Deep learning model for improving AIRS satellite ozone profile resolution in north pole region
The deep learning model is used to process the time field, space field and meteorological field of AIRS satellite data, and the problem of low time resolution of satellite data is solved, and the data resolution is improved and the observation accuracy is improved, helping to understand Arctic climate change.
Patent Information
- Application Number
- CN202411434138.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-15
- Publication Date
- 2025-02-11
AI Technical Summary
The AIRS satellite's ozone profile data in the Arctic region has a low temporal resolution, resulting in incomplete detection and high computational uncertainty.
A deep learning model is adopted, including data processing modules of time field, space field and meteorological field. The feature interaction relationship is captured through the denoising module and the Attention module, and the feature regression is used to predict the ozone concentration by using a linear residual neural network.
The time resolution of AIRS satellite data has been improved from daily to hourly level, greatly improving observation accuracy, and better observing STE phenomena and ozone level delivery in the Arctic, helping to understand the mechanisms of Arctic climate change.
Smart Images
Figure CN119416145A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of meteorological data processing technology, and in particular to a deep learning model for improving the resolution of AIRS satellite ozone profiles in the Arctic region. Background Art
[0002] About 90% of atmospheric ozone is located in the stratosphere and 10% in the troposphere. Tropospheric ozone is a short-lived climate factor that contributes to global and Arctic warming. Tropospheric ozone is greatly affected by human activities and has increased steadily over the past century. Elevated ozone levels can affect the physiological functions of plants and humans and may affect the climate system as a potent greenhouse gas. The polar regions are known as the "natural laboratories" of the Earth due to their special geographical location and unique environment. Compared with other regions, the polar regions are less affected by human activities, and climate change can better reflect the trend of global climate change. Therefore, it is necessary to monitor changes in polar ozone in a timely manner.
[0003] AIRS on NASA's Aqua satellite has been operating in a sun-synchronous polar orbit since May 2002. It is a hyperspectral infrared detector that provides daily global measurement coverage (two observations a day, at 1:30 UTC and 13:30 UTC). AIRS atmospheric measurements in the ozone u3 band provide sensitivity for estimating atmospheric ozone column density, and this product provides a spatial resolution of 1°×1° for weather forecasting and environmental monitoring. Although AIRS can provide daily global measurement coverage, there are still large gaps in the time dimension due to the presence of clouds. Missing single-day ozone data will result in incomplete detection. For annual or seasonal products, missing data will result in a large deviation between the product value and the true value, thereby increasing the uncertainty and bias of the calculation. Therefore, it is of practical significance to improve the temporal resolution of ozone satellite data.
[0004] With the development of GPU, a branch of machine learning, deep learning, has made great progress. In the field of deep learning, convolutional residual neural network has become a powerful tool to improve spatial resolution by intelligently fusing multiple observational data sets. Deep learning uses the network to capture complex spatial structures and functional relationships through a data-driven approach, ultimately producing enhanced spatial details of atmospheric components.
[0005] Currently, there are a variety of methods used to improve the data resolution of various atmospheric components, including tropospheric ozone. A common approach in traditional methods is to use data fusion and mixed observations from multiple sources to enhance the temporal resolution of the final product. This method is highly dependent on various observational data sets and has high requirements on the algorithm. Machine learning methods provide an alternative solution for improving resolution, which is less dependent on data than traditional methods. Common and effective methods include the use of random forests and stacked machine learning models, which can effectively extract patterns and relationships from input data to generate higher resolution outputs. However, ordinary machine learning methods have a poor fit to the data and are not consistent with the actual situation.
[0006] In summary, this application proposes a deep learning model for improving the resolution of AIRS satellite ozone profiles in the Arctic region. Summary of the invention
[0007] The purpose of the present invention is to propose a deep learning model for improving the resolution of AIRS satellite ozone profiles in the Arctic region in response to the problem of low temporal resolution of AIRS satellites in the background technology.
