A multi-temporal radiation signal inversion method based on conditional Transformer

Abstract

This invention relates to a multi-temporal radiation signal inversion method based on Conditional Transformer, belonging to the field of satellite remote sensing data processing / spatial information science and technology. The method includes: acquiring meteorological time-series data of the target area to establish a dataset; designing a Conditional Transformer deep learning network; combining the target time-series data and meteorological data encoded by time and meteorological code generators with the output of the Encoder layer, and inputting this combination into the Decoder layer to train the Conditional Transformer deep learning network to obtain a surface radiation signal inversion model; inverting surface radiation brightness temperature data and evaluating the inversion accuracy; efficiently capturing long-distance spatiotemporal dependencies of radiation signals; flexibly inverting radiation changes according to the target time and meteorological conditions; removing nonlinear interference from the atmosphere on radiation; and more accurately inverting the true radiation characteristics of the surface.

Description

Technical Field

[0001] This invention relates to the field of satellite remote sensing data processing / space information science and technology, specifically to a multi-temporal radiometric signal inversion method based on conditional Transformer. Background Technology

[0002] The Moderate Resolution Imaging Spectroradiometer (MODIS) offers advantages such as global coverage and multispectral comprehensive observation, making it a crucial data source for global climate change research and ecological environment monitoring. Taking the MODIS sensor aboard the Terra and Aqua satellites as an example, it can acquire 36 channels of radiation signals covering the visible to thermal infrared bands. Accurately modeling the variation patterns of these time-series radiation signals and gaining a deeper understanding of the evolution trends of surface and atmospheric radiation over time is of significant scientific importance and practical value for compensating for the limitations of temporal resolution in polar-orbiting satellites, improving the accuracy of surface parameter inversion, and facilitating long-term environmental monitoring.

[0003] Traditional radiation signal simulations typically rely on existing radiative transfer models (MODTRAN, 6SV, etc.). While these models are based on rigorous physical equations and have clear physical meanings, their computational processes are extremely complex and computationally intensive, making it difficult to meet the demands of real-time processing of massive amounts of satellite data. Furthermore, physical models heavily depend on accurate atmospheric profiles and surface parameters as input, which are often difficult to obtain in real-time and with high precision in actual operational scenarios, leading to significant uncertainties in the simulation results.

[0004] In recent years, with the development of deep learning technology, deep learning networks, represented by Convolutional Neural Networks (CNN) and Recurrent Neural Networks (LSTM), have been widely used in remote sensing image inversion tasks. However, existing technologies have significant limitations. First, CNNs mainly focus on local features of images, making it difficult to capture long-distance spatial dependencies between distant pixels in large-scale weather systems. Second, recurrent neural networks such as LSTM require sequential computation at time steps when processing time-series data, resulting in low training efficiency and a tendency for gradient vanishing, making it difficult to effectively capture long-period nonlinear variation patterns. Third, most current models can only perform deterministic inversion of the next frame, lacking the ability to dynamically adjust the inversion results based on a specified target time or specific meteorological conditions, thus failing to meet the needs of complex and ever-changing meteorological simulation and extrapolation. Summary of the Invention

[0005] The purpose of this invention is to address the problems of low long-distance spatiotemporal dependence, inflexible prediction of radiation changes, and difficulty in solving the aliasing of radiation signal sources in current inversion methods. This invention provides a multi-time radiation signal inversion method based on conditional Transformer, which can efficiently capture the long-distance spatiotemporal dependence of radiation signals and flexibly invert radiation changes according to the target time and meteorological conditions. It can also remove the nonlinear interference of the atmosphere on radiation and more accurately invert the true radiation characteristics of the Earth's surface.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a multi-time radiation signal inversion method based on conditional Transformer, comprising the following steps: S1. Obtain meteorological time series data of the target area, and establish a dataset after preprocessing; S2. Design a conditional Transformer deep learning network, including an Encoder layer and a Decoder layer, as well as a time coding generator and a meteorological coding generator for encoding the target time data and meteorological data in the dataset; Specifically, the target time data and meteorological data encoded by the time code generator and the meteorological code generator are combined with the output of the Encoder layer and input into the Decoder layer to train the conditional Transformer deep learning network to obtain the surface radiation signal inversion model. S3. Based on the surface radiation signal inversion model obtained in step S2, invert the surface radiation brightness temperature to obtain surface radiation brightness temperature data, and evaluate the inversion accuracy.

