Calibration reference source data simulation method and device, electronic equipment and program product

Abstract

The embodiment of the invention discloses a calibration reference source data simulation method and device, electronic equipment and a program product, and the method comprises the steps: preprocessing first calibration related data of a to-be-calibrated infrared instrument and calibration reference source data of a reference infrared instrument, and generating a model training data set; the infrared instrument to be calibrated and the reference infrared instrument are infrared instruments respectively loaded on different satellites; the model training data set comprises a plurality of data matching pairs, and the data matching pairs comprise sample input features obtained by the first calibration related data and corresponding label features obtained by the calibration reference source data; constructing a mixed attention multilayer perceptron model based on the model training data set, and learning an end-to-end mapping relation from first calibration related data to calibration reference source data in a data driving mode; and inverting the second calibration related data of the to-be-calibrated infrared instrument into observation data of the reference infrared instrument by using the constructed mixed attention multilayer perceptron model.

Description

Technical Field

[0001] This disclosure relates to the field of satellite remote sensing technology, specifically to a calibration reference source data simulation method, apparatus, electronic equipment, and program product. Background Technology

[0002] In-orbit radiometric calibration of satellite remote sensing instruments is the cornerstone of ensuring the accuracy of their quantitative applications. Cross-calibration, as an important in-orbit calibration method, transfers the calibration accuracy of the reference infrared instrument to the instrument being calibrated by conducting joint observations of the same Earth target with a high-precision, high-stability reference infrared instrument at the same time. The GMI (GPM Microwave Imager) instrument on the GPM (Global Precipitation Observation Mission) satellite is often used as a calibration reference source for other microwave radiometers due to its extremely high calibration accuracy and stability.

[0003] However, achieving high-precision cross-calibration faces many challenges: First, it is difficult for two satellites to achieve strict spatiotemporal and observational geometry matching, and small differences can introduce significant errors; second, the observational data coverage of the reference infrared instrument (such as GMI) is limited and cannot meet the high-frequency calibration requirements of the infrared instrument to be calibrated (such as FY-3D / MWRI); third, directly using the observational matching data for analysis is limited by the quantity and quality of the matching samples, making it difficult to establish a robust and universal calibration relationship.

[0004] Existing techniques typically rely on establishing physical models based on radiative transfer theory or simplified empirical statistical models to correlate observations from two instruments. While physical models have clear mechanisms, they are computationally complex and require precise input of a large number of atmospheric and surface parameters, which themselves contain uncertainties. Empirical statistical models (such as multiple linear regression), although simple, struggle to capture the complex nonlinear relationship between instrument observation signals and the final radiative brightness temperature, as well as the combined influence of numerous internal instrument state parameters (such as temperature) on the observations. This leads to a decrease in model accuracy and robustness when instrument conditions change or long-term operational performance degrades.

[0005] In recent years, deep learning technology, especially deep neural networks, has demonstrated powerful capabilities in nonlinear relationship modeling and complex pattern recognition. Multilayer perceptrons (MLPs), as a fundamental type of deep neural network, possess powerful general function approximation capabilities. However, standard MLPs often underperform when processing high-dimensional data with multiple features and strong internal correlations and dependencies, such as satellite instrument parameters. Standard MLPs treat all input features equally, failing to adaptively focus on features crucial to the output, and struggle to effectively model complex interaction sequences between features.

[0006] Therefore, a solution is needed that can overcome the above-mentioned defects, deeply integrate instrument physical state information, adaptively learn complex nonlinear mappings, and thus simulate and calibrate reference source data with high fidelity. Summary of the Invention

[0007] This disclosure provides a method, apparatus, electronic device, and program product for simulating calibration reference source data.

[0008] In a first aspect, this disclosure provides a method for simulating calibration reference source data, comprising: The first calibration-related data of the infrared instrument to be calibrated and the calibration reference source data of the reference infrared instrument are preprocessed to generate a model training dataset; the infrared instrument to be calibrated and the reference infrared instrument are infrared instruments mounted on different satellites respectively; the model training dataset includes multiple data matching pairs, and the data matching pairs include sample input features obtained from the first calibration-related data and their corresponding label features obtained from the calibration reference source data; Based on the model training dataset, a hybrid attention multilayer perceptron model is constructed to learn the end-to-end mapping relationship between the first calibration-related data and the calibration reference source data in a data-driven manner. Using the constructed hybrid attention multilayer perceptron model, the second calibration correlation data of the infrared instrument to be calibrated is inverted into the observation data of the reference infrared instrument.

[0009] The hybrid attention multilayer perceptron model combines channel attention, self-attention, and cross-attention mechanisms.

[0010] The hybrid attention multilayer perceptron model includes a feature augmentation module, a balancing layer, and a feature extraction module, wherein the feature augmentation module and the feature extraction module have a symmetrical structure. The feature augmentation module is used to perform linear transformation and feature augmentation on the sample input features, progressively increasing the feature dimension and generating multi-scale features; The balancing layer is used to map the output of the last layer of the feature augmentation module to a higher-dimensional latent feature vector; The feature extraction module takes the latent space vector as the initial input, and combines the multi-scale features output by the feature expansion module to perform dimensionality reduction and reconstruction layer by layer to generate dense features.

[0011] The feature augmentation module comprises N layers, each layer including a channel attention submodule, a fully connected transformation submodule, and a self-attention submodule connected in sequence. The channel attention submodule learns the correlation between the current data matching pair and each channel in the sample input features through adaptive channel weighting. The fully connected transformation submodule is used to perform basic linear transformation and feature augmentation operations on the features weighted by the channel attention submodule; The self-attention submodule is used to treat the output of the fully connected transformation submodule as a sequence and model the dependencies between the sequences.

[0012] The feature extraction module includes N layers, each of which includes a cross-attention submodule, a residual connection submodule, and a fully connected transformation submodule connected in sequence. The cross-attention submodule is used to selectively retrieve relevant information from the multi-scale features extracted by the feature expansion module, so as to simultaneously focus on the low-level detailed information output by the feature expansion module and the high-level comprehensive information output by the feature extraction module. The residual connection submodule is used to ensure that the multi-scale features obtained by the feature augmentation module can be losslessly transmitted to the decoding process of the feature extraction module; The fully connected submodule is used to perform dimensionality reduction and linear transformation on the output of the residual connected submodule to keep the feature dimension the same as the output of the corresponding layer of the feature expansion module.