[0008] The technical solution of the present invention: a deep learning model for improving the resolution of AIRS satellite ozone profiles in the Arctic region, comprising:
[0009] Time field, processing the data of cross-day time field and intraday time field to extract time features;
[0010] Spatial field, used to process ozone concentration data at different longitudes and latitudes around the current site to extract spatial features;
[0011] Meteorological field, which processes meteorological condition data at the same height and location to extract meteorological features;
[0012] The denoising module is used to strengthen the correlation between the features of the time field, space field and meteorological field, capture the interactive relationship of the features, and input the extracted results of the time field, space field and meteorological field into the ATTENTION module for interaction;
[0013] The linear residual neural network module uses the interaction results of the denoising module to perform feature regression and predict ozone concentration.
[0014] Optionally, the cross-day time field of the time field is the ozone concentration data of the station at the same simulated altitude for the previous three days, and the current-day time field is the ozone concentration data of the station for the previous three hours at the same simulated altitude.
[0015] Optionally, the spatial field contains ozone concentration data of nine points within the range of ±0.25° longitude and latitude at the same simulation altitude of the current site. The setting rules of the nine points are as follows: the longitude and latitude of the observation point are multiples of 0.25, that is, the two decimal places are 0.00, 0.25, 0.50 or 0.75. When the longitude and latitude of the prediction point do not meet the above four conditions, the prediction is made based on the principle of proximity. The longitude and latitude of the prediction point are ±0.25°, forming a nine-point area [(+0.25, 0, -0.25) × (+0.25, 0, -0.25)]. The nine points represent the changes in ozone values in the spatial field, which can effectively avoid the extreme value problem.
[0016] Optionally, the denoising module consists of a convolutional layer, a ReLU activation function, a batch normalization (BN) layer and a residual structure, which is used to remove noise during the feature extraction process while retaining the data characteristics of the time field, spatial field and meteorological field as much as possible.
[0017] Optionally, the linear residual neural network module consists of a multi-head attention layer and multiple residual linear blocks, captures the dependencies between input elements through the attention mechanism, and outputs a feature representation that includes the comprehensive impact of temporal and spatial meteorological conditions.
[0018] Optionally, the meteorological condition data processed by the meteorological field processing module include data on temperature, vortex, ozone concentration, relative humidity, wind speed in the U direction, wind speed in the V direction and pressure meteorological conditions.
[0019] Compared with the prior art, the present invention has at least one of the following beneficial technical effects:
[0020] The deep learning model of the present invention predicts data more smoothly and realistically, relying only on reanalysis data and using mature and stable ERA5 data to ensure the reliability and continuity of the data source, making it easy to use. Compared with machine learning, the model of the present invention can better handle the fitting between complex relationships and achieve better results.
[0021] Compared with traditional methods, this model has lower requirements on algorithms, lower requirements on data quality, and lower requirements on data sources.
[0022] Compared with common machine learning methods, this model fits the data better, is smoother and more in line with actual conditions.
[0023] The model of the present invention can improve the resolution of AIRS satellite data from the daily level to the hourly level, greatly improving the observation accuracy. In addition, the model can be used to better observe the STE phenomenon and ozone horizontal transport in the Arctic on an hourly time scale, and can better understand the mechanism of Arctic climate change. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 A diagram of the deep learning model architecture used to improve the resolution of AIRS satellite ozone profiles in the Arctic region;
[0025] Figure 2 This is the block structure diagram of the CRC module;
[0026] Figure 3 This is the structural diagram of RLblock;
[0027] Figure 4 The potential vorticity profile and ozone profile diagram over the Eureka station within 24 hours in the embodiment;
[0028] Figure 5 The ozone distribution diagram of the concentration change at 300hPa above the Eureka station in the embodiment;
[0029] Figure 6 It is the ozone distribution diagram of 400hPa concentration change over Eureka Station in the embodiment. DETAILED DESCRIPTION
[0030] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0031] It should be noted that the following are the full English names and Chinese translations of some abbreviations in the specification and the drawings:
[0032] Multi-headAttention: multi-head attention mechanism; Rlblock (Residual Linear block): residual linear block; Conv1d(x): one-dimensional convolution layer with input dimension x; LSTM: long short-term memory network; Conv: convolution layer (the model uses two-dimensional); block: denoising module, which is composed of various basic modules; BN (BatchNormalization): batch normalization; ReLu: rectified linear activation function; Output: output data; nn.Linear: linear transformation layer; layernorm: layer normalization.