[0007] Preferably, the meteorological time series data includes: Multi-time radiation signals, including radiation brightness temperature data in the thermal infrared band; Meteorological data, including temperature profiles, specific humidity profiles, ozone mass mixture ratios, surface temperature, and surface air pressure; Target time data.

[0008] Preferably, the preprocessing includes: The DN values ​​recorded by the sensor are converted into satellite-observed radiation values ​​using a calibration formula, and the radiation brightness temperature data are obtained using the inverse function of Planck's blackbody radiation law. Based on the time and latitude / longitude coordinates of each data acquisition, data from the same geographical location at the same time are matched, using image pixels as the unit. Pixels with clouds were removed, and the interquartile range was used to determine the range of normal data, thus eliminating residual noise.

[0009] Preferably, the time encoding generator uses an attention mechanism and residual structure to map discrete observation times into orthogonal target time feature vectors; the meteorological encoding generator uses a fully connected neural network to perform nonlinear feature extraction on multidimensional meteorological parameters to obtain meteorological data feature vectors.

[0010] Preferably, the encoder layer includes six stacked encoder layers with identical structures. Each encoder layer includes a multi-head attention mechanism module and a position-based feedforward neural network module. After the output of each encoder layer, residual connections and layer normalization are introduced for processing. The decoder layer comprises six stacked decoder layers with identical structures, each of which includes: A multi-head attention mechanism module with masks, including padding masks and sequence masks; The encoder-decoder attention mechanism module, and the query comes from the output of the previous decoder layer; Feedforward neural network module.

[0011] Preferably, the output of the Decoder layer is mapped through a linear layer, and the probability is calculated using the Softmax activation function.

[0012] Preferably, the training of the conditional Transformer deep learning network includes: using the radiation signal at the target time as a label, the surface radiation signal at the input time, the input meteorological data, and the data at the target time as features, and feeding them into the conditional Transformer deep learning network for training to obtain a surface radiation signal inversion model.

[0013] Preferably, during the training of the Conditional Transformer deep learning network, mean squared error is used as the loss function, and the Adam optimizer is used for parameter updates; and the feedforward neural networks of the encoder and decoder in the original Conditional Transformer deep learning network are replaced with convolutional neural networks.

[0014] Preferably, the surface radiation signal inversion model for inverting surface radiation brightness temperature specifically includes: The target time data to be inverted and the corresponding meteorological data are fed into the corresponding encoding generator to obtain the target time feature vector and the meteorological data feature vector; Keep the context feature matrix extracted by the Encoder layer from the historical time-time radiation signals unchanged, and use the target time feature vector and the meteorological data feature vector as conditional constraints input to the Decoder layer; The Decoder layer decodes the context feature matrix and outputs the surface radiation brightness temperature data at the corresponding time. The surface radiation brightness temperature data obtained by inversion is compared with the actual surface radiation brightness temperature data, and the accuracy is evaluated by average error deviation and symmetrical average absolute percentage error.

[0015] Beneficial effects: This invention obtains a surface radiation signal inversion model by designing and training a conditional Transformer deep learning network. It can efficiently capture the long-distance spatiotemporal dependence of radiation signals and flexibly invert radiation changes according to the target time and meteorological conditions. It removes the nonlinear interference of the atmosphere on radiation and more accurately inverts the true radiation characteristics of the surface, significantly improving the inversion accuracy. In particular, the introduction of a conditional generation mechanism embeds the encoded time series and meteorological data as guiding conditions into the model, so that the model is not limited to the inversion of the next fixed frame, but can accurately and directionally invert the surface radiation signal at any target time by combining the specified time series and meteorological data. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multi-temporal radiation signal inversion method based on conditional Transformer of the present invention; Figure 2 This is a flowchart of the time encoding method of the present invention; Figure 3 This is a flowchart of the meteorological coding method of the present invention. Detailed Implementation

