of student model data complexity, especially under complex terrain conditions. This method enables the small model to simulate the behavior pattern of the large model, exclude abnormal data during the inversion process, and improve the performance of the small model in specific application scenarios by reducing background noise, improving feature extraction and enhancing the generalization ability of the model. Finally, the student model is interactively integrated into the teacher model to form a high-quality joint optimization model, which further enhances the model's targeting and generalization capabilities.

[0053] The present 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. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0055] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0056] Figure 2 It is a scatter plot and accuracy display diagram of the LST&LSE inversion of the teacher model in an embodiment of the present invention;

[0057] Figure 3It is a scatter plot and precision display diagram of LST&LSE inversion of a small model 1 of an embodiment of the present invention (the horizontal axis is the real data, and the vertical axis is the model inversion data);

[0058] Figure 4 It is a scatter plot and accuracy display diagram of the 2LST&LSE inversion of the small model of the embodiment of the present invention (the horizontal axis is the real data, and the vertical axis is the model inversion data);

[0059] Figure 5 A schematic diagram of the relative contribution of knowledge distillation to improving model accuracy under different weights in an embodiment of the present invention;

[0060] Figure 6 Optimal weight box plots of different inversion results of the embodiments of the present invention;

[0061] Figure 7 It is the scatter plot and accuracy display diagram of LST&LSE inversion of the student model 1 of the embodiment of the present invention (the horizontal axis is the real data, and the vertical axis is the model inversion data);

[0062] Figure 8 It is the scatter plot and accuracy display diagram of the LST&LSE inversion of the student model 2 of the embodiment of the present invention (the horizontal axis is the real data, and the vertical axis is the model inversion data);

[0063] Fig. 9 A comparison diagram of the inversion accuracy of the student model 1 and the student model 2 according to an embodiment of the present invention;

[0064] Fig.10 The surface temperature map, cross-validation scatter plot and prediction error distribution diagram of the embodiment of the present invention (the abscissa of the scatter plot is MODIS data, and the ordinate is model inversion data);

[0065] Fig.11 This is a site verification scatter plot of an embodiment of the present invention (wherein the horizontal axis represents site data, the vertical axis represents model inversion data, and the unit is K). DETAILED DESCRIPTION

[0066] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0067] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0068] Embodiment 1

[0069] In this embodiment, physical methods and statistical methods are integrated as the theoretical basis for remote sensing parameter inversion, and the latest artificial intelligence technologies such as automatic machine learning, large models and knowledge distillation are combined to enhance the theoretical and practical application performance of the model. For details of the technical flow chart, see Figure 1 .

[0070] In terms of theoretical basis ( Figure 1 In A), the physical method provides the necessary input band parameters for the inversion algorithm, and the statistical derivation based on the physical method provides theoretical support for large model training to process high-dimensional multi-sample data. Given that remote sensing parameter inversion is usually data-driven, this embodiment uses the automatic machine learning tool AutoKeras to optimize the model structure and effectively select the most suitable neural network architecture for a specific data set ( Figure 1 On this basis, the preprocessed data is used to train the model with the optimal neural network architecture. In view of the limited accuracy of the small model, this strategy adopts knowledge distillation technology and uses the high-precision large model as the teacher model to optimize the small model, thereby improving its performance ( Figure 1 The process of knowledge transfer is detailed in Figure 1 Finally, this example fully verifies the effect of the AK-KD strategy ( Figure 1 E), including theoretical verification, cross-validation of MODIS data and field verification of station data, to ensure the effectiveness and reliability of the proposed strategy.

[0071] This embodiment provides an AI size model surface temperature and emissivity inversion method, including:

[0072] S1, obtain the inversion data set of geophysical parameters;

[0073] Specifically, the large model data set is based on the ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) data set, and uses its high resolution and multi-spectral capabilities to finely characterize surface features. High-quality ASTER data covering the main surface types in China in summer are selected to ensure the comprehensiveness and representativeness of the data set. Using a diversified strategy, this embodiment collects a total of about 9.6 million samples, spanning 27 scenes of imagery, providing a solid foundation for building a large-scale model with high generalization capabilities, and ensuring the balanced performance of the model on various types of data. To ensure data quality, all samples undergo strict quality control and preprocessing. Combined with MODIS data, ground observation station data and assimilated data, ASTER data consistent with high-credibility data are selected to ensure data accuracy. Furthermore, this embodiment selects data sets with high inversion accuracy and adds them to the large model training library through pre-training evaluation of pre-processed data for each scene, ensuring the generalization ability of the model while ensuring the accuracy of the model.

[0074] For specific regional data, in the large model data screening stage of this embodiment, two data sets with low inversion accuracy due to complex data were specially selected. It can be seen from the visible light images of the data that these regions have diverse surface types and large terrain differences, so they are key cases for optimizing small models. These regional data are limited by the inherent complexity of the earth system and its nonlinear characteristics, which directly affect the inversion accuracy of the small model. These specific scenarios provide a suitable environment for using knowledge distillation technology to analyze and improve the performance of small models. Through this process, it aims to explore the adaptability and optimization methods of small models when faced with complex data and limited accuracy conditions, and lay the foundation for subsequent model evaluation and application research.

[0075] In this example, MODIS land surface temperature data is used as a key auxiliary data set for cross-validation to ensure the reliability and accuracy of the model inversion results. In particular, MODIS land surface temperature data corresponding to the time and space points corresponding to the two small models are carefully selected as the benchmark for cross-validation. This selection is intended to evaluate and optimize the performance of the model to ensure its accuracy and robustness at different time and space scales.