[0033] Example
[0034] like Figure 1 As shown in the figure, the deep learning model proposed in the present invention for improving the resolution of AIRS satellite ozone profiles in the Arctic region has the following architecture description: the model extracts features and concentration trends by processing data in three fields, including time field, space field, and meteorological field. In the ARL module, attention is used to complete the interaction, and then regression is performed to obtain the predicted concentration.
[0035] The architecture of this deep learning model consists of three parts: time field, space field and meteorological field. The following is a detailed introduction to these three fields and their connection and workflow.
[0036] Time field: The time field is divided into two sub-fields: the cross-day time field and the same-day time field. After the Reshape operation, it is input into the CRC module for processing. The cross-day time field is the ozone concentration data of the station at the same simulated altitude for the previous three days, that is, using the 300hPa reanalysis data to predict the 300hPa satellite data. For example, when predicting the ozone concentration at 1:00 p.m. on the fourth day, the input is the ozone concentration data at 1:00 p.m. on the first, second, and third days. The same-day time field is the ozone concentration data of the station in the previous three hours at the same simulated altitude. For example, when predicting the ozone concentration at 1:00 p.m. on the fourth day, the input data is the ozone concentration at 10:00, 11:00, and 12:00 a.m. on the same day.
[0037] Spatial field: The spatial field represents the ozone concentration values at different longitudes and latitudes around the current station at the same altitude. Specifically, the spatial field contains the ozone concentration data of nine points within the range of ±0.25° longitude and latitude at the same simulation altitude of the current station. This setting enables the model to capture spatial changes and distribution characteristics. The setting rule for the nine points is: the longitude and latitude of the observation point are multiples of 0.25, that is, the two decimal places are 0.00, 0.25, 0.50 or 0.75. When the longitude and latitude of the prediction point do not meet the above four conditions, the nearest principle is used for prediction. The longitude and latitude of the prediction point are ±0.25°, forming an area of nine points [(+0.25,0,-0.25)×(+0.25,0,-0.25)]. The nine points represent the changes in ozone values in the spatial field, which can effectively avoid extreme value problems.
[0038] Meteorological field: Meteorological field is the introduction of meteorological conditions at the same location at the same altitude, including the following meteorological conditions: temperature, vortex, ozone concentration, relative humidity, wind speed in U direction, wind speed in V direction, and pressure.
[0039] In the entire architecture, features are extracted through different modules. The extracted time field, space field and meteorological field results are uniformly input into the ATTENTION module for interaction. The purpose of this stage is to use the multi-head attention mechanism to strengthen the correlation between features and capture complex feature interactions. After the interaction is completed, the results are regressed through the linear residual neural network (ARL layer). This design enables the model to better learn the changing laws of ozone concentration and make accurate predictions.
[0040] Dividing the data into different fields and processing them independently ensures the independence of each data feature. The advantages of this method are:
[0041] Improve feature expression capabilities: By independently extracting various features, the characteristics and trends of the data can be more clearly reflected, making the model more responsive to different factors.
[0042] Stable prediction: Through comprehensive interaction, the impact of meteorological conditions and other factors on ozone concentration can be effectively reflected, so that it can be kept within a stable value range during prediction.
[0043] Solving extreme value problems: The design of this architecture helps to better predict the extreme value phenomenon of ozone concentration and enhances the robustness of the model under extreme weather conditions.
[0044] In this embodiment, if Figure 2 The block structure of the CRC module is a denoising module, which is composed of various basic modules. It can be seen that the architecture of the denoising module uses a residual network structure to reduce the impact on the source features. The multi-layer perceptron (MLP) has a small field of view and has a poor denoising effect on direct data. Therefore, a convolutional layer is used for denoising, and multiple denoising modules are superimposed for iterative denoising to further improve the denoising effect. In order to avoid the excessive impact of the large dimension of the time field data on the final processing, an additional convolutional layer is introduced at the end of the module to compress and fuse the features, so that its dimension is appropriately compressed while retaining the key time change information.
