CN117593666A - Geomagnetic station data prediction method and system for aurora image - Google Patents

Geomagnetic station data prediction method and system for aurora image Download PDF

Info

Publication number
CN117593666A
CN117593666A CN202410081481.7A CN202410081481A CN117593666A CN 117593666 A CN117593666 A CN 117593666A CN 202410081481 A CN202410081481 A CN 202410081481A CN 117593666 A CN117593666 A CN 117593666A
Authority
CN
China
Prior art keywords
layer
station data
module
features
geomagnetic station
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202410081481.7A
Other languages
Chinese (zh)
Other versions
CN117593666B (en
Inventor
肖晖
田馨沁
李俊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Aeronautics and Astronautics
Original Assignee
Nanjing University of Aeronautics and Astronautics
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Aeronautics and Astronautics filed Critical Nanjing University of Aeronautics and Astronautics
Priority to CN202410081481.7A priority Critical patent/CN117593666B/en
Publication of CN117593666A publication Critical patent/CN117593666A/en
Application granted granted Critical
Publication of CN117593666B publication Critical patent/CN117593666B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0499—Feedforward networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Multimedia (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Astronomy & Astrophysics (AREA)
  • Remote Sensing (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses a geomagnetic station data prediction method and a geomagnetic station data prediction system of an aurora image, wherein the geomagnetic station data prediction method and the geomagnetic station data prediction system comprise the steps of obtaining polar region ultraviolet aurora image data of a satellite and geomagnetic station data of a ground station; preprocessing the polar ultraviolet image data to construct a data set, and dividing a training set, a verification set and a test set; constructing a multi-layer convolution network module to extract local characteristics of aurora eggs; constructing a transducer coding network module to carry out global interaction on aurora images to code aurora ovum global features, constructing a hierarchy feature fusion module, aggregating the attention weights among hierarchies, screening coding layer discriminant features, and fusing local and global feature representations to obtain prediction features; constructing a regression prediction module and outputting geomagnetic station data; defining a loss function, and training a geomagnetic station data prediction model; and predicting geomagnetic station data by using a geomagnetic station data prediction model, and analyzing and evaluating a prediction result. The method and the device improve the accuracy of the model in predicting the geomagnetic station data.

Description

Geomagnetic station data prediction method and system for aurora image
Technical Field
The invention belongs to the field of aurora feature extraction and geomagnetic station data prediction, and particularly relates to a geomagnetic station data prediction method and system of an aurora image.
Background
With the successful application of the data-driven deep learning method in multiple fields such as computer vision, natural language processing and the like, space-time data mining based on the deep learning method is a new technology development trend. The ultraviolet aurora image data and the geomagnetic station data are both space-time data, and have a certain complementary relationship in the time dimension and the space dimension. Geomagnetic index prediction based on deep learning is considered as a powerful complement to conventional methods. Deep learning networks such as an Artificial Neural Network (ANN), a BP neural network, a generalized regression neural network, and an extreme learning machine are widely used for predicting geomagnetic indexes such as AE and Dst. However, the above prediction model mostly adopts the structural data such as the inter-planetary magnetic field component, the solar wind density, the speed and the like as input, and based on the physical process driven by solar wind, the correlation relationship between the aurora image and geomagnetic data on the spatial scale is ignored. Research shows that the polar light image and the geomagnetic station data have a large-scale spatial association relationship, and the local change characteristics of the polar light egg shape and the brightness are also strongly related to the intensity of the geomagnetic station data. In recent years, the Transformer exhibits superior performance in general image classification, image retrieval, semantic segmentation, and the like. Vision Transformer (ViT), through its inherent attentiveness mechanisms, is able to globally characterize an image and automatically identify distinctive feature areas in the image. However, the receptive field of ViT is not effectively extended and the length of the attention of the individual pixel blocks does not change as the number of encoder layers increases. In addition, viT inputting a fixed-size block of pixels is detrimental to network capture of critical regional attention information. The self-attention mechanism of the transducer has a natural good modeling capability on the global characteristics of the input information, but the acquisition of the local information is not as strong as that of CNN.
Therefore, there is a need to solve the above-mentioned problems.