[0076] This embodiment uses ground station data to verify the accuracy of the model trained based on the ASTER dataset. In the two target study areas, areas with flat terrain and single surface type were selected to collect a series of high-quality ground observation data. Given that surface temperature data is more widely available than emissivity, this embodiment focuses on verifying the effectiveness of the proposed method by comparing the surface temperature data inverted by the optimized student model with ground measurements. The spatiotemporal alignment of the ground station data with the ASTER satellite observation records ensures the rigor and accuracy of the verification process.

[0077] S2, based on the data-driven nature of the inversion data set, using the automatic machine learning tool AutoKeras to obtain the optimal neural network architecture for model adaptation and train the model, and based on the optimal neural network architecture, introducing the trained large model as a teacher model;

[0078] S21, in the derivation of parameter inversion theory, this embodiment will explain in detail the physical significance of deep learning as the core algorithm of model training based on the combination of physical methods and statistical methods. See the schematic diagram for details. Figure 1A. First of all, the physical method ensures that the model has the necessary theoretical basis for practical applications. The interdependence between geophysical parameters determines that the inversion algorithm needs to meet three conditions. First, the retrieval model should have physical meaning. Secondly, the retrieval equation should be mathematically solvable in theory. Third, the retrieval accuracy is high and meets the application requirements. When performing surface temperature inversion, this process is based on the thermal radiation of the ground and the transmission of thermal radiation from the ground through the atmosphere to the remote sensing satellite. The radiation transfer equation (RTE) plays an indispensable role and provides a theoretical framework for the propagation process of surface thermal radiation in the atmosphere, which is of great significance for the accurate analysis and simulation of this complex physical process. The radiation transfer equation is as follows:

[0079] B λ (T λ )=(1-τ λ (θ))(1+(1-ε λ )τ λ (θ))B λ (T a )+B λ (T s ) (1)

[0080] The expression of Planck function is as follows:

[0081]

[0082] Among them, B λ (T λ ) is the on-satellite radiation, τ λ (θ) is the atmospheric transmittance, T s is the surface temperature, T a is the near-surface air temperature, ε λ is the surface emissivity. Among these five parameters, τ λ (θ), τ λ (θ), T s , T a are all unknowns. Theoretically, without any prior knowledge, LST can be retrieved from the four RTEs established from the four TIR bands. When solving mathematical equations, these relationships cannot be described by strict mathematical equations, so the errors will be amplified, and the simplification of the Planck function will also lead to some errors. Therefore, in order to construct a closed set of equations, the number of equations must be greater than or equal to the number of unknowns. In order to significantly improve the accuracy of prediction and inversion, deep learning technology is used. By applying remote sensing technology, atmospheric interference can be effectively separated and corrected from remote sensing data. Taking Aster BT10~14 as an example to construct a closed set of equations to accurately estimate the surface temperature.

[0083] Next, the application of statistical methods is based on physical models for parameter estimation to enhance the practicality and predictive accuracy of the model. Bubeck and Sellke mathematically proved that increasing the number of parameters in the model can improve robustness and thus improve generalization. In the field of data fitting, introducing more data and expanding the dimensions of equations are key strategies to improve model fitting accuracy. In terms of data dimensions, introducing more feature variables can enrich the basic information of the model. Specifically, the least squares method is used to seek to minimize the overall sum of squared errors. (where y i Predicted value, is the model prediction value), thereby optimizing the model parameters and improving the accuracy of the fit. Regression analysis further expands this framework by solving To estimate the model parameters (where X is the design matrix and y is the response vector), the observation sample dimension is expanded, so that the model parameters can be accurately adjusted within the framework of minimizing the sum of squared errors. In terms of sample size, relying on the law of large numbers, that is, when the sample size n tends to infinity, the sample mean Converges to the overall expectation μ with probability, that is For any ∈>0, the expansion of the sample size makes the sample mean of the model gradually approach the overall mean, further improving the generalization ability of the model. According to the central limit theorem, as the sample size increases, the distribution of the sample mean will tend to the normal distribution. This approximation of the normal distribution provides a basis for the use of statistical methods (such as hypothesis testing and confidence intervals), thereby enhancing the reliability of the model's evaluation and prediction on unknown data, which will effectively improve the model's generalization ability. Through this fusion analysis of physical and statistical methods, a proper balance is maintained between model complexity and data volume, providing a solid theoretical foundation for model training in the AK-KD algorithm.

[0084] S22, in terms of automatic machine learning strategies, AutoKeras is an advanced open source software library developed by the DATA Laboratory of Texas A&M University. It uses automatic machine learning (AutoML) technology to automatically find the best neural network architecture in a predetermined search space. Different from other AutoML systems, AutoKeras focuses on automatically identifying the most suitable model structure and hyperparameters through neural architecture search (NAS) to optimize the performance of deep learning models. Its core is to use Bayesian optimization to guide neural structure search, combined with Gaussian process to control network deformation (network morphism), so as to efficiently explore the preset search space.

[0085] The neural architecture search problem can be formulated as follows: Given a neural architecture search space Divide the input data D into training data D t and verification data Dv The cost function cost is used to evaluate the efficiency of the neural network architecture. The goal is to find an optimal neural network To achieve the best performance on the data set D. Therefore, find the optimal neural network f and its learning parameters θ that meet the following conditions:

[0086]

[0087] Where Cost is the evaluation metric function (such as correlation coefficient, absolute error), and θ is the learning parameter of f. Search space By gradually adjusting the initial architecture to cover all neural network architectures, the search space is specifically expressed as:

[0088]

[0089] Where n is the number of neural network layers, u i is the number of units in the i-th layer, φ i is the activation function of the i-th layer, p i is the dropout probability of the i-th layer, and its value is between [0,1]. Under the guidance of the Bayesian optimization (BO) algorithm, the system searches in the search space Search for the best architecture.