[0045] The core purpose of the denoising module is to effectively remove noise during feature extraction while retaining key information. This module is mainly composed of convolutional layers (Conv), rectified linear activation functions (ReLU), batch normalization (BN) layers, and residual structures. The following is a detailed description of the components of the denoising module, as well as their connections and workflow.
[0046] The denoising module mainly uses convolutional layers (Conv) to preserve information features, uses the rectified linear activation function (ReLU) for denoising, and uses batch normalization (BN) layers to ensure the stability of training and avoid problems such as gradient disappearance or explosion and internal covariate shift. At the same time, the residual structure is used to ensure that information and gradients can be transmitted more effectively, avoiding the problem of gradient disappearance, thereby supporting the deep design of the network.
[0047] The denoising module consists of the following:
[0048] 1. Composition characteristics
[0049] Convolutional layer (Conv): The convolutional layer is the basis of the denoising module and is responsible for local perception and feature extraction of input features. Through the convolution operation, the spatial features and local patterns in the input data can be effectively captured, thereby helping to retain useful information.
[0050] ReLU activation function: The ReLU activation function is applied after the output of the convolutional layer to make the negative output zero. This nonlinear transformation helps to increase the expressive power of the network. At the same time, in the denoising process, ReLU can suppress some noise signals and improve the robustness of the model.
[0051] Batch Normalization (BN): The batch normalization layer is used to normalize the output of the convolutional layer to ensure that the input distribution of each layer remains stable. This can effectively avoid the problem of gradient disappearance or explosion and reduce the impact of internal covariate shift, thereby improving the training stability and convergence speed of the model.
[0052] Residual structure: To ensure that information and gradients can be effectively propagated, the denoising module adopts a residual structure. This structure introduces skip connections, which makes it easier for the network to transmit information during training and avoid the gradient vanishing problem. In this way, the depth of the network can be effectively improved, promoting the learning of more complex features.
[0053] After ablation testing, the denoising module can significantly remove noise from features while maintaining feature integrity. This effect not only improves the prediction accuracy of the model, but also enhances its resistance to interference information. Through effective structural design, the denoising module makes the feature extraction process more efficient and provides clearer and more accurate input data for subsequent model training and prediction.
[0054] In summary, the denoising module achieves the goal of effective denoising in feature extraction through carefully designed components and efficient workflow, while retaining key information to the greatest extent, laying the foundation for improving the overall performance of the model.
[0055] Among them, the construction of RLblock is as follows Figure 3 As shown in the figure, in order to further extract and fuse data features in the fields of time, space and meteorology, this study designed an ARL module based on the attention mechanism, and its structure is shown in the figure above. The module consists of a multi-head attention layer and multiple residual linear blocks (RLBlocks). Through the attention mechanism, the model can fully capture the complex dependencies between input elements, thereby outputting feature representations that include the comprehensive impact of time and space meteorology.
[0056] The ARL (Attention Residual Learning) module is designed to enhance feature interaction and information storage capabilities, thereby improving the prediction of ozone concentration. The module is mainly composed of a comprehensive interactive Attention head and multiple Residual Learning Blocks (Rlblock). The following is a detailed introduction to the components of the ARL module, as well as their connections and workflows.
[0057] 1. Composition characteristics
[0058] Attention head: The core of the ARL module is the comprehensive interactive Attention head. This component deepens the feature interaction of the three input fields (time field, space field and meteorological field) through the attention mechanism. The Attention head can dynamically weigh the importance of different features and highlight the most helpful information for ozone concentration prediction. This feature weighting enables the model to better capture complex correlations and patterns and improve the accuracy of predictions.
