Satellite infrared multi-probe metadata homogenization method based on variational self-coding

Abstract

The invention discloses a satellite infrared multi-probe metadata homogenization method based on variational self-encoding, which comprises the following steps of: acquiring multichannel observation brightness temperature data of a satellite large-array infrared detector, constructing a variational self-encoder model, designing a comprehensive loss function, adopting a dynamic weight scheduling mechanism, gradually adjusting the weight of each loss item in a training process, and carrying out uniformization on the multi-channel observation brightness temperature data of the satellite large-array infrared detector. And carrying out iterative training by adopting Adam and a piecewise learning rate, storing an optimal model, reconstructing observation data by utilizing the trained model, and outputting homogenized multi-detection-element hyperspectral brightness temperature data. The method can automatically extract and learn the complex nonlinear characteristics of the hyperspectral data, has higher accuracy and robustness compared with an empirical correction method, and improves the consistency and stability of the data of the large-array detector; and the method has relatively high expandability and universality.

Description

Technical Field

[0001] This invention relates to the field of meteorological satellite remote sensing and data processing technology, specifically to a method for homogenizing satellite infrared multi-sensor data based on variational autoencoding. Background Technology

[0002] Satellite hyperspectral infrared (HIR) instruments have important applications in meteorological monitoring and numerical weather prediction, typically including typhoon monitoring, climate research, and data assimilation. However, for example, the Geostationary Interferometric Infrared Sounder (GIIRS) aboard the Fengyun-4B satellite is composed of a large-scale array of infrared detectors. Systematic biases and response inhomogeneities often exist between different detectors. These differences can lead to inconsistencies in the data acquired for the same observation scene, thereby reducing overall observation accuracy and affecting the reliability and application value of the data.

[0003] In existing satellite data processing technologies, common detector homogenization methods mainly rely on empirical correction and detector-by-detector statistical averaging. While these methods can reduce inter-detector differences to some extent, they still have limitations: on the one hand, they are insufficient in characterizing the complex nonlinear relationships of high-dimensional spectral features; on the other hand, these methods are poorly adaptable to complex observation environments (such as atmospheric changes and observational geometric differences), making it difficult to simultaneously guarantee reconstruction accuracy and inter-detector consistency. This restricts the application of hyperspectral infrared observation data in scenarios such as high-precision data assimilation and extreme weather monitoring. Therefore, a satellite infrared multi-detector data homogenization method based on variational autoencoding is needed. Summary of the Invention

[0004] The purpose of this invention is to propose a satellite infrared multi-detector data homogenization method based on variational autoencoder to solve the problems of existing technologies. This method can fully exploit the multi-channel information of hyperspectral satellite observations, effectively suppress systematic biases between detectors through adaptive learning, and enhance the consistency of multi-detector data while ensuring reconstruction accuracy, thereby better supporting atmospheric sounding and numerical forecasting applications.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention includes the following steps: 1) Acquire multi-channel brightness temperature observation data and corresponding radiative transfer simulation background data from the satellite large array infrared detector, and preprocess the observation data; 2) Construct a variational autoencoder model, which includes an encoder, a reparameterization layer, and a decoder. The encoder compresses the preprocessed high-dimensional spectral data into the latent space, and the decoder reconstructs the latent variables to recover the brightness temperature data. 3) Define the comprehensive loss function, which includes a channel weighted mean square error term, a KL divergence constraint term, and a probe residual standard deviation constraint term; 4) A dynamic weight scheduling mechanism is adopted to adjust the weights of each loss term in the comprehensive loss function during model training; 5) The variational autoencoder model is iteratively trained using a preset optimizer and a piecewise learning rate, and the optimal model is saved; 6) The multi-channel brightness temperature observation data are reconstructed using the optimal model, and after inverse transformation, uniformized multi-element hyperspectral brightness temperature data are output.

[0006] Furthermore, the preprocessing in step 1) includes spectral channel clipping, physical quantity conversion, standardization, and the division of the training set and validation set.