[0090] In the AK-KD strategy, the data collected from the Aster satellite is first preprocessed and then input into the AutoKeras system. The system will determine the optimal neural network structure for the data and then use the structure to formally train the model. Figure 1 Part B. Specifically, a basic search space FCN is defined within AutoKeras, and its parameters include: the number of neural network layers n∈N∩[1,9], the number of neurons in each layer u i ∈N∩[32,1024] and the search step is 16, the activation function φ i ∈{ReLU,Tanh,Sigmoid}, and the ratio of the Dropout layer p i ∈[0,1]. After the data is input into the system, the model is pre-trained with a lower epochs in the initial stage to verify the effectiveness of the neural network architecture for each search. During the search process, the network deformation is controlled by Gaussian process and Bayesian optimizer, and the system efficiently explores the predefined search space. The neural network kernel function of the edit distance is defined as follows:

[0091]

[0092] Among them, d(f a ,f b) represents the edit distance between two neural networks, and ρ is the mapping function. The kernel function guides the network deformation during the search process by evaluating the similarity between different neural network architectures. After a limited number of searches, the system automatically terminates the search and outputs the optimal neural network architecture. Finally, the data is formally trained using the optimal architecture to obtain the best performance model.

[0093] S3, using knowledge distillation technology, transfers the deep knowledge of the teacher model to the small model to form a student model;

[0094] Specifically, in the theoretical derivation of the knowledge distillation network, the core differences between traditional model training methods and knowledge distillation technology and the advantages of the latter are first compared. As an efficient knowledge transfer mechanism, knowledge distillation allows key information to be extracted from large and complex models, thereby building more compact and efficient models. In the field of deep learning, the size, complexity and performance of a model are usually key indicators for evaluating its practicality. By optimizing these indicators, knowledge distillation technology effectively balances model performance and complexity and expands the scope of application of the model.

[0095] In practical applications, a response-based distillation strategy is adopted, focusing on the neural response of the last layer of the teacher model, aiming to directly learn the final prediction of the teacher model. Specifically, a large model with strong generalization ability and high accuracy is used as the teacher model to guide the small model to learn and imitate the prediction behavior of the teacher model. In this way, the student model uses the prediction of the teacher model as a soft label to produce similar predictions based on the same input, thereby imitating the behavior of the teacher model. The distillation loss of response-based knowledge transfer can be expressed as:

[0096]

[0097] in, represents the evaluation loss. s is the output of the student model. t is the output of the teacher model. T is the control z t factor of importance.

[0098] In this embodiment, the mean square error (MSE) is used as the specific implementation of the loss function L. The MSE is chosen mainly because of its good performance in convex optimization and nonlinear regression models. The mean square error (MSE) is defined as the difference between the observed value y and the model prediction value The mean squared difference between the two, where N is the number of samples. This metric provides an effective way to quantify the prediction accuracy of the model. The precise statement of mean squared error (MSE) is as follows:

[0099]

[0100] In order to construct a comprehensive and integrated optimization goal, the Global Optimization Criterion (GOC) is introduced to calculate the knowledge distillation training. The formula is as follows. This criterion is a fusion of two key metrics: data fidelity, ) and model consistency measures (Model Consistency, ) is the weighted sum of .

[0101]

[0102] The weight parameter α is an adjustable parameter between 0 and 1, which is used to balance the relative contribution of these two metrics in the overall optimization objective. Data approximation metrics focus on minimizing the difference between the true output y and the student model prediction The model consistency metric mainly focuses on minimizing the teacher model With student models The difference in output between . In this loss function, the first term Quantifies the difference between the student model output and the true label. The second term The difference between the output of the student model and the teacher model is quantified. In this way, the student model will try to imitate the teacher model as much as possible during training, while also trying to get as close to the true label as possible.

[0103] At the same time, knowledge distillation is also widely used to achieve label smoothing, evaluate the accuracy of teacher models, and determine the optimal output layer structure. Including improving the entropy of labels in the data argumentation process, it can effectively enhance the predictive ability of the model. Therefore, under the framework of knowledge distillation, the concept of applying knowledge transfer to training data compression, namely dataset distillation, is further explored. By converting the key knowledge of large datasets into smaller and more efficient student datasets with certain distillation weights, it aims to reduce the burden of deep model training. In the knowledge distillation process of small models, especially the application of surface temperature datasets, the surface temperature data under ideal clear sky conditions is corrected by dataset distillation through the adjustment of knowledge distillation weight factors, which makes up for the deviation between actual values ​​and predicted values, and provides accurate measured values ​​for surface temperature inversion under conditions of complex data and limited accuracy.

[0104] Specifically, in the process of knowledge distillation, we first determine the most appropriate knowledge distillation weight in the global optimization strategy. The schematic diagram is as follows: Figure 1As shown in C. In the preliminary stage, the pre-distillation method is adopted to guide the student model with the help of the teacher model. In this process, the direct distillation of the data set is omitted, PCCS and MAE are used as key evaluation indicators, and the weight range of 0.1 to 0.9 is explored. A series of pre-distillation experiments are carried out for the surface temperature and surface emissivity (LST, LSE10~14) of the student model, and the optimal knowledge distillation weights are determined through statistical test analysis. Subsequently, the selected weights are applied to the formal distillation process. The teacher model will gradually optimize the student model in each round of iteration based on the knowledge distillation loss generated by the global optimization strategy based on this weight. Next, the data set of the student model is distilled through the determined knowledge distillation weights to achieve high-quality data set optimization reconstruction. This method combines the weights of the soft targets of the teacher model and the hard targets of the student model into a comprehensive label. These labels not only retain the original category information, but also incorporate the teacher model's understanding of the underlying structure of the data, providing a richer training target for the student model. Using these comprehensive labels, the training of the student model is no longer limited to pursuing consistency with hard targets, but is approaching more detailed and comprehensive goals, including identifying and simulating the subtle relationships between categories revealed by the teacher model. This knowledge distillation comprehensive labeling mechanism injects a more refined and comprehensive supervision signal into the training of the student model, promoting the improvement of the model's generalization ability.