Disclosure of Invention
The invention aims to: the first object of the invention is to provide a geomagnetic station data prediction method of an aurora image, which utilizes a multi-layer convolution network to extract local features of the aurora image and retain position information, utilizes a transform coding network to learn global features of the aurora image, utilizes a hierarchical feature fusion module to fuse the local feature representation and the global feature representation layer by layer, and aggregates discrimination features among layers, thereby improving the expression capability of a model on fine granularity features of the aurora image, and further improving the accuracy of the model on geomagnetic station data prediction.
A second object of the present invention is to provide a geomagnetic station data prediction system of an aurora image.
The technical scheme is as follows: in order to achieve the above object, the present invention discloses a geomagnetic station data prediction method of an aurora image, comprising the following steps:
(1) Acquiring polar ultraviolet image data of a satellite region and geomagnetic station data of a ground station;
(2) Preprocessing the polar ultraviolet image data to construct a data set, and dividing a training set, a verification set and a test set;
(3) Constructing a multi-layer convolution network module to extract local characteristics of aurora eggs, wherein the multi-layer convolution network of the multi-layer convolution network module totally comprises four convolution blocks, the first two convolution blocks consist of two convolution layers and one pooling layer, and the second two convolution blocks consist of three convolution layers and one pooling layer;
(4) Constructing a transducer coding network module to carry out global interaction on an aurora image so as to code the global characteristics of aurora eggs, wherein a coding network of the transducer coding network module consists of an embedded layer and a coding layer, the embedded layer divides the input aurora image into patches with fixed sizes, then the patch vectors are converted into embedded vectors with fixed lengths through linear mapping, and the embedded position codes are used for conforming to the input normal form of the transducer coding layer;
(5) Constructing a hierarchical feature fusion module, aggregating the attention weights among the hierarchies, screening the discriminant features of the coding layers, and fusing the local feature representation and the global feature representation to obtain predicted features; the hierarchical feature fusion module is divided into feature aggregation between front and back layers and fusion of convolution block features and coding features, wherein the feature aggregation between the front and back layers is realized by fusing attention weights between the front and back layers and screening marks so as to extract coding layer discriminant features after hierarchical aggregation; the fusion of the convolution block features and the coding features is to fuse the convolution block features and the coding layer discriminant features after hierarchical aggregation to form fusion features;
(6) Constructing a regression prediction module, expanding the final prediction features into vectors, inputting the prediction features into the regression prediction module, and outputting geomagnetic station data;
(7) Defining a loss function, training and verifying a geomagnetic station data prediction model consisting of a multi-layer convolution network module, a transform coding network module, a hierarchical feature fusion module and a regression prediction module by using the training set and the verification set constructed in the step (2), and adjusting super parameters for training to obtain a trained geomagnetic station data prediction model;
(8) Predicting geomagnetic station data by using a trained geomagnetic station data prediction model, and analyzing and evaluating a prediction result; inputting the geomagnetic station data prediction model trained in the step (7) by using the test set constructed in the step (2) to obtain a geomagnetic station data prediction result; and combining the true value to perform error analysis and evaluation on the prediction result.
In the step (1), region ultraviolet aurora image data shot by an ultraviolet imager UVI are acquired from an SPDF website of a spatial physical data facility of NASA; geomagnetic station data which are in the same time period, have magnetic latitude of 60-80 degrees and are uniformly distributed along magnetic longitude are obtained from a world data center WDC, and the geomagnetic station data refer to minute-by-minute data of a geomagnetic field horizontal H component.
Preferably, in the step (2), preprocessing the polar ultraviolet image data of the polar region shot by UVI, and reading the polar ultraviolet image data by using a python script, obtaining imaging time, an image matrix and geographic longitude and latitude and geomagnetic longitude and latitude of each single sample by using a cdf data file, carrying out coordinate transformation on the polar light image, wherein a target coordinate system takes a geomagnetic pole as a center, the magnetic latitude ranges from 50 degrees to 90 degrees MLAT, the magnetic local time ranges from 0 to 24MLT, and finally resetting negative pixels of the image to 0; the LBHL band 160nm-180nm image of the polar ultraviolet image data in the polar ultraviolet region is screened out and used as the input of a geomagnetic station data prediction model, and geomagnetic station data in the same period is output of the geomagnetic station data prediction model; the aurora image and geomagnetic station data are put into different folders to form a data set, and the data set is represented by 7:2: the scale of 1 is divided into a training set, a validation set and a test set.