[0013] Specifically, the hybrid attention multilayer perceptron model is used to invert the second calibration correlation data of the infrared instrument to be calibrated into the observation data of the reference infrared instrument, including: The second calibration-related data is preprocessed to obtain multi-channel input features; The preprocessed input features are fed into the constructed hybrid attention multilayer perceptron model to obtain normalized brightness temperature values; The normalized brightness temperature value output by the hybrid attention multilayer perceptron model is denormalized to restore it to a brightness temperature value with actual physical meaning and units.

[0014] Secondly, this disclosure provides a calibration reference source data simulation apparatus, comprising: The preprocessing module is configured to preprocess the first calibration-related data of the infrared instrument to be calibrated and the calibration reference source data of the reference infrared instrument to generate a model training dataset; the infrared instrument to be calibrated and the reference infrared instrument are infrared instruments mounted on different satellites respectively; the model training dataset includes multiple data matching pairs, and the data matching pairs include sample input features obtained from the first calibration-related data and their corresponding label features obtained from the calibration reference source data; The building module is configured to build a hybrid attention multilayer perceptron model based on the model training dataset, and learn the end-to-end mapping relationship between the first calibration-related data and the calibration reference source data in a data-driven manner. The inversion module is configured to use the constructed hybrid attention multilayer perceptron model to invert the second calibration correlation data of the infrared instrument to be calibrated into the observation data of the reference infrared instrument.

[0015] The function can be implemented by hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above function.

[0016] In one possible design, the above-described device includes a memory and a processor. The memory stores one or more computer instructions that support the device in performing the corresponding methods described above, and the processor is configured to execute the computer instructions stored in the memory. The device may also include a communication interface for communicating with other devices or communication networks.

[0017] Thirdly, embodiments of this disclosure provide an electronic device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described in any of the above aspects.

[0018] Fourthly, embodiments of this disclosure provide a computer-readable storage medium for storing computer instructions used by any of the above-described devices, which, when executed by a processor, are used to implement the methods described in any of the above aspects.

[0019] Fifthly, embodiments of this disclosure provide a computer program product comprising computer instructions which, when executed by a processor, are used to implement the methods described in any of the preceding aspects.

[0020] The technical solutions provided in this disclosure may have the following beneficial effects: 1. High-precision simulation: By introducing a hybrid attention mechanism (including a combination of channel attention, self-attention, and cross attention), the model can adaptively focus on key instrument state parameters, capture the complex nonlinear interactions and time-series dependencies between parameters, and thus learn a deeper and more accurate mapping function from the original state of the infrared instrument to be calibrated to the calibration reference simulation data of the reference infrared instrument. The simulation accuracy is higher than that of existing linear or simple nonlinear statistical models.

[0021] 2. Strong feature fusion capability: The multi-dimensional input features constructed in this disclosure, combined with the channel attention mechanism, enable the model to intelligently fuse multi-source heterogeneous information from different subsystems of the instrument (optical path, receiver, antenna, calibration source), making full use of all available information for inference.

[0022] 3. Excellent generalization and robustness: The cross-attention and residual connections designed in the network structure ensure that the multi-scale features extracted in the encoding stage can be effectively utilized by the feature extraction module, which enhances the model's adaptability to different instrument states (such as temperature fluctuations) and observation conditions, reduces the risk of overfitting, and improves the model's generalization performance on unseen data.

[0023] 4. End-to-end automation: It realizes end-to-end modeling from raw observation data to calibration reference source data, avoiding complex manual feature engineering and physical model parameter debugging. It has a high degree of automation and is easy to integrate into business and promote to cross-calibration scenarios of other sensor pairs.

[0024] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0025] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments, taken in conjunction with the accompanying drawings. In the drawings: Figure 1 A flowchart is shown for a calibration reference source data simulation method according to an embodiment of the present disclosure.

[0026] Figure 2 A flowchart illustrating a method for simulating FY-3D / MWRI observation data into GPM / GMI according to an embodiment of the present disclosure is shown.

[0027] Figure 3 A structural block diagram of a calibration reference source data simulation apparatus according to an embodiment of the present disclosure is shown.

[0028] Figure 4 This is a schematic diagram of the structure of an electronic device suitable for implementing a calibration reference source data simulation method according to an embodiment of the present disclosure. Detailed Implementation

[0029] In the following, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings to enable those skilled in the art to readily implement them. Furthermore, for clarity, portions unrelated to the description of the exemplary embodiments have been omitted from the drawings.

[0030] In this disclosure, it should be understood that terms such as “comprising” or “having” are intended to indicate the presence of features, figures, steps, behaviors, components, parts or combinations thereof disclosed in this specification, and do not preclude the possibility of the presence or addition of one or more other features, figures, steps, behaviors, components, parts or combinations thereof.

[0031] It should also be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0032] This disclosure aims to address the problems of related technologies in simulating calibration reference source data, such as strong dependence on complex physical models, high computational costs, and / or difficulty in capturing nonlinear and high-dimensional correlation features. Its core idea is to utilize historical observation data as a training set to construct a multilayer perceptron model that integrates channel attention and spatial attention. This model learns the end-to-end mapping relationship between input conditions (such as instrument status, environmental parameters, and target parameters) and high-precision calibration reference source data (such as radiometric values, spectra, and images) in a data-driven manner, thereby achieving fast, flexible, and high-fidelity simulation of the calibration reference source.

[0033] This disclosure provides a method for simulating calibration reference source data based on a hybrid attention multilayer perceptron model. This method constructs a hybrid attention multilayer perceptron network (MLP) that deeply integrates channel attention and self / cross-attention mechanisms. This network can automatically learn and extract a deep, nonlinear mapping relationship between the raw observation signals and internal state parameters of the infrared instrument to be calibrated and the observation data of the reference infrared instrument. This enables high-precision and high-efficiency simulation of the calibration reference source data, providing stable, reliable, and sufficient data support for on-orbit cross-calibration of satellite instruments.

[0034] To achieve the above objectives, the technical solution adopted in this disclosure is a calibration reference source data simulation method based on a hybrid attention multilayer perceptron model. The core of this method is to use a hybrid attention multilayer perceptron network to invert the calibration-related data of FY-3D / MWRI into observation data of the cross-calibration reference sources GPM / GMI, thereby increasing the cross-calibration reference source data and providing a stable reference source for cross-calibration.

[0035] The details of the embodiments of this disclosure are described in detail below through specific examples.