[0105] This strategy not only enables the student model to inherit the decision-making ability of the teacher model, but also significantly improves the performance while maintaining low complexity and computational cost, and even surpasses the teacher model. This process effectively compensates for the deviation between the model prediction value and the actual value, ensuring the high accuracy of the inversion results, demonstrating the significant advantages of knowledge distillation technology in improving data inversion accuracy, and further ensuring that the inversion results of the small model surface temperature and emissivity are closely aligned with the representation of physical properties.

[0106] S3.1, Knowledge distillation weight selection,In the field of knowledge distillation, weight optimization is one of the core issues, which determines the strength of knowledge transfer between the teacher model and the student model, and directly affects the ability of the student model to absorb the knowledge of the teacher model and maintain its own learning characteristics. The schematic diagram is as follows Figure 1 As shown in D. In theory, the search for an ideal alpha value aims to achieve the best balance between teacher guidance and student autonomous learning, without sacrificing the adaptability of the student model, while maximizing the guidance benefits obtained from the teacher model. This process emphasizes the dynamic balance between the self-exploration ability of the student model at a lower alpha value and the guiding role of the teacher model at a higher alpha value. The key is to find an optimal alpha value to promote the effective integration of teacher guidance and student autonomous exploration, thereby improving the performance of the student model on specific tasks.

[0107] In theory, after the initial training of the model is completed, incremental learning of the model using the original data can continuously obtain useful knowledge and further improve the accuracy. The selection of knowledge distillation weights takes advantage of this feature. In the knowledge distillation of the student model, only the original labels are used as a reference without considering the additional knowledge distillation labels. In this case, when the knowledge distillation weight is zero, that is, the teacher model guidance is not introduced, this is equivalent to incremental learning, and the model accuracy of the student model is improved. When the weight is 1, it means that it is completely dependent on the teacher model, which deviates from the incremental learning of the original data set. Due to the deviation of the training loss from the original label, the accuracy will decrease. To this end, a dynamic balance point is sought, that is, the contribution of the teacher model and the contribution of the student model work together, and then the knowledge distillation label under this weight is used to formally optimize the student model.

[0108] In this framework, the independent sample t-test is used to compare the mean difference between two independent sample groups (i.e., original accuracy and accuracy after weight adjustment). The purpose is to find the weight with the lowest t value, that is, the dynamic equilibrium point of knowledge distillation contribution. The formula for calculating the t statistic is as follows:

[0109]

[0110] in, represents the sample average of the improved accuracy, μ is the accuracy before improvement, is the sample standard deviation of the improved accuracy, and n is the sample size. In order to deeply analyze the contribution of different weights to the final model accuracy improvement, the relative utility of the weights is quantified by the statistic t, and the evaluation function is designed as:

[0111]

[0112] In order to gain a deeper understanding of the distribution of the t statistic under given degrees of freedom, the probability density function (PDF) of the t distribution is introduced, which is expressed as follows:

[0113]

[0114] Here, v represents the degrees of freedom, and Γ is the gamma function, which is used to expand the concept of factorial in this context. Further, the cumulative distribution function (CDF) of the t distribution provides the probability of occurring at a specific t value or more extreme values, and its calculation formula is as follows:

[0115]

[0116] Combining the t statistic and the corresponding degrees of freedom, the p-value is determined by the CDF of the t distribution, that is, the probability of the observed statistic (or more extreme cases) occurring under the premise that the original hypothesis is true. For a two-tailed test, the calculation formula for the p-value is:

[0117] p=2×(1-F(|t|)) (14)

[0118] Taking all factors into consideration, this section establishes the theoretical basis for achieving effective knowledge transfer between the teacher model and the student model through an in-depth analysis of the knowledge distillation weight selection mechanism. By precisely adjusting the weight parameters, this embodiment aims to balance the relationship between teacher guidance and student self-learning, thereby optimizing the learning path of the student model.

[0119] S4, interactively integrating the student model into the teacher model to form a joint optimization model, and utilizing the joint optimization model to invert the geophysical parameters to be measured.

[0120] In this embodiment, the theoretical accuracy analysis and verification

[0121] 1. Automatic machine learning network optimization

[0122] In the field of remote sensing inversion research, especially in the inversion task using deep learning algorithms, the selection of neural network architecture often relies on an exhaustive trial-and-error method, and the configuration of neurons between network layers lacks flexibility. In view of this, this embodiment adopts the AutoKeras automated network search framework, searches 30 times in a preset search space to explore the optimal network architecture, and uses PCCS (Pearson correlation coefficient) and MAE (mean absolute error) as evaluation indicators of model performance.

[0123] This embodiment uses MAE as the evaluation index because it intuitively reflects the average deviation between the predicted value and the true value and is a common tool for measuring prediction accuracy. At the same time, PCCS is used to evaluate the linear correlation between the predicted value and the actual value, which helps to reveal the correlation strength between the model output and the real data, thereby deeply evaluating the explanatory power of the model. By combining these two indicators, not only can the absolute error of the model be evaluated, but also the correlation between the model prediction results and the actual results can be seen, thereby comprehensively evaluating the performance of the model.

[0124] In particular, the optimal neural network architecture was searched for the surface temperature inversion data of the teacher model as an example. In order to fully demonstrate the high efficiency and configuration flexibility of AutoKeras's search, 8 rounds of 30 rounds of neural network architecture search were carried out. The specific results are shown in Table 1.