[0007] Furthermore, the standardization process constructs a standardizer by calculating the mean and standard deviation of each channel in the training set, and the observed data and background data are processed using the same standardizer.

[0008] Further, in step 2), the encoder outputs the mean and logarithmic variance of the latent variables, the reparameterization layer obtains the latent variables based on standard normal noise sampling, and the encoder and decoder use the ReLU activation function.

[0009] Further, in step 3), the channel weighted mean square error term assigns a first weight to the critical channel for business assimilation and a second weight between 0 and 1 to the non-critical channel.

[0010] Furthermore, the dynamic weight scheduling mechanism in step 4) includes: the weight of the KL divergence constraint term is linearly increased to the maximum value, and the weight of the probe residual standard deviation constraint term and the weight of the non-critical channel are exponentially decayed to a preset lower limit through piecewise exponential decay.

[0011] Furthermore, in step 5), the preset optimizer is the Adam optimizer, and the zero-point score based on the absolute median of the business assimilation channel residuals on the validation set is used as the optimal model selection index.

[0012] Furthermore, the iterative training employs an early stopping mechanism, stopping training when the score remains unchanged after a preset number of consecutive zero-score epochs.

[0013] Furthermore, the inverse transformation process described in step 6) restores the reconstructed data to the brightness temperature physical domain using the mean and standard deviation of the normalizer.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This method introduces a deep generative model, which can automatically extract and learn the complex nonlinear features of hyperspectral data, making it more accurate and robust than empirical correction methods. While maintaining the accuracy of brightness temperature reconstruction, it effectively reduces the systematic bias between detectors, improving the consistency and stability of data from large-array detectors. The method has strong scalability and versatility, and can be applied to other types of satellite hyperspectral detectors, providing higher-quality observational data for numerical weather prediction, climate research, and extreme weather monitoring. Attached Figure Description

[0015] Figure 1 This is a flowchart of a satellite infrared multi-detector data homogenization method based on variational autoencoder provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the variational autoencoder (VAE) model structure provided according to an embodiment of the present invention; Figure 3 This is a comparison curve of O–B average values ​​by channel provided according to an embodiment of the present invention; Figure 4 This is a comparison curve of the standard deviation of the O-B residuals of the probes by channel, provided according to an embodiment of the present invention. Detailed Implementation

[0016] Reference Figure 1 The present invention includes the following steps: 1) Acquire multi-channel brightness temperature observation data and corresponding radiative transfer simulation background data from the satellite large array infrared detector, and preprocess the observation data; 2) Construct a variational autoencoder model, which includes an encoder, a reparameterization layer, and a decoder. The encoder compresses the preprocessed high-dimensional spectral data into the latent space, and the decoder reconstructs the latent variables to recover the brightness temperature data. 3) Define the comprehensive loss function, which includes a channel weighted mean square error term, a KL divergence constraint term, and a probe residual standard deviation constraint term; 4) A dynamic weight scheduling mechanism is adopted to adjust the weights of each loss term in the comprehensive loss function during model training; 5) The variational autoencoder model is iteratively trained using a preset optimizer and a piecewise learning rate, and the optimal model is saved; 6) The multi-channel brightness temperature observation data are reconstructed using the optimal model, and after inverse transformation, uniformized multi-element hyperspectral brightness temperature data are output.

[0017] The “GIIRS observation” refers to the brightness temperature observation of the FY-4BGIIRS long-wave channel; “background” refers to the brightness temperature simulated by the Fast Radiative Transfer for TOVS (RTTOV) model based on the fifth-generation atmospheric reanalysis data (ECMWFReanalysisv5, ERA5) from the European Centre for Medium-Range Weather Forecasts (ECMWF); and “detectors” refers to the 128 detector units of the GIIRS fixed array.

[0018] The method for uniformity of multi-detector data of satellite large-array infrared detectors based on variational autoencoders provided in this embodiment of the invention refers to... Figure 2 The schematic diagram of the model structure is obtained by performing the following steps S1 to S6.