[0017] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0018] Example, reference Figure 1 As shown, a multi-temporal radiation signal inversion method based on conditional Transformer includes the following steps: Step 1: Obtain meteorological time series data for the target area from MODIS data, including: Multi-time radiation signals: MODIS02 calibrated radiation data (data obtained after calibration in the thermal infrared band from satellite observations); Meteorological data, ERA5 atmospheric reanalysis data, including temperature profile, specific humidity profile, ozone mass mixture ratio, surface temperature and surface pressure; Target time data: Feature data based on satellite observation time; Step 2: Preprocess the data and create a dataset: (1) The DN values ​​recorded by the sensor are converted into satellite-observed radiometric values ​​using a calibration formula: ; In the formula, L is the calibrated spectral radiance, DN is the digital count value of the original satellite observation, Gain is the gain coefficient, and Offset is the intercept coefficient; The radiation brightness temperature data were obtained using the inverse function of Planck's blackbody radiation law: ; In the formula, T b The target is the surface radiation brightness temperature. The center wavelength of the sensor, , is the radiation constant.

[0019] (2) Spatiotemporal registration: Based on the time and latitude and longitude coordinates of each data acquisition, the data of the same geographical location at the same time are matched in units of image pixels; for the problem of low spatial resolution of ERA5 data, bilinear interpolation is used to achieve spatial alignment of pixels; for the problem of high temporal resolution of Fengyun satellite, representative images are selected and matched with ERA5 time. (3) Conditional coding: The time coding generator uses the attention mechanism and residual structure to map discrete observation times into orthogonal target time feature vectors; the meteorological coding generator uses a fully connected neural network to perform nonlinear feature extraction on multidimensional meteorological parameters to obtain meteorological data feature vectors; (4) Outlier removal: Using MODIS35 cloud mask data, distinguish between cloudy and cloudless pixels and remove cloudy pixels; use the interquartile range (IQR) method to determine the range of normal data and remove residual noise.

[0020] Step 3: Design a conditional Transformer deep learning network, including an Encoder layer, a Decoder layer, a time encoding generator, and a meteorological encoding generator. Combine the target time data and meteorological data encoded by the time encoding generator and the meteorological encoding generator with the output of the Encoder layer, and input this combination into the Decoder layer to train the conditional Transformer deep learning network, thereby obtaining the surface radiation signal inversion model. The encoder layer consists of six stacked, identical encoder layers. Each encoder layer includes two main sub-layers: a multi-head attention mechanism module and a position-based feedforward neural network module. During self-attention computation, the input vector (e.g., ...) is... and The query is mapped to three distinct feature space vectors: query, key, and value. This process is achieved through three learnable weight matrices. , , Multiplication is implemented. For any input vector... The conversion formula is as follows: ; ; ; Where q represents the query feature of the current input sequence, used to calculate the similarity with the key vectors at all positions to match the target information; k represents the key feature of the input sequence, used to calculate the similarity with the query vector; v represents the value feature of the input sequence, which is combined with the attention weights to perform a weighted summation to obtain the final attention output.

[0021] Subsequently, attention weights were used to... The vectors are weighted and summed to generate the transformed output vector (e.g., ...). , ).by For example, its calculation is as follows and linear combination ; in, , This represents the value vector obtained by mapping the input vector. , Indicates corresponding to respectively , Attention weight coefficients.