[0125] Table 1

[0126]

[0127] From the analysis of the results, it is found that the neural network architecture with seven hidden layers (32, 32, 720, 32, 32, 752, 32) has a higher inversion accuracy for the teacher model, so this architecture is selected in the inversion task of this model. Due to the data-driven nature of the inversion model, AutoKeras will be applied to specific data in subsequent inversion tasks to obtain the best neural network architecture for more targeted training in order to achieve the best accuracy.

[0128] 2. Input band selection and verification

[0129] Based on the physical logic deduction analysis, in order to construct a closed set of equations, the number of equations must be greater than or equal to the number of unknowns, so as to improve the fitting accuracy of the model. This embodiment will consider all possible brightness temperature band combinations. Least squares method and regression analysis further confirm this view, pointing out that involving more data categories associated with unknown variables will help minimize the overall error. To verify this theory, the Aster simulation data set with a surface temperature step size of 4 is first used. By flexibly selecting all possible combinations of the five brightness temperature bands BT10 to 14 as input, the AutoKeras neural network is used to search and train the model and conduct experiments on surface temperature inversion. MAE and PCCS are used as evaluation indicators to explore the relationship between input band combination and model accuracy. Comprehensive band combination (BC) analysis not only helps to reduce the overall model error and improve prediction accuracy, but is also a key step in establishing an effective model. The relationship table between input band combination and model accuracy, the experimental results are shown in Table 2, from which it can be clearly seen that different combinations have a specific impact on model accuracy.

[0130] Table 2

[0131]

[0132]

[0133] In-depth analysis of the surface temperature inversion verifies the theoretical expectation that the increase in band combinations can significantly improve model performance. The analysis of PCCS shows that the performance improvement is not simply dependent on the increase in the number of bands. Band 14 performs best in the single-band case, reaching 0.966. Specific two-band and three-band combinations, such as 11-13 (PCCS is 0.981) and 12-13-14 (PCCS is 0.991), surpass some four-band and five-band combinations, emphasizing the key role of information complementarity in the selected bands in model accuracy. In particular, the analysis of MAE shows that multi-band combinations further optimize performance, especially the 10-12-13-14 combination, which reaches 0.996 on PCCS and reduces MAE to 0.774. The excellent performance of the five-band combination 10-11-12-13-14 reduces MAE to the lowest 0.478, highlighting the importance of integrating multi-band information for accuracy improvement, thus substantially supporting the initial theoretical analysis.

[0134] In order to verify the law of large numbers and explore the impact of sample size on model performance, the sample data was increased, the surface temperature step was reduced to 1, and the band combinations with better effects in Table 2 (4BC 10-12-13-14, 5BC 10-11-12-13-14) were selected as experimental objects. In order to further evaluate the impact of sample size on inversion accuracy, these results were compared and analyzed with the inversion accuracy of the simulated data with a step size of 4. The relationship between sample data and model accuracy is shown in Table 3. The detailed results are shown in Table 3:

[0135] Table 3

[0136]

[0137] The analysis results show that the accuracy of the surface temperature inversion model is significantly enhanced by reducing the step size to expand the sample size. When the step size is reduced from 4 to 1 and the sample data increases fourfold, the inversion accuracy PCCS of the four-band combination 10-12-13-14 is improved from 0.996 to 0.997, and the MAE is reduced from 0.774K to 0.609K. The PCCS of the five-band combination 10-11-12-13-14 is increased from 0.968 to 0.999, and the MAE is reduced from 0.478K to 0.348K, which is used as a simulation verification of large model training under the AK-KD strategy. This trend confirms the importance of increasing the sample size to improve the model inversion accuracy and reduce errors under the condition of constant data quality, and further verifies the key role of the law of large numbers in the application of surface temperature inversion. Based on this, the five-band combination 10-11-12-13-14 is selected as the optimal band combination to train the large model and the small model, so as to optimize the LST&LSE inversion accuracy using the AK-KD strategy.

[0138] 3. Dataset model training and analysis

[0139] In the process of training and optimizing the surface temperature and emissivity inversion model using the AK-KD method, the model trained with the large model data set is called the teacher model, and the model trained with the two specific area data sets is called the small model. The small model optimized by the large model through knowledge distillation is called the student model.

[0140] The neural network architecture determined by AutoKeras was used to train the selected teacher model data, aiming to accurately invert the land surface temperature (LST) and the emissivity of the brightness temperature bands (BT10, BT11, BT12, BT13, BT14). Through this process, an independent inversion model for each inversion target was constructed, and the corresponding scatter plots and accuracy displays were generated accordingly (such as Figure 2 ) to visually demonstrate the model performance.

[0141] Specifically, the performance of the obtained inversion model in terms of the Pearson correlation coefficient (PCCS) and mean absolute error (MAE) is as follows: the PCCS of the LST model is 0.967 and the MAE is 0.685K, showing good correlation and reasonable error level; the PCCS of the emissivity model in the LSE10 to LSE14 bands are 0.934, 0.949, 0.951, 0.844 and 0.788, respectively, and the MAE is also maintained at a low level, 0.007, 0.007, 0.007, 0.003 and 0.001, respectively. These results support the theoretical basis of the large model as a teacher model, and its multi-input features and large data volume contribute to the model accuracy. At the same time, it verifies the effectiveness of the selected neural network architecture and reveals the performance differences between the inversion models of each band. This progress has laid a solid foundation for using knowledge distillation technology to improve the accuracy of student models.