[0019] Step S1: Data Acquisition and Matching. Based on time and field-of-view matching, acquire GIIRS long-wavelength channel observations and corresponding spatiotemporal simulated brightness temperatures; simultaneously extract the detector numbers (1-128) of the samples. After trimming the original long-wavelength channel spectral ranges to the operational assimilation effective range, channels 33-278 are retained, forming a total of 246-dimensional spectral vectors. Channels 1-66 are the operational assimilation key channels (denoted as the key channel set C). 66 Channels 67 to 246 are non-critical channels, with a total of 10,000 samples. Each sample contains information including (246-dimensional observation brightness temperature vector o, 246-dimensional background brightness temperature vector b, and corresponding probe number d).

[0020] The specific steps of step S1 are as follows: Step S1.1: Spectral channel trimming and physical quantity homogenization. The GIIRS long-wavelength channels are trimmed to 33–278 channels according to the effective range of service assimilation, forming a 246-dimensional spectral vector; the input radiance values ​​are converted into brightness temperature according to Planck's formula.

[0021] Step S1.2: Construct training samples. Associate the cropped and transformed 246-dimensional observed brightness temperature vector *o* and 246-dimensional background brightness temperature vector *b* with their corresponding detector numbers *d* to form complete training sample units; where samples 1-78 correspond to detector 1, samples 79-156 correspond to detector 2, ..., samples 9926-10000 correspond to detector 128, ensuring that the samples for each detector are evenly distributed in the training set, providing sufficient data support for the model to learn detector characteristics. Let the observed brightness temperature vector be denoted as... The background brightness temperature vector is denoted as , records the index set of 66 business assimilation channels and probe number .

[0022] Step S1.3: Standardization. Calculate the mean of the training set by channel. with standard deviation ( ), forming a standardizer For any sample Its standardized vector is defined as: in Small constant ( The normalizer fits the data using only the training set and remains unchanged throughout the training, validation, and inference phases; the reconstructed output is obtained through inverse transformation. The data were restored to the brightness temperature domain. The same normalizer was used for both observational and background data. To ensure the comparability of indicators.

[0023] Step S1.4: Data partitioning. All samples are randomly shuffled under a fixed random seed to ensure reproducibility of results. Then, the data is partitioned into training and validation sets in a 9:1 ratio. The channel mean and standard deviation vectors, key channel masks, and channel range information used for standardization are persistently saved for subsequent training, validation, and result reproduction experiments.

[0024] Step S2: Construct a variational autoencoder (VAE). The input is the normalized spectral vector. VAE is composed of encoders. Reparameterized sampling and decoder Composition, latent dimensions .

[0025] The specific steps of step S2 are as follows: (1) Encoder: Maps the 246-dimensional spectral data input to the mean and log-variance of the latent variables: , ,

[0026] in The activation function used is ReLU. get .

[0027] (2) Reparameterized sampling: using standard normal noise : ,

[0028] (3) Decoder: Reconstructs the latent variables into 246-dimensional spectral data output: ,

[0029] in , From a probabilistic modeling perspective, the conditional likelihood can be set as an isotropic Gaussian: ,in It is a constant.

[0030] Step S3: Define the comprehensive loss function This involves simultaneously constraining reconstruction accuracy and detector uniformity within a standardized space. Let a training mini-batch be... ,sample The probe number is Input is Reconstructed into The key channel set is as follows .

[0031] The specific details of step S3 are as follows: (1) Channel-weighted MSE (precision term): To highlight key channels, weights are assigned to each channel. : Among them, for If it belongs to The weight is 1 if it is assigned, otherwise a smaller weight between 0 and 1 is assigned. .

[0032] Key channel average The weighted error of the 66 key channels is approximately 0.0004, and the sum of the errors is 66 × 1 × 0.0004 = 0.0264; the average error of the non-key channels is approximately 0.0264. The weighted error of 180 non-critical channels is approximately 0.0009, which is 180 × 0.3 × 0.0009 = 0.0486. The weighted sum of errors for a single sample is 0.0264 + 0.0486 = 0.075; for a batch... ≈9.5×10 −6 .