[0022] To accelerate convergence and alleviate gradient problems, the Transformer introduces residual connections and layer normalization (LN) after the output of each sublayer (including the self-attention module and the feedforward network module). The core objective of normalization is to adjust the input feature distribution to a standard normal distribution with a mean of 0 and a variance of 1. The LN operation is performed independently within a single sample, and its mathematical expression is: ; in, and These represent the sample mean and variance, respectively. To prevent tiny amounts of division by zero, and These are learnable affine transformation parameters; The decoder layer also consists of six stacked decoder layers with identical structures, each of which includes: The multi-head attention mechanism module with masking prevents data leakage by introducing a mask to shield information from future moments, ensuring that the model relies only on known historical information during inversion; specifically, it includes two forms: padding mask and sequence mask. The encoder-decoder attention mechanism module, in which the query comes from the output of the previous decoder layer. The key and value are derived from the encoder's output features. This interaction mechanism enables the transmission of encoded information to the decoding end, and its calculation logic is as follows: ; ; ; The output of the Decoder layer is mapped through a linear layer, and the probability is calculated using the Softmax activation function. refer to Figure 2 As shown, the Time Encoding Generator (TEG) is used to generate the target time-time encoding. The TEG consists of an attention mechanism, residual connections, and a BN layer. The attention mechanism allows the model to dynamically focus on different parts of the input when processing sequential or ensemble data, enabling the network to allocate different attention weights in different contexts. The residual connections and BN layer perform normalization operations on the data, thereby achieving the encoding of the target time. refer to Figure 3 As shown, the Climatic Encoding Generator (CEG) is used to generate the target time meteorological data encoding. CEG consists of a multi-layer fully connected network. Each neuron in the network is connected to all neurons in the previous layer and also to all neurons in the next layer, allowing information to be freely transmitted in the network. Each neuron affects the output of the entire network, thereby better integrating the influence of meteorological data on the condition vector.

[0023] The training of the conditional Transformer deep learning network includes: using the radiation signal at the target time as a label, the surface radiation signal at the input time, the input meteorological data, and the data at the target time as features, and feeding them into the conditional Transformer deep learning network for training to obtain a surface radiation signal inversion model.

[0024] During the training of the Conditional Transformer deep learning network, mean squared error (MSE) is used as the loss function, and the Adam optimizer is used for parameter updates. In addition, the feedforward neural networks of the encoder and decoder in the original Conditional Transformer deep learning network are replaced with convolutional neural networks. This achieves the complementary advantages of CNN's local feature extraction capability and Transformer's global dependency modeling capability, effectively overcoming the limitations of a single model.

[0025] Step 4: Based on the obtained surface radiation signal inversion model, invert the surface radiation brightness temperature to obtain surface radiation brightness temperature data, and evaluate the inversion accuracy; including: The target time data to be inverted and the corresponding meteorological data are fed into the corresponding encoding generator to obtain the target time feature vector and the meteorological data feature vector; Keep the context feature matrix extracted by the Encoder layer from the historical time-time radiation signals unchanged, and use the target time feature vector and the meteorological data feature vector as conditional constraints input to the Decoder layer; The Decoder layer decodes the context feature matrix and outputs the surface radiation brightness temperature data at the corresponding time. The retrieved surface radiance temperature data were compared with actual surface radiance temperature data. The mean mean absolute error (MBE) and symmetric mean absolute percentage error (SMAPE) were used as accuracy evaluation formulas to assess the model's retrieval capability. The formulas are as follows: Mean Error Deviation (MBE): ; Symmetric Mean Absolute Percentage Error (SMAPE): ; In the formula, N is the total number of samples. These are the values ​​retrieved from the model. The value represents the true value; the larger the average error deviation, the more severe the systematic bias of the model. Symmetric average absolute percentage error data are usually between 0% and 200%, and the smaller the value, the higher the prediction accuracy.

[0026] The embodiments of the present invention have been described in detail above with reference to the examples. However, the present invention is not limited to the above embodiments. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.

Claims

1. A multi-temporal radiation signal inversion method based on conditional Transformer, characterized in that: Includes the following steps: S1. Obtain meteorological time series data of the target area, and establish a dataset after preprocessing; S2. Design a conditional Transformer deep learning network, including an Encoder layer and a Decoder layer, as well as a time coding generator and a meteorological coding generator for encoding the target time data and meteorological data in the dataset; Specifically, the target time data and meteorological data encoded by the time code generator and the meteorological code generator are combined with the output of the Encoder layer and input into the Decoder layer to train the conditional Transformer deep learning network to obtain the surface radiation signal inversion model. S3. Based on the surface radiation signal inversion model obtained in step S2, invert the surface radiation brightness temperature to obtain surface radiation brightness temperature data, and evaluate the inversion accuracy.