[0142] 4. Research area small model training and analysis

[0143] In this example, the neural network architecture selected by the AutoKeras automation framework is used to perform inversion analysis of surface temperature and emissivity on two small model cases. The inversion scatter plots and accuracy of small model 1 and small model 2 are shown as follows: Figure 3 , Figure 4. These small model datasets are derived from datasets where the inversion accuracy is limited due to the complexity of the data, which is limited by the inherent complexity of the Earth system and its nonlinear characteristics. Compared with the teacher model of the high-quality large dataset, small model 1 and small model 2 show a certain performance gap in the key performance indicators PCCS and MAE, especially in specific bands such as LSE13 and LSE14. This comparison reveals the impact of data quality and quantity on inversion accuracy, and also emphasizes the potential of knowledge distillation technology in improving the performance of small models, providing new ideas for the application of models in complex environments.

[0144] 5. Knowledge distillation weight design and analysis (statistical test)

[0145] To guide the formulation of efficient distillation strategies, this example adopts a rigorous statistical method, the independent sample t-test. This method aims to evaluate the statistical significance of model performance improvement under different weight configurations. To this end, for each weight configuration, 30 sets of knowledge distillation experiments without data set distillation were implemented, thereby collecting a series of accuracy data on the model in its original state and after adjusting the weights.

[0146] By carefully applying the above statistical framework, this example conducts an in-depth analysis of 30 sets of experimental data under different weight configurations to evaluate the specific impact of weight adjustment on model performance. The statistical results of teacher relative contribution and student relative contribution based on PCCS and MAE under different weights are shown in Table 4, which summarizes the knowledge distillation statistical analysis results of surface temperature inversion of small model 1, and obtains the t value and p value calculated based on PCCS and MAE under different weights.

[0147] Table 4

[0148]

[0149] By analyzing the t-statistics and p-values ​​of PCCS and MAE under different alphas, we can observe the relative changes in the contributions of the teacher and student models. The t-statistic changes from positive to negative, indicating that as alpha increases, the contribution of the teacher model gradually decreases, while the contribution of the student model gradually increases. At the same time, the p-values ​​under all alphas are significant, confirming the reliability of the statistical results.

[0150] In order to accurately define the optimal weights in the knowledge distillation process, 30 rounds of distillation experiments with differentiated weights were conducted on six different inversion results of LST, LSE10, LSE11, LSE12, LSE13, and LSE14 of small model 1 and small model 2. In order to deeply analyze the contribution of different weights to the improvement of model accuracy, the evaluation function is The purpose is to quantify the relative utility of the weights through the statistic t, and further normalize the data to draw the corresponding radar chart ( Figure 5 ).

[0151] The analysis of the radar chart reveals the relative contribution of different inversion results to the improvement of MAE and PCCS indicators under various weights. Observations show that for the improvement of MAE, the best weights for optimizing small models 1 and 2 are mainly concentrated at 0.2 and 0.3. For the PCCS indicator, the best weights extend to 0.2, 0.3, and 0.4. As the weight deviates from the optimal value, its contribution to the improvement of model performance decreases. Specifically, the greater the deviation of the weight from the optimal value, the weaker its positive impact on performance improvement.

[0152] Finally, the optimal weights of different inversion results are counted and box plots are drawn ( Figure 6 ) for further analysis.

[0153] In the box plot analysis, the median of small model 1 is 0.35, and the interquartile range is between 0.275 and 0.5. The median of small model 2 is 0.3, and the interquartile range is 0.2 to 0.425. The interquartile range represents the clustering area of ​​the middle 50% of the data, providing a basis for evaluating the central tendency and dispersion of the data. The weight of 0.3 is in the center of both models and is also at the lower quartile edge in small model 1.

[0154] In view of this, 0.3 is identified as the optimal weight for improving the small model in the knowledge distillation process, which not only reflects the centralization trend of the data, but also maintains the consistency across models. Such a weight not only ensures that the small model obtains effective guidance from the teacher model, but also maintains its own learning motivation as a student model, achieving the optimal effect of knowledge distillation.

[0155] 6. Student model optimization analysis

[0156] Based on the previous statistical test analysis, 0.3 was selected as the key weight of knowledge distillation so that the teacher model can guide the two small models while maintaining the accuracy improvement. And by reconstructing the inversion target data set, the difference between the true value and the predicted value is made up, and the inversion target value close to the true value is restored. Then, the scatter plots and accuracy display of student model 1 and student model 2 after the small model is improved by knowledge distillation are shown ( Figure 7 , Figure 8 ).

[0157] Compare and analyze the bar charts of the changes in various inversion indicators before and after knowledge distillation, such as Fig. 9 shown.

[0158] Through the performance analysis of the student models after knowledge distillation, the improved student models 1 and 2 showed significant performance improvements in all inversion indicators. Especially in the key LST band, the PCCS of student model 1 increased by 0.014 to 0.984; the MAE accuracy increased by 0.368 to 0.731. The PCCS accuracy of student model 2 increased by 0.025 to 0.973; MAE decreased by 0.313K to 0.716K. This highlights the significant effect of knowledge distillation in improving model accuracy and reducing prediction errors. In the emissivity bands from LSE10 to LSE14, the two improved models also showed significant performance improvements, among which the PCCS of student model 1 in the five emissivity bands increased by an average of 0.058 to 0.917, and the MAE decreased by an average of 0.0020 to 0.0044. The PCCS of student model 2 increased by an average of 0.083 to 0.879, and the MAE decreased by an average of 0.0024 to 0.0064. At the same time, the percentage improvement of LST&LSE inversion accuracy between student model 1 and student model 2 is shown in Table 5, which shows the percentage improvement of accuracy of the smaller model of the student model in each inversion parameter.

[0159] Table 5

[0160]

[0161] Through the above analysis of the results, it can be concluded that the knowledge distillation strategy adopted in this embodiment effectively improves the accuracy of surface temperature and emissivity inversion, especially when dealing with complex data and limited resources. It shows great potential and provides important support for improving the robustness and reliability of the model in practical applications. Finally, the high-precision and generalized teacher model is interactively integrated with the optimized small model to form a high-quality joint optimization model. In the joint optimization model, a regionally adaptive model is selected according to the source partition of the data to further enhance the pertinence and generalization ability of the model.