[0036] Figure 1 A flowchart illustrating a calibration reference source data simulation method according to an embodiment of this disclosure is shown. Figure 1 As shown, the calibration reference source data simulation method includes the following steps: In step S101, the first calibration-related data of the infrared instrument to be calibrated and the calibration reference source data of the reference infrared instrument are preprocessed to generate a model training dataset; the infrared instrument to be calibrated and the reference infrared instrument are infrared instruments mounted on different satellites respectively; the model training dataset includes multiple data matching pairs, and the data matching pairs include sample input features obtained from the first calibration-related data and their corresponding label features obtained from the calibration reference source data; In step S102, a hybrid attention multilayer perceptron model is constructed based on the model training dataset to learn the end-to-end mapping relationship between the first calibration-related data and the calibration reference source data in a data-driven manner. In step S103, the constructed hybrid attention multilayer perceptron model is used to invert the second calibration correlation data of the infrared instrument to be calibrated into the observation data of the reference infrared instrument.

[0037] In this embodiment, the infrared instrument to be calibrated and the reference infrared instrument can be infrared instruments mounted on different satellites. The first calibration-related data may include, but is not limited to, the instrument parameters of the infrared instrument to be calibrated and the raw observation data. The raw observation data may be the raw observation count (i.e., DN value) recorded by the infrared instrument to be calibrated. The instrument parameters of the infrared instrument to be calibrated may include, but are not limited to, the instrument status, environmental parameters, and raw observation counts. The calibration reference source data may include the brightness temperature value of the reference infrared instrument or the image corresponding to the brightness temperature value.

[0038] In some embodiments, the first calibration-related data and the calibration reference source data are preprocessed and standardized to obtain multidimensional feature data corresponding to the first calibration-related data, i.e., sample input features, and 1-dimensional feature data corresponding to the brightness temperature value, i.e., label features.

[0039] The purpose of preprocessing the first calibration-related data and the calibration reference source data is to obtain data matching pairs for model training and validation. Each data matching pair is a sample used for model training and validation, which includes a matching sample input feature pair and label features.

[0040] In some embodiments, the above data matching pairs are obtained by the following method: The first calibration-related data and the calibration reference source data of the reference infrared instrument are matched in time and space to obtain a data matching pair; The data matching pairs are standardized to obtain data matching pairs.

[0041] The first calibration-related data and the calibration reference source data have corresponding temporal and spatial information, and data matching pairs are formed by matching both time and space. After standardizing all data pairs in the data pair set, data matching pairs are obtained. In this step, the maximum and minimum values ​​of the standardized data can also be used to de-standardize the normalized brightness temperature values ​​output by the subsequent model, thereby generating the calibration reference source data of the simulated reference infrared instrument.

[0042] In some embodiments, the first calibration-related data and the calibration reference source data of the reference infrared instrument are matched in time and space to obtain a data matching pair, including: Unify the geographic coordinates of the first calibration-related data and the calibration reference source data; Based on unified geographic coordinates, determine the spatial candidate pairs with the closest spatial distance between the first calibration-related data and the calibration reference source data; Eliminate spatial candidate pairs whose spatial distance exceeds a set threshold; Spatial candidate pairs whose observation time difference exceeds the time threshold are removed, and the remaining spatial candidate pairs are determined as the data matching pairs.

[0043] The following describes a feasible preprocessing procedure using the FY-3D / MWRI infrared instrument to be calibrated and the GPM / GMI reference infrared instrument as an example. The first calibration-related data uses the raw observation data of the FY-3D / MWRI and the instrument parameters of the GMI as examples, while the calibration reference source data uses the brightness temperature value of the GMI as an example.

[0044] See Figure 2 As shown, for the cross-calibration scenario of FY-3D / MWRI and GPM / GMI, the observation data of GMI is used as the calibration reference ground truth. A mapping dataset from the calibration-related data of MWRI to the brightness temperature data of GMI is constructed. The specific sub-steps include: Sub-step 1.1: Spatiotemporal matching of cross-calibrated data pairs

[0045] The goal of this sub-step is to establish a precise one-to-one correspondence between FY-3D / MWRI observation data and GPM / GMI observation data, which serve as the calibration reference source. Considering the inherent structure of standard L1 level data products, this embodiment employs a precise index matching method based on a Geolocation Look-up Table (GLT). This method directly utilizes the geographic coordinate grid provided by the instrument data, which is strictly aligned with the observation values, to establish a matching relationship through coordinate comparison. It offers advantages such as high accuracy and no resampling error. Details are as follows: 1. Data understanding and confirmation of index relationships: FY-3D / MWRI data: Using FY-3D / MWRI's L1 A-level data products as calibration-related data, it typically consists of two core parts: 1) Scientific Dataset (SDS): Contains raw observation counts (DN values) for each channel, as well as instrument status parameters (such as thermal mirror temperature, receiver temperature, calibration count values, etc.). The scientific dataset is organized and stored in a two-dimensional grid (Grid_Science) of scan lines × pixels.

[0046] 2) Geolocation Dataset: This dataset contains the precise latitude and longitude coordinates, observation time, etc., corresponding to each scan line and each pixel. The data is organized in a two-dimensional grid (Grid_Geo_MWRI) of the same dimension and corresponds strictly one-to-one with the grid of the scientific dataset (Grid_Science). That is, the geographic coordinates of the observation value at position (i, j) in the scientific data grid are the latitude and longitude values ​​of the same position (i, j) in the geolocation grid.

[0047] GPM / GMI data: Using the L1 level data product of GPM / GMI as the calibration reference source data, its structure is similar to that of the L1 A level data product of FY-3D / MWRI, and includes: 1) Scientific dataset: Contains brightness temperature data (TB) for each channel, which is organized and stored in a two-dimensional grid (Grid_TB) of scan lines × pixels.

[0048] 2) Geolocation Dataset: Contains corresponding latitude and longitude values ​​and observation time. The grid (Grid_Geo_GMI) strictly corresponds one-to-one with the brightness temperature data grid Grid_TB. That is, the geographic coordinates of the observation value at position (i, j) in the scientific data grid are the latitude and longitude values ​​of the same position (i, j) in the geolocation grid.

[0049] 2. Matching strategy: Based on the above data structure, the core of the matching lies in directly comparing the coordinate values ​​in the geographic positioning grid of the infrared instrument to be calibrated and the reference infrared instrument, without any resampling or interpolation of their respective scientific datasets.

[0050] Inputs: Grid_Geo_MWRI (MWRI grid: including latitude and longitude, observation time), Grid_Science (MWRI raw observation counts and instrument status parameters); Grid_Geo_GMI (GMI grid: including latitude and longitude, observation time), Grid_TB (GMI brightness temperature value).

[0051] Matching objective: Find all (i_mwri, j_mwri) and (i_gmi, j_gmi) index pairs that satisfy the conditions of spatial proximity (i.e., spatial distance less than a set threshold) and time synchronization.

[0052] 3. Matching process: To improve efficiency, this embodiment first uses space to partition the area, then performs nearest neighbor search within each partition, and finally uses time filtering.