2. The multi-temporal radiation signal inversion method based on conditional Transformer according to claim 1, characterized in that: The meteorological time series data includes: Multi-time radiation signals, including radiation brightness temperature data in the thermal infrared band; Meteorological data, including temperature profiles, specific humidity profiles, ozone mass mixture ratios, surface temperature, and surface air pressure; Target time data: Feature data based on satellite observation time.

3. The multi-temporal radiation signal inversion method based on conditional Transformer according to claim 2, characterized in that: The preprocessing includes: The DN values ​​recorded by the sensor are converted into satellite-observed radiation values ​​using a calibration formula, and the radiation brightness temperature data are obtained using the inverse function of Planck's blackbody radiation law. Based on the time and latitude / longitude coordinates of each data acquisition, data from the same geographical location at the same time are matched, using image pixels as the unit. Pixels with clouds were removed, and the interquartile range was used to determine the range of normal data, thus eliminating residual noise.

4. The multi-temporal radiation signal inversion method based on conditional Transformer according to claim 2, characterized in that: The time encoding generator uses an attention mechanism and residual structure to map discrete observation times into orthogonal target time feature vectors; the meteorological encoding generator uses a fully connected neural network to perform nonlinear feature extraction on multidimensional meteorological parameters to obtain meteorological data feature vectors.

5. The multi-temporal radiation signal inversion method based on conditional Transformer according to claim 4, characterized in that: The encoder layer consists of six stacked encoder layers with identical structures. Each encoder layer includes a multi-head attention mechanism module and a position-based feedforward neural network module. After the output of each encoder layer, residual connections and layer normalization are introduced for processing. The decoder layer comprises six stacked decoder layers with identical structures, each of which includes: A multi-head attention mechanism module with masks, including padding masks and sequence masks; The encoder-decoder attention mechanism module, and the query comes from the output of the previous decoder layer; Feedforward neural network module.

6. The multi-temporal radiation signal inversion method based on conditional Transformer according to claim 5, characterized in that: The output of the Decoder layer is mapped through a linear layer, and the probability is calculated using the Softmax activation function.

7. The multi-temporal radiation signal inversion method based on conditional Transformer according to claim 5, characterized in that: The training of the conditional Transformer deep learning network includes: using the radiation signal at the target time as a label, the surface radiation signal at the input time, the input meteorological data, and the data at the target time as features, and feeding them into the conditional Transformer deep learning network for training to obtain a surface radiation signal inversion model.

8. The multi-temporal radiation signal inversion method based on conditional Transformer according to claim 7, characterized in that: During the training of the Conditional Transformer deep learning network, mean squared error is used as the loss function, and the Adam optimizer is used for parameter updates; in addition, the feedforward neural networks of the encoder and decoder in the original Conditional Transformer deep learning network are replaced with convolutional neural networks.

9. The multi-temporal radiation signal inversion method based on conditional Transformer according to claim 1, characterized in that: The surface radiation signal inversion model specifically includes the inversion of surface radiation brightness temperature, which includes: The target time data to be inverted and the corresponding meteorological data are fed into the corresponding encoding generator to obtain the target time feature vector and the meteorological data feature vector; Keep the context feature matrix extracted by the Encoder layer from the historical time-time radiation signals unchanged, and use the target time feature vector and the meteorological data feature vector as conditional constraints input to the Decoder layer; The Decoder layer decodes the context feature matrix and outputs the surface radiation brightness temperature data at the corresponding time. The surface radiation brightness temperature data obtained by inversion is compared with the actual surface radiation brightness temperature data, and the accuracy is evaluated by average error deviation and symmetrical average absolute percentage error.

Images