[0162] 7. Cross-validation and site validation

[0163] Cross-validation is valuable in identifying spatial differences between products and retrieved values, especially in the context of spatial resolution and time series variation of different satellite datasets. This example selects MODIS surface temperature datasets corresponding to two specific regional data and performs a detailed cross-validation on the small model and its student model after knowledge distillation optimization. The analysis includes four sets of comparison graphs ( Fig.10 )--small model 1 vs. student model 1, small model 2 vs. student model 2, covering the MODIS surface temperature map (the map of the student model is the surface temperature inversion map), the scatter plot of the model predictions, and the distribution map of the prediction errors, so as to comprehensively evaluate the model performance.

[0164] Through cross-validation, it was found that the MAE of surface temperature inversion of student model 1 before and after knowledge distillation decreased from 2.416K to 1.730K, and PCCS increased from 0.832 to 0.912; the MAE of LST inversion of student model 2 decreased from 2.345 to 1.666, and PCCS increased from 0.777 to 0.893, proving the effectiveness of the knowledge distillation optimization strategy in improving prediction accuracy. A careful observation of the error distribution diagram reveals that the error range of the optimized model has been reduced, and the frequency of zero errors has increased significantly, indicating that the consistency between model predictions and actual observations has been significantly enhanced.

[0165] In addition, by comparing the MODIS temperature distribution map with the predicted temperature distribution map, the optimized model is not only similar to the MODIS surface temperature distribution in general, but also shows richer details by using the high resolution of Aster data. This analysis highlights the practical value of the knowledge distillation strategy in fine-tuning the surface temperature inversion model to transfer teacher knowledge.

[0166] Study Site Data Site Validation

[0167] Compared with emissivity, site data of surface temperature have higher availability, so this embodiment mainly focuses on verifying the effectiveness of the proposed method by comparing the surface temperature inverted by satellite radiation with the ground site measurement data. Since the monitoring method of ground field measurement directly affects the accuracy of data acquisition, in order to eliminate interference, through strict data quality control, representative and high-precision field measurement data with good clear sky quality and flat terrain were selected. Further considering the complexity of site verification, obtaining robust statistics requires appropriate outlier removal strategies. To this end, the new 3-sigma criterion proposed by Sobrino et al. (2016) based on the theoretical uncertainty associated with the algorithm was adopted. Specifically, the outliers of the difference between the retrieved and site surface temperatures were removed using a threshold of three times the theoretical accuracy. Subsequently, site verification was performed on small model 1, student model 1, small model 2, and student model 2, and their scatter plots and accuracy are shown as follows. Fig.11 shown.

[0168] Through the optimization of the knowledge distillation strategy, student model 1 showed significant performance improvement in the site data verification, with its PCCS accuracy increased from 0.963 to 0.986, and the MAE loss decreased from 1.953K to 1.315K. Similarly, after optimization, student model 2 improved from 0.940 to 0.970 on PCCS, and the MAE loss decreased from 1.988K to 1.414K. Although the accuracy in the site data verification is slightly lower than the verification result based on ASTER satellite data, considering the impact of the real physical conditions represented by the site data on the surface temperature, such as local terrain changes, vegetation coverage, and thermal properties of ground materials, the results of the site verification have confirmed the reliability and practical value of the proposed method to a certain extent. This further confirms the robustness and effectiveness of the knowledge distillation strategy in the face of complex real physical conditions.

[0169] Beneficial effects of this embodiment:

[0170] This example uses the AutoKeras automated neural network architecture search tool in automatic machine learning (AutoML) to reduce manual intervention and optimize the model structure. Through its automated model architecture selection and hyperparameter adjustment, the model development cycle and computing resource consumption can be effectively reduced, providing an efficient and flexible solution for remote sensing parameter inversion.

[0171] This embodiment uses the cutting-edge technology of artificial intelligence to solve the problem of limited precision of small model inversion due to the high complexity of data caused by the diversification of surface coverage due to the complexity and nonlinear characteristics of the earth system. A new strategy of large model-small model joint optimization inversion that integrates knowledge distillation technology is adopted, called AK-KD method, which aims to study a method paradigm for solving the difficulty of small model inversion using large models. First, based on the data-driven nature of the inversion data set, the optimal neural network architecture for model adaptation is obtained and the model is trained based on AutoKeras. Then the trained high-precision and generalized large model is introduced as the teacher model, and the small model in the local area where the precision is still limited is used as the student model. First, through pre-distillation, the best guidance weight is obtained by relying on statistical index verification, and then the knowledge distillation technology is formally used to pass the deep knowledge of the large model to the small model, so as to improve the inversion accuracy of the small model in the precision-limited area, so as to be closer to the representation of the actual physical characteristics. In order to realize a high-precision parameter inversion model for a specific area.

[0172] In particular, knowledge distillation has many advantages over traditional model training methods. By adopting the strategy of knowledge distillation, the teacher model demonstrates generalization capabilities on a global scale, which can effectively solve the challenges of student model data complexity, especially under complex terrain conditions. This method enables the small model to simulate the behavior patterns of the large model, exclude abnormal data during the inversion process, and improve the performance of the small model in specific application scenarios by reducing background noise, improving feature extraction, and enhancing the generalization ability of the model. Finally, the student model is interactively integrated into the teacher model to form a high-quality joint optimization model, which further enhances the model's targeting and generalization capabilities.