[0053] Spatial matching:

[0054] 1) Traverse each valid position P_mwri (i_m, j_m) in Grid_Geo_MWRI and obtain its latitude and longitude coordinates (Lon_m, Lat_m).

[0055] 2) In Grid_Geo_GMI, quickly find the point P_gmi (i_g, j_g) that is spatially closest to P_mwri. This is typically achieved using spatial indexed data structures (such as KD-Tree) to handle massive amounts of data.

[0056] 3) Calculate the distance Dist between P_mwri and P_gmi.

[0057] If Dist ≤ spatial threshold (e.g., 0.02° corresponds to approximately 2.2 km), then record this spatial candidate pair [(i_m, j_m), (i_g, j_g)] and its distance Dist.

[0058] Time matching: 1) For each spatial candidate pair obtained in the previous step, extract the observation times T_m and T_g from Grid_Geo_MWRI(i_m, j_m) and Grid_Geo_GMI(i_g, j_g), respectively.

[0059] 2) Calculate the observation time difference ΔT = |T_m - T_g|.

[0060] 3) If ΔT ≤ time threshold (e.g., 300 seconds), then the index pair becomes a valid data matching pair through spatiotemporal matching.

[0061] 4. Matching result output: Finally, a matching index list is output for each data matching pair. Each row of the list corresponds to a valid matching pair, and this list serves as the overall index for all subsequent data reading operations. Based on MWRI_Line (scan line) and MWRI_Pixel (pixel), 28 input features can be extracted from the MWRI L1A scientific dataset; based on GMI_Line and GMI_Pixel, the ground truth brightness temperature of the target channel can be extracted from the GMI L1 scientific dataset. This list must contain at least the following fields: MWRI_Line: The scan line number (or row index i_m) in the MWRI data. MWRI_Pixel: The pixel number (or column index j_m) in the MWRI data. GMI_Line: Scan line number (or row index i_g) in the GMI data. GMI_Pixel: The pixel number (or column index j_g) in the GMI data. Distance_km: Spatial distance (kilometers) (i.e., the observation distance between two data points in a data matching pair) Time_Delta_s: Time difference (seconds) (i.e., the observation time difference between the two data points in the data matching pair) Thus, through the above steps, a precise "bridge" connecting the MWRI source data and the GMI reference data has been obtained—namely, the list of matching indices for the data matching pairs. Subsequent data extraction, cleaning, and standardization operations will be strictly performed according to this list of matching indices to ensure that every data pair input into the model training has strict spatiotemporal consistency.

[0062] Sub-step 1.2: Construction of multi-dimensional input features

[0063] From the data matching pairs, the features corresponding to the FY-3D / MWRI data (the 28 input features extracted above) are selected as model inputs, and the true brightness temperature values ​​of the corresponding channels of GPM / GMI are selected as the simulation targets for model output. For example, the brightness temperature data of the H polarization and 10.7 GHz channel of GMI are selected for simulation, and the corresponding input of MWRI is the raw observation count (DN value) of the H polarization and 10.65 GHz channel as the core signal. To achieve high-precision simulation, the model input is not limited to a single signal, but comprehensively utilizes multiple parameters that reflect the overall working state of the instrument.

[0064] The model constructed in this embodiment contains 28 input features, each of which is one-dimensional data excluding location information, and one label feature, specifically: Feature Name I / O Number of channels Remark MWRI raw signal (DN value) I 1 H-polarization, 10.65 GHz channel MWRI thermal mirror temperature I 9 Nine thermal mirror temperatures reflect the thermal state of optical path components. MWRI receiver temperature I 5 Five receiver temperatures reflect the thermal state of the receiver link. MWRI Space and Interior Target Calibration Observations I 1 The average value of H-polarization and 10.65 GHz channel calibration observation points is used to provide a real-time calibration reference point. MWRI antenna and feed horn temperature I 6 Six antenna and feedhorn temperatures reflect the thermal state of the antenna system. MWRI Earth's Center of Mass Vector I 2 The two axes of the vector reflect the attitude of the satellite platform. MWRI Moon Vector Information I 3 Used to assist in attitude or interference determination input. MWRI background radiation brightness temperature estimate I 1 As input of prior physical knowledge GMI calibration brightness temperature O 1 H-polarization, 10.65 GHz channel, used for training labels. Sub-step 1.3: Data cleaning and alignment

[0065] In this step, records in the data matching pair that have missing, invalid, or obviously outlier values ​​(such as brightness temperature exceeding the physical reasonable range) are removed.

[0066] Sub-step 1.4: Data Standardization

[0067] The cleaned 28-dimensional input features and the 1-dimensional sample label brightness temperature ground truth are subjected to min-maximum standardization, normalizing the data of each dimension to the [0, 1] interval. The calculation formula is: X_norm = (X - X_min) / (X_max - X_min). X_norm is the normalized value, X is the currently standardized data (which can be the data corresponding to one dimension of the 28-dimensional input features or the brightness temperature ground truth), X_min is the minimum value of the currently standardized data among all data matching pairs, and X_max is the maximum value of the currently standardized data among all data matching pairs. This step can eliminate the influence of differences in physical units and numerical ranges on the stability and speed of model training. It should be noted that the maximum and minimum values ​​corresponding to each dimension of the input features and the brightness temperature ground truth can be saved during this process for subsequent validation, testing, and new data transformations.

[0068] Sub-step 1.5: Dataset Construction and Partitioning

[0069] The standardized data pairs are shuffled chronologically to avoid time-related biases in model training. They are then divided into training, validation, and test sets according to a preset ratio (e.g., 7:2:1). The training set is used for model parameter learning, the validation set is used for monitoring the training process and tuning hyperparameters, and the test set is used for independent evaluation of the final model performance.

[0070] In some alternative embodiments, the hybrid attention multilayer perceptron model combines channel attention, self-attention, and cross-attention mechanisms.

[0071] In some optional embodiments, the hybrid attention multilayer perceptron model includes a feature augmentation module, a balancing layer, and a feature extraction module, wherein the feature augmentation module and the feature extraction module have a symmetrical structure; The feature augmentation module is used to perform linear transformation and feature augmentation on the sample input features, progressively increasing the feature dimension and generating multi-scale features; The balancing layer is used to map the output of the last layer of the feature augmentation module to a higher-dimensional latent feature vector; The feature extraction module takes the latent space vector as the initial input, and combines the multi-scale features output by the feature expansion module to perform dimensionality reduction and reconstruction layer by layer to generate dense features.