[0033] (2) KL divergence (latent regularization term): for the dimensions of latent variables Let the mean and standard deviation of the encoder output be denoted as . , : The absolute value of sample μn,i is <0.2, σn,i is close to 1, and Lkl,n ≈ 0.035 for a single sample; batch =0.035.

[0034] (3) Probe STD constraint (homogeneity term): First define the sample-channel residual Calculate the sample-level average for key channels. Aggregate samples from the same probe: Its probe average residual The collection of probes that appeared only in this batch The dispersion across the probe element is calculated above, denoted as: ,

[0035] The above uniformity terms are used to suppress systematic differences between different probes.

[0036] Calculate the average residual of the key channels in the sample; this batch involves 8 probes, and the samples are grouped according to the probe number to calculate the average residual of each probe. (respectively 0.032, 0.024, 0.018, 0.005, -0.002, -0.015, -0.022, -0.029); Then calculate the average value across the probe element. =1 / 8×(0.032+0.024+0.018+0.005−0.002−0.015−0.022−0.029)≈0.0014. ≈0.021.

[0037] (4) The total loss function is expressed as:

[0038] Step S4: Employ a weighted scheduling mechanism to balance the contributions of the three losses. To balance the contributions of the three losses during training, let the total training epochs be... The current epoch is Define the weight sequence. The total loss is approximately L ≈ 0.1225.

[0039] The specific operation of step S4 is as follows: (1) warm-up (latent regularization gradually strengthens): in the preceding The epoch linearly rises to Then keep: ; At epoch 0, β=0; at epoch 10, β=0.5; and at epoch 20 and thereafter, β=1.0.

[0040] (2) Attenuation (gradual relaxation of uniformity term): Piecewise exponential decay is used, with a step size of... attenuation rate And set a lower limit : ; λ = 10.0 for epochs 0-19, λ = 5.0 for epochs 20-39, λ = 2.5 for epochs 40-59, and λ = 1.0 for epochs 80 and beyond.

[0041] (3) Attenuation (weight reduction of non-critical channels): Similarly, set the step size. Attenuation rate Lower limit : ; ε = 0.5 from epoch 0 to 24, ε = 0.3 from epoch 25 to 49, ε = 0.18 from epoch 50 to 74, and ε = 0.1 from epoch 100 onwards.

[0042] in Starting from small to large helps stabilize the learning potential space; , Starting from large to small can avoid premature and excessive uniformity constraints and interference from non-critical channels.

[0043] Step S5: Iterative training is performed using the Adam optimizer and multiple learning rates. At the end of each epoch, the optimal model is selected by setting zero scores on the validation set, and early stopping can be used in conjunction with this.

[0044] The specific operation of step S5 is as follows: (1) The Adam optimizer is used with an initial learning rate of Weight decay is Under a fixed random seed, the samples are randomly shuffled and divided into training and validation sets in a 9:1 ratio. DataLoaders are then built to perform iterative training and parameter updates in batches.

[0045] (2) Validation metrics (zero score): Reconstruction will be performed only on 66 assimilation channels. With background Denormalize to brightness temperature using the StandardScaler of the training set, and calculate... The sample mean is used to define the residuals (standardized space). , Then take the absolute median of the channels as the score: If the score refreshes the minimum value in each epoch, then the current weights are saved as the optimal model; otherwise, the score remains unchanged. If the individual epochs are not improved, the project will be discontinued sooner rather than later.

[0046] Step S6: Reconstruction and Evaluation Step S6.1: Reconstruct the test samples using the optimal model to obtain... Then, the brightness temperature is restored through inverse normalization transform, generating a uniform brightness temperature product while preserving the optimal model weights and normalizer parameters. To support the reproduction of experiments and subsequent batch processing.