And (3) each convolution layer is followed by a ReLU activation function, and the convolution layers uniformly adopt 3 multiplied by 3 convolution kernels.
Further, the transducer encoder of the transducer encoding network module in the step (4) is composed of four identical encoding layers, each encoding layer is sequentially composed of a layer normalized LN, a multi-head self-attention module MHSA, a residual connection, a layer normalized LN, a multi-layer perceptron MLP and a residual connection, and the multi-layer perceptron MLP is composed of two layers of convolution functions and a ReLU activation function.
Preferably, in the step (5), the hierarchical feature fusion module fuses the current layer attention weight and the previous layer attention weight through matrix multiplication, extracts the coding layer discriminant features after hierarchical aggregation by using a Max function, fuses the convolution block features and the coding layer discriminant features after hierarchical aggregation to form fusion features through matrix multiplication, and finally inputs a transform coding layer again to obtain the prediction features.
In step (6), the regression prediction module is composed of three layers of fully-connected networks, including two hidden layers and an output layer, wherein the number of neurons of the three layers of fully-connected networks is 1024, 1024 and 12 respectively; firstly, expanding the prediction features obtained by the transform coding layer into vectors, inputting the prediction features into a regression prediction module, and finally outputting geomagnetic station data.
Further, if in step (7)Representing the real sequence->Representing the predicted sequence, the loss function is defined as the mean square error between the geomagnetic station data predicted value and the true value:
。
preferably, the evaluation criteria in step (8) include a root mean square error RMSE, an average relative variance ARV and a decision coefficient R 2 。
The invention relates to a geomagnetic station data prediction system of an aurora image, which comprises the following components:
the original data acquisition module is used for acquiring polar ultraviolet image data of a satellite and geomagnetic station data of a ground station;
the data set constructing module is used for preprocessing the polar ultraviolet image data of the polar region to construct a data set and dividing a training set, a verification set and a test set;
the multi-layer convolution network module is constructed to extract the local characteristics of aurora eggs, and the multi-layer convolution network of the multi-layer convolution network module totally comprises four convolution blocks, wherein the first two convolution blocks consist of two convolution layers and one pooling layer, and the second two convolution blocks consist of three convolution layers and one pooling layer;
the method comprises the steps of constructing a transducer coding network module, performing global interaction on an aurora image to code aurora ovum global characteristics, wherein a coding network of the transducer coding network module consists of an embedded layer and a coding layer, the embedded layer divides the input aurora image into patches with fixed sizes, then converts patch vectors into embedded vectors with fixed lengths through linear mapping, and embeds position codes to accord with an input paradigm of the transducer coding layer;
the hierarchical feature fusion module is used for constructing a hierarchical feature fusion module, aggregating the attention weights among the hierarchies, screening the discriminant features of the coding layer, and fusing the local feature representation and the global feature representation to obtain the prediction features; the hierarchical feature fusion module is divided into feature aggregation between front and back layers and fusion of convolution block features and coding features, wherein the feature aggregation between the front and back layers is realized by fusing attention weights between the front and back layers and screening marks so as to extract coding layer discriminant features after hierarchical aggregation; the fusion of the convolution block features and the coding features is to fuse the convolution block features and the coding layer discriminant features after hierarchical aggregation to form fusion features;
the regression prediction module is used for constructing a regression prediction module, expanding the final prediction characteristics into vectors, inputting the prediction characteristics into the regression prediction module and outputting geomagnetic station data;
training a geomagnetic station data prediction model module, defining a loss function, training and verifying a geomagnetic station data prediction model formed by a multi-layer convolution network module, a Transformer coding network module, a hierarchical feature fusion module and a regression prediction module by utilizing a training set and a verification set constructed in a construction data set module, and adjusting super parameters for training to obtain a trained geomagnetic station data prediction model;