[0072] In some optional embodiments, the feature augmentation module includes N layers, each layer including a channel attention submodule, a fully connected transformation submodule, and a self-attention submodule connected in sequence. The channel attention submodule learns the correlation between the current data matching pair and each channel in the sample input features through adaptive channel weighting. The fully connected transformation submodule is used to perform basic linear transformation and feature augmentation operations on the features weighted by the channel attention submodule; The self-attention submodule is used to treat the output of the fully connected transformation submodule as a sequence and model the dependencies between the sequences.

[0073] In some optional embodiments, the feature extraction module includes N layers, each layer including a cross-attention submodule, a residual connection submodule and a fully connected transformation submodule connected in sequence. The cross-attention submodule is used to selectively retrieve relevant information from the multi-scale features extracted by the feature expansion module, so as to simultaneously focus on the low-level detailed information output by the feature expansion module and the high-level comprehensive information output by the feature extraction module. The residual connection submodule is used to ensure that the multi-scale features obtained by the feature augmentation module can be losslessly transmitted to the decoding process of the feature extraction module; The fully connected submodule is used to perform dimensionality reduction and linear transformation on the output of the residual connected submodule to keep the feature dimension the same as the output of the corresponding layer of the feature expansion module.

[0074] In this embodiment, a deep neural network model specifically designed for processing instrument state sequence data is constructed and trained using data matching pairs obtained during data preprocessing. For ease of description, this disclosure refers to this model as a hybrid attention multilayer perceptron model. This model integrates channel attention mechanism, sequence self-attention mechanism, and encoder-decoder cross-attention mechanism, aiming to accurately capture the complex and nonlinear mapping relationship between the multi-source input of the infrared instrument to be calibrated and the brightness temperature of the reference infrared instrument, and to model the dependencies between parameters.

[0075] The hybrid attention MLP network model disclosed herein adopts a symmetrical "feature augmentation module - balancing layer - feature extraction module" structure, which achieves a good balance between information compression and feature reconstruction. The following explanation uses a network model comprising 10 computational layers (excluding the input layer) as an example. The output feature dimensions (number of neurons) of each computational layer are as follows: 32, 64, 128, 256, 512, 256, 128, 64, 32, 1. Assuming the total number of computational layers in the network model is M, and the number of layers in the feature augmentation module and feature extraction module is N, then N = (M-2) / 2. It should be noted that the above-mentioned computational layers do not refer to MLP network layers; a computational layer may include at least one MLP layer and / or other computational modules. The following explanation uses M=10 and N=4 as an example.

[0076] The feature augmentation module corresponds to layers 1-4 of this network model: it is responsible for taking the preprocessed 28-dimensional input feature vector. Linear transformations and feature augmentation are performed, progressively increasing the feature dimension to 256 dimensions, ultimately generating a high-level feature representation. The feature augmentation module outputs 32-dimensional, 64-dimensional, 128-dimensional, and 256-dimensional multi-scale features at each layer, namely layers 1-4.

[0077] The balancing layer corresponds to the 5th layer of the network model: as an information bottleneck and convergence point, it integrates the 256-dimensional high-level feature representation output from the last layer of the feature augmentation module. Further mapping to 512-dimensional latent feature vectors It is the most fundamental representation of input data for online learning.

[0078] The feature extraction module corresponds to layers 6-9 of the network model: outputting balanced layers. Starting from this point, and combining multi-scale features from the feature expansion module, the dimensionality is gradually reduced and the information is reconstructed, ultimately outputting a dense 32-dimensional feature.

[0079] The output layer corresponds to the 10th layer of the network model: it can be implemented by a linear fully connected layer, used to map the 32-dimensional dense features output by the feature extraction module into 1-dimensional normalized brightness temperature predictions. .

[0080] The network structures of the feature augmentation module, the balancing layer, and the feature extraction module are described below.

[0081] Each layer of the feature extension module ( Each module comprises three sub-modules connected in sequence: a channel attention sub-module, a fully connected transformation sub-module, and a self-attention sub-module, which process the input features... (in Processing: Channel Attention Submodule: This module emphasizes the relevance of key instrument parameters to the current context through adaptive channel weighting. In other words, this channel attention submodule learns the weights of each channel in the input features in relation to the simulated calibration reference source data, where each channel corresponds to one dimension of the input features. Its mathematical expression is as follows:

[0082] GAP refers to Global Average Pooling. This is the channel average, which is obtained by applying the input corresponding to the current channel of the attention submodule for that channel. The average value. Channel weights are used to learn the importance of the current channel when simulating calibration reference source data. The constraint is in the interval (0, 1). This channel attention submodule can include two layers of MLP, ( )and( , These are the network parameters of the two MLP layers, respectively. Activation function , For the Sigmoid function; This indicates element-wise multiplication. This disclosure uses a channel attention mechanism that enables the network model to dynamically evaluate the relative importance of input instrument parameters such as "thermal mirror temperature" and "receiver temperature" under different observation conditions.

[0083] Fully connected transformation submodule: Performs basic linear transformations on the features output by the channel attention submodule to increase the feature dimensionality. For example, the first layer of the feature augmentation module will increase the input features from 28 channels to 32 channels.

[0084]

[0085] in, The features are those that have been upgraded in dimensionality by the fully connected transformation submodule; These are the model parameters for the fully connected layer, BN represents batch normalization, and ReLU is the activation function.

[0086] The self-attention submodule treats the current batch of training data as a sequence and models the dependencies between sequences. For data containing... A batch of training datasets containing sample data. ( (To expand the feature dimension of the module layer for the current features), the calculation is as follows:

[0087] in, For query sequence, For keyword sequence, To prepare the numerical sequence for self-attention adjustment, This is the projection matrix, which is also the model parameter of the self-attention submodule. Key vector The feature dimensions. This is the output of the fully connected transformation submodule. This module enables the model to capture the overall patterns of instrument status over time (such as the systematic impact of diurnal temperature cycles on observations), rather than relying on a single independent sample; that is, to establish dependencies between various time-series sequences.

[0088] Feature extension module The final output of the layer is: , where LayerNorm is the layer normalization.

[0089] The balancing layer receives the final output of the feature augmentation module. (This fact has been summarized) =4), and then a linear transformation is performed through a fully connected layer:

[0090] in, , which are the model parameters of the balancing layer. This operation projects features to a higher-dimensional latent space, encouraging information to be fully mixed and compressed here, forming a robust representation of the input state.

[0091] Each layer of the feature extraction module ( Previous layer output (initial The input is ) and the corresponding layer (the first layer) of the feature augmentation module is integrated. Features of layers This involves information reconstruction and dimensionality reduction. Assuming both the feature augmentation module and the feature extraction module have N layers, the total number of layers in the network model is M = 2N + 2 layers. The nth layer of the feature extraction module... Layer (i.e., the first layer in layer M) The feature extension module of layer M corresponds to the M-th layer. layer.