[0059] Multiple Rlblocks: After the Attention head, the ARL module is connected to multiple ResidualLearningBlocks (Rlblocks). Each Rlblock consists of multiple convolutional layers, ReLU activation functions, and batch normalization layers, and uses residual connections to enhance information flow. Through these Rlblocks, the ARL module is able to process deeper information and capture subtle changes and complex relationships in the data.
[0060] 2. Connections between features
[0061] The connection method of ARL module is as follows:
[0062] Input features: The module receives basic ozone concentration values and their corresponding time fields, space fields and meteorological fields as input.
[0063] Feature interaction: The input features are first processed by the Attention head to perform feature weighting and interaction to generate enhanced feature representation.
[0064] Hierarchical information processing: The enhanced feature representation is passed to multiple Rlblocks. Each Rlblock further extracts hierarchical information through convolution and nonlinear transformation, and uses residual connections to ensure effective propagation of information.
[0065] Output features: After being processed by multiple Rlblocks, the final output features are passed to the subsequent regression layer to predict the ozone concentration.
[0066] The design of the ARL module significantly improves the performance of the model, especially when processing complex data. Experimental verification shows that when residual connections are not used, the information storage capacity of the model will be greatly reduced, and even lead to the phenomenon of non-regression in prediction. This shows the key role of residual connections in enhancing information flow and gradient propagation.
[0067] Through the ARL module, the model can effectively integrate characteristic information from different fields and enhance its sensitivity to changes in ozone concentration. Ultimately, the ARL module improves the prediction accuracy and robustness of the model, allowing it to maintain excellent regression capabilities when facing diverse inputs.
[0068] Performance Test:
[0069] The troposphere is divided into two heights, the upper troposphere and the lower troposphere. The corresponding heights are Height1 (300, 400, 500, 600, 700 hPa) and Height2 (850, 925, 1000 hPa). We used R2, RRMSE, and RMAE for scoring.
[0070] The calculation formulas for R2, RRMSE, and RMAE are as follows:
[0071]
[0072] Among them, y i is the true observed value, is the model prediction value, is the mean of the true observations, y e is the ERA5 reanalysis data, and n represents the number of samples.
[0073] The results are shown in the following table:
[0074]
[0075] Application Example 1 Ozone exchange in STE (Stratosphere-Tropospheric Exchange):
[0076] like Figure 5 As shown, an example over the Eureka station on October 16, 2023, shows the potential vorticity profile and ozone profile within 24 hours, when a typical STE phenomenon occurred. Due to the invasion of the stratosphere, the ozone in the upper troposphere increased abnormally. Compared with 1:00 UTC, the ozone at the bottom of the stratosphere decreased by 0.009ppmv, while the ozone at the top of the troposphere increased by 0.03ppmv. At the strongest time of the STE, the ozone in the upper troposphere increased by 36% compared with the initial value of the day. The fine changes in the ozone profile provide support for a better assessment of the impact of STE on the change of ozone in the Arctic spring.
[0077] Application Example 2OHT is the abbreviation of Ozone Horizontal Transport, which means ozone horizontal transport in Chinese;
[0078] To observe the horizontal transport of ozone, the model of the present invention was run within 1° latitude and 6° longitude centered on the station, covering an area of approximately 50 km × 50 km around the station.
[0079] It is worth noting that since the spatial resolution of the reanalysis data is 0.25°×0.25°, the spatial resolution of our model has also been improved to 0.25°×0.25°, making the changes in ozone horizontal transport more detailed and obvious. The ozone distribution over the Eureka station was observed on July 5 and July 6, 2022. The concentrations at the bottom of the stratosphere and the top of the troposphere are shown in Figure 2. Figure 5 shown.
[0080] like Figure 5-6 As shown in the figure, the horizontal ozone flow trend in the area near the station, the color represents the ozone concentration during the evolution process, and the arrow represents the wind vector. The concentration changes greatly from 00:00 to 06:00 UTC on the 6th, so the observation is made with hourly resolution at this time.