[0047] Step S6.2: Calculate the O-B mean and standard deviation for each channel based on the probe element, generating the output for the embodiment. Figure 3 and Figure 4 The statistical comparison charts of the mean O-B values ​​before and after homogenization for each channel and the statistical comparison chart of the standard deviation of the O-B residuals of the probes shown support the beneficial effect of this invention in significantly reducing the differences between probes while maintaining reconstruction accuracy.

[0048] This invention is based on the VAE model to automatically learn the nonlinear characteristics of hyperspectral data. By combining dynamic weight scheduling and comprehensive loss constraints, it can effectively suppress systematic biases between detectors and improve the uniformity of multi-detector data while maintaining the accuracy of spectral reconstruction. It can provide high-quality observation data for applications such as numerical weather prediction and climate research. It is also scalable and can be applied to other satellite hyperspectral detectors.

[0049] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for homogenizing satellite infrared multi-detector data based on variational autoencoder, characterized in that, Includes the following steps: 1) Acquire multi-channel brightness temperature observation data and corresponding radiative transfer simulation background data from the satellite large array infrared detector, and preprocess the observation data; 2) Construct a variational autoencoder model, which includes an encoder, a reparameterization layer, and a decoder. The encoder compresses the preprocessed high-dimensional spectral data into the latent space, and the decoder reconstructs the latent variables to recover the brightness temperature data. 3) Define the comprehensive loss function, which includes a channel weighted mean square error term, a KL divergence constraint term, and a probe residual standard deviation constraint term; 4) A dynamic weight scheduling mechanism is adopted to adjust the weights of each loss term in the comprehensive loss function during model training; 5) The variational autoencoder model is iteratively trained using a preset optimizer and a piecewise learning rate, and the optimal model is saved; 6) The multi-channel brightness temperature observation data are reconstructed using the optimal model, and after inverse transformation, uniformized multi-element hyperspectral brightness temperature data are output.

2. The satellite infrared multi-detector data homogenization method based on variational autoencoder according to claim 1, characterized in that, Step 1) The preprocessing includes spectral channel clipping, physical quantity conversion, standardization, and the division of the training set and validation set.

3. The satellite infrared multi-detector data homogenization method based on variational autoencoder according to claim 2, characterized in that, The standardization process constructs a standardizer by calculating the mean and standard deviation of each channel in the training set, and the observed data and background data are processed using the same standardizer.

4. The satellite infrared multi-detector data homogenization method based on variational autoencoder according to claim 1, characterized in that, In step 2), the encoder outputs the mean and logarithmic variance of the latent variables, the reparameterization layer obtains the latent variables based on standard normal noise sampling, and the encoder and decoder use the ReLU activation function.

5. The satellite infrared multi-detector data homogenization method based on variational autoencoder according to claim 1, characterized in that, In step 3), the channel weighted mean square error term assigns a first weight to the critical channels of business assimilation and a second weight between 0 and 1 to the non-critical channels.

6. The satellite infrared multi-detector data homogenization method based on variational autoencoder according to claim 1, characterized in that, Step 4) The dynamic weight scheduling mechanism includes: the weight of the KL divergence constraint term is linearly increased to the maximum value, and the weight of the probe residual standard deviation constraint term and the weight of the non-critical channel are exponentially decayed to a preset lower limit through piecewise exponential decay.

7. The satellite infrared multi-detector data homogenization method based on variational autoencoder according to claim 1, characterized in that, Step 5) The preset optimizer is the Adam optimizer, and the zero-tap score based on the absolute median of the business assimilation channel residuals on the validation set is used as the optimal model selection index.

8. The satellite infrared multi-detector data homogenization method based on variational autoencoder according to claim 7, characterized in that, The iterative training employs an early stopping mechanism, stopping training when the score remains unchanged after a preset number of consecutive zero-score epochs.

9. The satellite infrared multi-detector data homogenization method based on variational autoencoder according to claim 1, characterized in that, Step 6) The inverse transformation process restores the reconstructed data to the brightness temperature physical domain by using the mean and standard deviation of the normalizer.

Images