[0092] Cross-attention submodule: This is the key to connecting encoded and decoded information.

[0093] The output comes from the feature augmentation module, that is, the output of the corresponding layer of the feature augmentation module. and Output from the feature extraction module, i.e., the layer above the feature extraction module:

[0094] in, , and is the projection matrix, which represents the model parameters of the cross-attention submodule. This cross-attention mechanism allows the feature extraction module to selectively retrieve relevant information from the multi-scale features extracted by the feature expansion module during the prediction process. By fusing the features output by the corresponding layer of the feature expansion module with the features from the layer above the feature extraction module, it simultaneously focuses on the low-level details of the features output by the feature expansion module and the high-level comprehensive information of the features output by the feature extraction module. In other words, the model influences the final prediction result by fusing low-level details and high-level comprehensive information during the prediction process.

[0095] Residual Connectivity Submodule: To ensure that features from the feature augmentation module can be losslessly transferred to the decoding process, an identity residual connectivity is introduced.

[0096] in, `<matrix>` is an optional linear projection matrix used to adjust the dimensions (when needed), and `<parameter>` represents the model parameters of the residual connective submodule. The residual connective submodule is designed to mitigate the vanishing gradient problem.

[0097] The fully connected transformation submodule first performs dimensionality reduction and linear transformation to make the feature dimension the same as the output feature of the corresponding layer in the feature augmentation module.

[0098] in, and These are the model parameters for the fully connected transformation submodule.

[0099] Output layer: The final output of the feature extraction module (That is, the output of the fully connected transformation submodule in the last layer of the feature extraction module) undergoes a simple linear transformation to obtain the final prediction:

[0100] in, , which are the model parameters of the output layer. This is the normalized brightness temperature prediction value.

[0101] The training process is described below.

[0102] Training configuration: Use the training dataset generated in the above steps for training. Set the batch size to batch_size = 1024 and the training epochs to 1000. Use a cosine annealing scheduler with an initial learning rate of... .

[0103] Loss function: Mean squared error loss is used to measure the predicted brightness temperature. (After denormalization) and GMI true brightness temperature Differences between them:

[0104] in This represents the number of samples within a batch.

[0105] Optimizer: Use the AdamW optimizer (β1 = 0.9, β2 = 0.999) and set the weight decay to... To prevent overfitting.

[0106] Training Process and Regularization: During training, the loss is monitored on independent validation sets. When the validation loss is continuous... Each round (e.g.) If the model does not decrease, an early stopping mechanism is initiated, and the model parameters are restored to the level with the lowest validation loss. Simultaneously, a recall rate (e.g., set to 0.1) is applied before the fully connected transformation submodules of the feature expansion and feature extraction modules to further improve the model's generalization ability.

[0107] Normalization and denormalization: During training, the sample input features and label features All values ​​have been normalized to [0,1]. Model output Inverse normalization (i.e., denormalization) is required to obtain the physical brightness temperature value: ,in The global maximum and minimum values ​​of the target brightness temperature for the training dataset.

[0108] Through the above design, the network model can fully explore the deep information and complex correlations in the 28-dimensional input features, achieve high-fidelity and high-precision simulation of the GMI reference brightness temperature, and provide a reliable data foundation for subsequent cross-calibration.

[0109] In some embodiments, the constructed hybrid attention multilayer perceptron model is used to invert the second calibration correlation data of the infrared instrument to be calibrated into the observation data of the reference infrared instrument, including: The second calibration-related data is preprocessed to obtain multi-channel input features; The preprocessed input features are fed into the constructed hybrid attention multilayer perceptron model to obtain normalized brightness temperature values; The normalized brightness temperature value output by the hybrid attention multilayer perceptron model is denormalized to restore it to a brightness temperature value with actual physical meaning and units.

[0110] In this embodiment, the second calibration-related data is new data from an infrared instrument to be calibrated that has not participated in training. Like the first calibration-related data, it includes the raw observation counts and instrument parameters output by the infrared instrument to be calibrated. To simulate the calibration reference source data corresponding to the second calibration-related data, the second calibration-related data can be preprocessed. This preprocessing includes extracting 28-dimensional data from the second calibration-related data, performing outlier checks and invalid value processing on the 28-dimensional data, and finally standardizing it to obtain 28-dimensional input features. These input features are also standardized values ​​obtained by normalizing the maximum and minimum values ​​of each dimension as mentioned above. After the 28-dimensional input features are input into the trained network model, a normalized brightness temperature prediction value is obtained. This brightness temperature prediction value, after inverse normalization (i.e., de-standardization), yields a simulated brightness temperature value, which has actual physical meaning and units.

[0111] The following example will still be used to illustrate the simulation of calibration reference source data (GMI brightness temperature) based on MWRI observation data.

[0112] Data preprocessing: For the new MWRI observation data to be simulated, extract the same 28-dimensional input data as the training data. Perform outlier checking and invalid value handling on the input data in exactly the same manner as described in the training process above.

[0113] Standardization: Using the minimum (X_min) and maximum (X_max) values ​​of each feature dimension calculated and saved from the training dataset in step 1, the new 28-dimensional input data is standardized and mapped to the interval [0, 1] to obtain 28-dimensional input features.

[0114] Model inference: The standardized 28-dimensional input features are input into the pre-trained and saved hybrid attention MLP model, and the model performs forward propagation calculations to obtain the model output.

[0115] Denormalization of Results: The normalized brightness temperature prediction value output by the model is denormalized using the minimum (Y_min) and maximum (Y_max) brightness temperatures (GMI brightness temperatures) saved during training, restoring them to brightness temperature values ​​with actual physical meaning and units (unit: Kelvin, K). The calculation formula is: Y = Y_pred * (Y_max - Y_min) + Y_min, where Y_pred is the normalized brightness temperature prediction value output by the model.

[0116] Output: Simulated GMI H polarization and 10.7 GHz channel brightness temperature data corresponding to the input MWRI instrument conditions. This data can be used as a high-precision calibration reference for cross-calibrating and correcting MWRI observations of this channel.

[0117] The following quantitative evaluation and comparative analysis of the accuracy and effectiveness of the model simulation results are conducted to verify the effectiveness of this disclosure.

[0118] Error statistics: On an independent test dataset, the brightness temperature data obtained from model inversion were compared with the actual GMI observed brightness temperature data. A series of statistical indicators were calculated, including: Bias: The average difference between the inverted value and the true value, reflecting systematic error.

[0119] Root mean square error (RMSE): The root mean square of the difference between the inverted value and the true value, which comprehensively reflects the overall level of error.