[0081] like Figure 6 As shown, from 12:00 on July 5 to 9:00 on July 6, ozone was transported from the Eureka station to the northwest and then to the southeast. From 0:00 to 4:00 on July 6, a large amount of ozone was transported from the northwest, causing the concentration at the station to rise sharply. After 4:00, ozone near the Eureka station was transported to the southeast, causing the concentration to drop sharply. Figure 6 The wind direction and wind speed are marked, which shows that the horizontal transport of ozone is consistent with the wind direction. At the same time, the hourly ozone value is calculated to obtain the horizontal ozone flux at the Eureka station. According to calculations, from 9:00 on July 5 to 12:00 on July 6, about 0.25ppmv of ozone passed through the bottom of the stratosphere, and about 0.11ppmv of flux passed through the top of the troposphere. The model successfully observed the horizontal movement of ozone and the amount of horizontal movement, providing a reference for studying polar environmental processes and assessing the impact of climate change.
[0082] The deep learning model of the present invention predicts data more smoothly and realistically, relying only on reanalysis data and using mature and stable ERA5 data to ensure the reliability and continuity of the data source, making it easy to use. Compared with machine learning, the model of the present invention can better handle the fitting between complex relationships and achieve better results.
[0083] Compared with traditional methods (such as data fusion), this model has lower requirements on algorithms, lower requirements on data quality, and lower requirements on data sources.
[0084] Compared with common machine learning methods (such as decision trees), this model fits the data better, is smoother and more in line with the actual situation.
[0085] The model of the present invention can improve the resolution of AIRS satellite data from the daily level to the hourly level, greatly improving the observation accuracy. In addition, the model can be used to better observe the STE phenomenon and ozone horizontal transport in the Arctic on an hourly time scale, and can better understand the mechanism of Arctic climate change.
[0086] The present invention uses reanalysis data to ensure the reliability and continuity of the data source, greatly improving its observation accuracy compared to the original satellite observation data, and can better understand the mechanism of Arctic climate change.
[0087] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.
Claims
1. A deep learning model for improving the resolution of AIRS satellite ozone profiles in the Arctic region, characterized by: include: Time field: Process the data of cross-day time field and intraday time field, and extract time features; Spatial field, used to process ozone concentration data at different longitudes and latitudes around the current site and extract spatial features; Meteorological field, processing meteorological condition data at the same height and location, extracting meteorological characteristics; The denoising module is used to strengthen the correlation between the features of the time field, space field and meteorological field, capture the interactive relationship of the features, and input the extracted results of the time field, space field and meteorological field into the ATTENT ION module for interaction; The linear residual neural network module uses the interaction results of the denoising module to perform feature regression and predict ozone concentration.
2. The deep learning model for improving the resolution of AIRS satellite ozone profiles in the Arctic region according to claim 1, characterized in that: The cross-day time field of the time field is the ozone concentration data of the station at the same simulated altitude for the previous three days at the same time, and the same-day time field is the ozone concentration data of the station for the previous three hours at the same simulated altitude.
3. The deep learning model for improving the resolution of AIRS satellite ozone profiles in the Arctic region according to claim 1, characterized in that: The spatial field includes ozone concentration data of nine points within the range of ±0.25° longitude and latitude at the same simulation altitude of the current station.
4. The deep learning model for improving the resolution of AIRS satellite ozone profiles in the Arctic region according to claim 1, characterized in that: The denoising module consists of a convolutional layer, a ReLU activation function, a batch normalization layer and a residual structure, and is used to remove noise during the feature extraction process while retaining the data characteristics of the time field, space field and meteorological field as much as possible.
5. The deep learning model for improving the resolution of AIRS satellite ozone profiles in the Arctic region according to claim 1, characterized in that: The linear residual neural network module consists of a multi-head attention layer and multiple residual linear blocks. It captures the dependencies between input elements through the attention mechanism and outputs feature representations containing the comprehensive effects of temporal and spatial meteorology.
6. The deep learning model for improving the resolution of AIRS satellite ozone profiles in the Arctic region according to claim 1, characterized in that: The meteorological condition data processed by the meteorological field processing module include data on temperature, vortex, ozone concentration, relative humidity, wind speed in the U direction, wind speed in the V direction and pressure meteorological conditions.