[0173] Embodiment 2

[0174] This embodiment provides an AI size model surface temperature and emissivity inversion system, which is used to implement the AI ​​size model surface temperature and emissivity inversion method described in Embodiment 1. The system includes:

[0175] A data acquisition module, used to obtain an inversion data set of geophysical parameters;

[0176] A large model building module is used to obtain the optimal neural network architecture for model adaptation and train the model based on the data drive of the inversion data set using the automatic machine learning tool AutoKeras, and introduce the trained large model as a teacher model based on the optimal neural network architecture;

[0177] A knowledge transfer module, used to transfer the deep knowledge of the teacher model to the small model to form a student model by using knowledge distillation technology;

[0178] The joint optimization module is used to interactively integrate the student model into the teacher model to form a joint optimization model, and use the joint optimization model to invert the geophysical parameters to be measured.

[0179] This embodiment has all the advantages of the first embodiment.

[0180] Embodiment 3

[0181] 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, the AI ​​size model surface temperature and emissivity inversion method described in Example 1 is implemented.

[0182] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. AI size model surface temperature and emissivity inversion method, characterized by: The following steps are involved: Acquire an inversion data set of geophysical parameters; the inversion data set is ASTER data; Based on the data-driven nature of the inversion dataset, the automatic machine learning tool AutoKeras is used to obtain the optimal neural network architecture for model adaptation and train the model. Based on the optimal neural network architecture, the trained large model is introduced as a teacher model. Using knowledge distillation technology, the deep knowledge of the teacher model is transferred to the small model to form a student model; interactively integrating the student model into the teacher model to form a joint optimization model, and using the joint optimization model to invert the geophysical parameters to be measured; The process of using the automatic machine learning tool AutoKeras to obtain the optimal neural network architecture for model adaptation includes: The inversion data set of geophysical parameters is obtained through Aster satellite acquisition; Preprocessing the inversion data set to obtain preprocessed data; Input the preprocessed data into the AutoKeras system and define the search space of the neural structure; Initialize neural architecture search using AutoKeras automatic machine learning technology; A Bayesian optimization algorithm is used to guide the neural architecture search, and a Gaussian process is used to control network deformation to explore a predefined search space; Based on the search space, automatically searching for an optimal neural network architecture within a predetermined number of iterations; Using knowledge distillation technology, the process of transferring the deep knowledge of the teacher model to the small model to form a student model includes: Through pre-distillation, the best guidance weights are obtained by relying on statistical index verification; Based on the optimal guidance weights, the deep knowledge of the teacher model is transferred to the small model to form a student model through formal distillation; The process of obtaining the optimal guidance weights includes: Based on the independent sample t-test method, the t statistic is obtained through the t statistic calculation formula; The probability density function of t distribution is introduced to understand the distribution of the t statistic under given degrees of freedom. Use the cumulative distribution function of the t distribution to determine the probability p value of the t statistic; According to the probability p-value and t-statistic, the best guidance weight is obtained.

2. The AI ​​size model surface temperature and emissivity inversion method according to claim 1 is characterized in that: The search space includes: the number of layers of the neural network, the number of units in each layer, the activation function and the dropout probability; The search space is specifically described as follows: Where n is the number of neural network layers, u i is the number of units in the i-th layer, φ i is the activation function of the i-th layer, p i is the Dropout probability of the i-th layer.

3. The AI ​​size model surface temperature and emissivity inversion method according to claim 1 is characterized in that: The process of using the Bayesian optimization algorithm to guide the neural structure search includes: Use the edit distance kernel function to evaluate the similarity between different neural network architectures and guide network deformation during the search process; The neural network kernel function of the edit distance is defined as follows: Among them, d(f a , f b ) represents the edit distance between two neural networks, and ρ is the mapping function.

4. The AI ​​size model surface temperature and emissivity inversion method according to claim 1 is characterized in that: It also includes a cost function, which is used to evaluate the search efficiency of the neural network architecture; The optimal neural network architecture is: The learning parameters of the optimal neural network architecture are: Among them, Cost is the evaluation metric function and θ is the learning parameter of f.

5. The AI ​​size model surface temperature and emissivity inversion method according to claim 1 is characterized in that: Also includes: An evaluation function is set to analyze the contribution of different weights to the final model accuracy improvement, and the relative utility of the weights is quantified through the t-statistic.

6. AI size model surface temperature and emissivity inversion system, characterized by: For implementing the AI ​​size model surface temperature and emissivity inversion method according to any one of claims 1 to 5, the system comprises: A data acquisition module, used to acquire an inversion data set of geophysical parameters; the inversion data set is ASTER data; A large model building module is used to obtain the optimal neural network architecture for model adaptation and train the model based on the data drive of the inversion data set using the automatic machine learning tool AutoKeras, and introduce the trained large model as a teacher model based on the optimal neural network architecture; A knowledge transfer module, used to transfer the deep knowledge of the teacher model to the small model to form a student model by using knowledge distillation technology; A joint optimization module, used for interactively integrating the student model into the teacher model to form a joint optimization model, and using the joint optimization model to invert the geophysical parameters to be measured; Using knowledge distillation technology, the process of transferring the deep knowledge of the teacher model to the small model to form a student model includes: Through pre-distillation, the best guidance weights are obtained by relying on statistical index verification; Based on the optimal guidance weights, the deep knowledge of the teacher model is transferred to the small model to form a student model through formal distillation; The process of obtaining the optimal guidance weights includes: Based on the independent sample t-test method, the t statistic is obtained through the t statistic calculation formula; The probability density function of t distribution is introduced to understand the distribution of the t statistic under given degrees of freedom. Use the cumulative distribution function of the t distribution to determine the probability p value of the t statistic; According to the probability p-value and t-statistic, the best guidance weight is obtained.

7. An electronic device, characterized in that: The device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the method for inverting the surface temperature and emissivity of the AI ​​size model as described in any one of claims 1 to 5 is implemented.

Images