[0120] Standard deviation (STD): The standard deviation of an error series reflects the degree of dispersion of the error or the magnitude of random error.

[0121] Correlation coefficient (R): The Pearson correlation coefficient between the inverted value and the true value, reflecting the degree of linear correlation between the two.

[0122] Scatter plot and fitted line: Plot the scatter plot of the inverted brightness temperature and the true brightness temperature, and calculate its linear regression equation (slope, intercept). Ideally, the slope should be 1 and the intercept should be 0.

[0123] Benchmarking: To highlight the advantages of this disclosure, a benchmarking method is established. This method involves directly comparing the raw brightness temperature data (converting DN values ​​to brightness temperature using its internal scaling factors) of the MWRI target channel (H-polarization 10.65 GHz) with the GMI brightness temperature. The same statistical metrics (Bias, RMSE, STD, R, etc.) are calculated.

[0124] Accuracy Comparison: The error statistics (such as RMSE, Bias) of the inversion results of the model disclosed herein are directly compared with the error statistics of the benchmark method. The significant reduction in RMSE and Bias demonstrates that the model disclosed herein can effectively correct systematic biases and random errors in MWRI observations, and more closely approximate the high-precision GMI reference value.

[0125] Robustness analysis: The distribution of model error under different instrument temperature ranges, different geographical regions, and different observation times (such as annual variations) is analyzed. If the model error remains stable and small, it indicates that the method disclosed herein has good robustness and generalization ability, and does not depend on specific instrument conditions or environmental conditions.

[0126] Uncertainty assessment: The range of uncertainty in the model simulation results can be estimated by analyzing the distribution of errors on the test set (such as histograms) or by using model ensemble (training multiple models) methods.

[0127] Through systematic statistical and comparative analysis, quantitative data proves that the simulation method based on hybrid attention MLP proposed in this disclosure can generate calibration reference data with significantly higher accuracy than the original MWRI brightness temperature data and closer to the true value of GMI reference. This provides a reliable technical means and data product for high-precision on-orbit cross-calibration of FY-3D / MWRI.

[0128] The disclosed solution has the following characteristics: 1. High-precision simulation: By introducing a hybrid attention mechanism (including a combination of channel attention, self-attention, and cross attention), the model can adaptively focus on key instrument state parameters, capture the complex nonlinear interactions and time-series dependencies between parameters, and thus learn a deeper and more accurate mapping function from the original state of the infrared instrument to be calibrated to the calibration reference simulation data of the reference infrared instrument. The simulation accuracy is higher than that of existing linear or simple nonlinear statistical models.

[0129] 2. Strong feature fusion capability: The multi-dimensional input features constructed in this disclosure, combined with the channel attention mechanism, enable the model to intelligently fuse multi-source heterogeneous information from different subsystems of the instrument (optical path, receiver, antenna, calibration source), making full use of all available information for inference.

[0130] 3. Excellent generalization and robustness: The cross-attention and residual connections designed in the network structure ensure that the multi-scale features extracted in the encoding stage can be effectively utilized by the feature extraction module, which enhances the model's adaptability to different instrument states (such as temperature fluctuations) and observation conditions, reduces the risk of overfitting, and improves the model's generalization performance on unseen data.

[0131] 4. End-to-end automation: It realizes end-to-end modeling from raw observation data to calibration reference source data, avoiding complex manual feature engineering and physical model parameter debugging. It has a high degree of automation and is easy to integrate into business and promote to cross-calibration scenarios of other sensor pairs.

[0132] The following are embodiments of the apparatus disclosed herein, which can be used to execute embodiments of the method disclosed herein.

[0133] Figure 3 A structural block diagram of a calibration reference source data simulation apparatus according to an embodiment of the present disclosure is shown. This apparatus can be implemented as part or all of an electronic device through software, hardware, or a combination of both. Figure 3 As shown, the calibration reference source data simulation device includes: The preprocessing module 301 is configured to preprocess the first calibration-related data of the infrared instrument to be calibrated and the calibration reference source data of the reference infrared instrument to generate a model training dataset; the infrared instrument to be calibrated and the reference infrared instrument are infrared instruments mounted on different satellites respectively; the model training dataset includes multiple data matching pairs, and the data matching pairs include sample input features obtained from the first calibration-related data and their corresponding label features obtained from the calibration reference source data; The construction module 302 is configured to construct a hybrid attention multilayer perceptron model based on the model training dataset, and learn the end-to-end mapping relationship between the first calibration related data and the calibration reference source data in a data-driven manner. The inversion module 303 is configured to use the constructed hybrid attention multilayer perceptron model to invert the second calibration correlation data of the infrared instrument to be calibrated into the observation data of the reference infrared instrument.

[0134] In an optional implementation of this embodiment, the hybrid attention multilayer perceptron model combines channel attention, self-attention, and cross-attention mechanisms.

[0135] In an optional implementation of this embodiment, the hybrid attention multilayer perceptron model includes a feature augmentation module, a balancing layer, and a feature extraction module, wherein the feature augmentation module and the feature extraction module have a symmetrical structure; The feature augmentation module is used to perform linear transformation and feature augmentation on the sample input features, progressively increasing the feature dimension and generating multi-scale features; The balancing layer is used to map the output of the last layer of the feature augmentation module to a higher-dimensional latent feature vector; The feature extraction module takes the latent space vector as the initial input, and combines the multi-scale features output by the feature expansion module to perform dimensionality reduction and reconstruction layer by layer to generate dense features.

[0136] In an optional implementation of this embodiment, the feature augmentation module includes N layers, each layer including a channel attention submodule, a fully connected transformation submodule, and a self-attention submodule connected in sequence. The channel attention submodule learns the correlation between the current data matching pair and each channel in the sample input features through adaptive channel weighting. The fully connected transformation submodule is used to perform basic linear transformation and feature augmentation operations on the features weighted by the channel attention submodule; The self-attention submodule is used to treat the output of the fully connected transformation submodule as a sequence and model the dependencies between the sequences.

[0137] In an optional implementation of this embodiment, the feature extraction module includes N layers, each layer including a cross-attention submodule, a residual connection submodule, and a fully connected transformation submodule connected in sequence. The cross-attention submodule is used to selectively retrieve relevant information from the multi-scale features extracted by the feature expansion module, so as to simultaneously focus on the low-level detailed information output by the feature expansion module and the high-level comprehensive information output by the feature extraction module. The residual connection submodule is used to ensure that the multi-scale features obtained by the feature augmentation module can be losslessly transmitted to the decoding process of the feature extraction module; The fully connected submodule is used to perform dimensionality reduction and linear transformation on the output of the residual connected submodule to keep the feature dimension the same as the output of the corresponding layer of the feature expansion module.