the geomagnetic station data prediction module predicts geomagnetic station data by using a trained geomagnetic station data prediction model, and analyzes and evaluates a prediction result; inputting a trained geomagnetic station data prediction model by using a test set constructed in a data set constructing module to obtain a geomagnetic station data prediction result; and combining the true value to perform error analysis on the prediction result, wherein the evaluation criteria comprise Root Mean Square Error (RMSE), average Relative Variance (ARV) and decision coefficient (R) 2 。
The beneficial effects are that: compared with the prior art, the invention has the following remarkable advantages:
(1) According to the invention, the local features of the aurora image are extracted by using a multi-layer convolution network, the position information is reserved, the global features of the aurora image are learned by using a Transformer coding network module, the local feature representation and the global feature representation are fused layer by using a hierarchical feature fusion module, the distinguishing features between layers are polymerized, the expression capability of the model on the fine granularity features of the aurora image is improved, and therefore, the prediction accuracy of the model on geomagnetic station data is improved;
(2) From the angle of the aurora image, the invention uses a deep learning model to replace a simple neural network, uses a convolution network to enhance the characteristic representation capability under a transducer self-attention mechanism, and extracts the global characteristic and the local characteristic of the aurora image in parallel; the hierarchical feature fusion module not only fuses the two types of features, but also re-aggregates features among the layers; finally, the prediction research of geomagnetic station data based on aurora images is realized by means of a fully connected network;
(3) The multi-layer convolution network module constructed by the invention utilizes a serial neural network structure to extract local characteristics of aurora eggs and retain position information; the constructed transducer coding network module learns global space characteristics between aurora ovum and geomagnetic station data; the constructed hierarchical feature fusion module aggregates the attention weights among the hierarchies, screens the discriminative feature areas, and fuses the convolution features and the coding features layer by layer; the constructed regression prediction module maps the two-dimensional feature matrix into a one-dimensional feature vector, and carries out final regression prediction on geomagnetic station data through a hidden layer and an output layer.
Drawings
FIG. 1 is a schematic diagram of a frame of the present invention;
FIG. 2 is a schematic diagram of a coding layer in a transform coding network module according to the present invention;
fig. 3 is a schematic structural diagram of a hierarchical feature fusion module according to the present invention.
Detailed Description
The technical scheme of the invention is further described below with reference to the accompanying drawings.
Example 1
The invention discloses a geomagnetic station data prediction method of an aurora image, which is arranged in a computer with the following configuration: AMDRyzen9 3950X 16 core processor, nvidia GeForceRTX3090 graphics processor, main frequency 3.49GHz, memory 64GB, and operating system windows10. The implementation of the geomagnetic station data prediction method of the aurora image is based on a Tensorflow2.0 deep learning framework toolkit.
The invention discloses a geomagnetic station data prediction method of an aurora image, which comprises the following steps of:
(1) Acquiring polar ultraviolet image data of a polar region of a satellite and geomagnetic station data of 12 ground stations; acquiring regional ultraviolet aurora image data shot by an ultraviolet imager UVI from an SPDF website of a spatial physical data facility of NASA; acquiring 12 geomagnetic station data which are uniformly distributed along magnetic longitude and within the range of 60-80 degrees of magnetic latitude in the same time period from a world data center WDC, wherein the geomagnetic station data refer to minute-by-minute data of a geomagnetic field horizontal H component;
the SPDF website of NASA provides UVI _level1 data products under UVI sensors carried by POLAR satellites, the pixel size of aurora ovum images of the products is 200 multiplied by 228, the spatial resolution is about 0.04 degrees per pixel, the world data center WDC provides minute-by-minute monitoring data of magnetometers of geomagnetic stations in all places, the invention selects 12 geomagnetic stations which have the same period as the POLAR satellites and have magnetic latitude in the range of 60-80 degrees and are uniformly distributed along the magnetic longitude, and the geomagnetic station data particularly refers to horizontal H components in magnetic field parameters.