[0138] In an optional implementation of this embodiment, the inversion module is implemented as follows: The second calibration-related data is preprocessed to obtain multi-channel input features; The preprocessed input features are fed into the constructed hybrid attention multilayer perceptron model to obtain normalized brightness temperature values; The normalized brightness temperature value output by the hybrid attention multilayer perceptron model is denormalized to restore it to a brightness temperature value with actual physical meaning and units.

[0139] The calibration reference source data simulation device in this embodiment corresponds to the calibration reference source data simulation method described above. For specific details, please refer to the calibration reference source data simulation method described above, which will not be repeated here.

[0140] Figure 4 This is a schematic diagram of the structure of an electronic device suitable for implementing a calibration reference source data simulation method according to an embodiment of the present disclosure.

[0141] like Figure 4 As shown, the electronic device 400 includes a processing unit 401, which can be implemented as a CPU, GPU, FPGA, NPU, or other processing unit. The processing unit 401 can execute various processes according to any of the methods described above in this disclosure, based on a program stored in the read-only memory (ROM) 402 or a program loaded from the storage portion 408 into the random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the electronic device 400. The processing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0142] The following components are connected to I / O interface 405: an input section 406 including a keyboard, mouse, etc.; an output section 407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN card, modem, etc. The communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from them can be installed into storage section 408 as needed.

[0143] In particular, according to embodiments of this disclosure, any of the methods described above in the embodiments of this disclosure can be implemented as a computer software program. For example, embodiments of this disclosure include a computer program product comprising a computer program tangibly embodied on a machine-readable medium, the computer program containing program code for performing any of the methods in the embodiments of this disclosure. In such an embodiment, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411.

[0144] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0145] The units or modules described in the embodiments of this disclosure can be implemented in software or hardware. The described units or modules can also be located in a processor, and the names of these units or modules do not necessarily constitute a limitation on the unit or module itself.

[0146] In another aspect, this disclosure also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the apparatus described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores one or more programs that are used by one or more processors to perform the methods described in this disclosure.

[0147] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

Claims

1. A method for simulating calibration reference source data, wherein, include: Preprocess the first calibration-related data of the infrared instrument to be calibrated and the calibration reference source data of the reference infrared instrument to generate the model training dataset; The infrared instrument to be calibrated and the reference infrared instrument are infrared instruments mounted on different satellites; the model training dataset includes multiple data matching pairs, and the data matching pairs include sample input features obtained from the first calibration related data and their corresponding label features obtained from the calibration reference source data; Based on the model training dataset, a hybrid attention multilayer perceptron model is constructed to learn the end-to-end mapping relationship between the first calibration-related data and the calibration reference source data in a data-driven manner. Using the constructed hybrid attention multilayer perceptron model, the second calibration correlation data of the infrared instrument to be calibrated is inverted into the observation data of the reference infrared instrument.

2. The method according to claim 1, wherein, The hybrid attention multilayer perceptron model combines channel attention, self-attention, and cross-attention mechanisms.

3. The method according to claim 1, wherein, The hybrid attention multilayer perceptron model includes a feature augmentation module, a balancing layer, and a feature extraction module, wherein the feature augmentation module and the feature extraction module have a symmetrical structure. The feature augmentation module is used to perform linear transformation and feature augmentation on the sample input features, progressively increasing the feature dimension and generating multi-scale features; The balancing layer is used to map the output of the last layer of the feature augmentation module to a higher-dimensional latent feature vector; The feature extraction module takes the latent space vector as the initial input, and combines the multi-scale features output by the feature expansion module to perform dimensionality reduction and reconstruction layer by layer to generate dense features.

4. The method according to claim 3, wherein, The feature augmentation module comprises N layers, each layer including a channel attention submodule, a fully connected transformation submodule, and a self-attention submodule connected in sequence. The channel attention submodule learns the correlation between the current data matching pair and each channel in the sample input features through adaptive channel weighting. The fully connected transformation submodule is used to perform basic linear transformation and feature augmentation operations on the features weighted by the channel attention submodule; The self-attention submodule is used to treat the output of the fully connected transformation submodule as a sequence and model the dependencies between the sequences.

5. The method according to claim 3, wherein, The feature extraction module includes N layers, each layer including sequentially connected cross-attention submodule, residual connection submodule and fully connected transformation submodule; The cross-attention submodule is used to selectively retrieve relevant information from the multi-scale features extracted by the feature expansion module, so as to simultaneously focus on the low-level detailed information output by the feature expansion module and the high-level comprehensive information output by the feature extraction module. The residual connection submodule is used to ensure that the multi-scale features obtained by the feature augmentation module can be losslessly transmitted to the decoding process of the feature extraction module; The fully connected submodule is used to perform dimensionality reduction and linear transformation on the output of the residual connected submodule to keep the feature dimension the same as the output of the corresponding layer of the feature expansion module.

6. The method according to claim 1, wherein, Using the constructed hybrid attention multilayer perceptron model, the second calibration correlation data of the infrared instrument to be calibrated is inverted into the observation data of the reference infrared instrument, including: The second calibration-related data is preprocessed to obtain multi-channel input features; The preprocessed input features are fed into the constructed hybrid attention multilayer perceptron model to obtain normalized brightness temperature values; The normalized brightness temperature value output by the hybrid attention multilayer perceptron model is denormalized to restore it to a brightness temperature value with actual physical meaning and units.

7. A calibration reference source data simulation device, wherein, include: The preprocessing module is configured to preprocess the first calibration-related data of the infrared instrument to be calibrated and the calibration reference source data of the reference infrared instrument to generate a model training dataset; the infrared instrument to be calibrated and the reference infrared instrument are infrared instruments mounted on different satellites respectively; the model training dataset includes multiple data matching pairs, and the data matching pairs include sample input features obtained from the first calibration-related data and their corresponding label features obtained from the calibration reference source data; The building module is configured to build a hybrid attention multilayer perceptron model based on the model training dataset, and learn the end-to-end mapping relationship between the first calibration-related data and the calibration reference source data in a data-driven manner. The inversion module is configured to use the constructed hybrid attention multilayer perceptron model to invert the second calibration correlation data of the infrared instrument to be calibrated into the observation data of the reference infrared instrument.

8. An electronic device, wherein, The method includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method of any one of claims 1-6.

9. A computer-readable storage medium having computer instructions stored thereon, wherein, When executed by a processor, the computer instructions implement the method described in any one of claims 1-6.

10. A computer program product comprising computer instructions, wherein, When executed by a processor, the computer instructions implement the method described in any one of claims 1-6.

Images