(2) Preprocessing the polar ultraviolet image data to construct a data set, and dividing a training set, a verification set and a test set; the method comprises the steps of preprocessing polar light image data of a polar region ultraviolet shot by UVI, reading the polar light image data by using a python script, obtaining imaging time and an image matrix of each single sample, and geographic longitude and latitude and geomagnetic longitude and latitude of the single sample, carrying out coordinate transformation on a polar light image, wherein a target coordinate system takes a geomagnetic pole as a center, the magnetic latitude ranges from 50 degrees to 90 degrees MLAT, the magnetic local time range is from 0 to 24MLT, and finally resetting negative value pixels of the image to 0 to eliminate satellite noise; the LBHL wave band 160nm-180nm image of the polar ultraviolet image data is screened out to be used as the input of a geomagnetic station data prediction model, and 12 geomagnetic station data in the same time period are output of the geomagnetic station data prediction model; the aurora image and geomagnetic station data are put into different folders to form a data set, and the data set is represented by 7:2:1 is divided into a training set, a verification set and a test set;
and the cdf data file is used for obtaining the imaging time, the image matrix and the geographic longitude and latitude and geomagnetic longitude and latitude of each single sample. And carrying out coordinate conversion on the polar light image, wherein the target coordinate system takes a geomagnetic pole as a center, the magnetic latitude range is 50-90 degrees MLAT, the magnetic local time range is 0-24MLT, the pixel size of the converted image is 241 multiplied by 241, and the gray scale range is 0-255. Since the LBHS band is severely absorbed by oxygen Schumann Rungeband, the embodiment uses the LBHL band (160-180 nm) of POLAR satellite, and the time resolution between two consecutive LBHL images is between 0.5-3 minutes in the normal observation mode. Screening LBHL (160 nm-180 nm) images of UVI imaging data as input of a geomagnetic station data prediction model, wherein 12 geomagnetic station data at the same moment are output of the geomagnetic station data prediction model; the aurora image data is image sequence data with specific spatial resolution, the imaging range comprises the whole arctic region, the lowest latitude is 50 DEG, and each sample imageMatching vectors consisting of 12 geomagnetic station data at the same time +.>The aurora image and geomagnetic station data are put into different folders, 46152 groups of data are added together, and the data set is represented by 7:2: the scale of 1 is divided into a training set, a validation set and a test set.
(3) Constructing a multi-layer convolution network module to extract local characteristics of aurora eggs, wherein the multi-layer convolution network of the multi-layer convolution network module totally comprises four convolution blocks, the first two convolution blocks consist of two convolution layers and one pooling layer, and the second two convolution blocks consist of three convolution layers and one pooling layer; each convolution layer is connected with a ReLU activation function, the convolution layers uniformly adopt 3 multiplied by 3 convolution kernels, the receptive field of the convolution block is enlarged by stacking the convolution layers, and meanwhile, network parameters are reduced; the multi-layer convolution network not only reserves the position relation in the image, but also obtains the local characteristic of the image, the multi-layer convolution network module utilizes a serial network structure to extract the characteristics of the aurora image, reserves the spatial distribution relation of the aurora intensity in the aurora image and obtains the local characteristic of the image;
as shown in fig. 1, the multi-layer convolutional network module is composed of four convolutional blocks, each of the four convolutional blocks further comprises two parts of a convolutional layer and a pooling layer, the first two convolutional blocks are composed of two layers of the convolutional layer and one layer of the pooling layer, the last two convolutional blocks are composed of three layers of the convolutional layer and one layer of the pooling layer, the convolutional kernel size is 3×3, each convolutional layer is followed by a ReLU activation function, as shown in formula (1), and parameter settings of each layer of the convolutional network are shown in table 1:
（1）
table 1 multi-layer convolutional network layer-by-layer parameter settings
(4) Constructing a transducer coding network module to carry out global interaction on an aurora image so as to code the global characteristics of aurora eggs, wherein a coding network of the transducer coding network module consists of an embedded layer and a coding layer, the embedded layer divides the input aurora image into patches with fixed sizes, then the patch vectors are converted into embedded vectors with fixed lengths through linear mapping, and the embedded position codes are used for conforming to the input normal form of the transducer coding layer, so that the problem of non-uniform characteristic latitude is solved; the method comprises the steps that a transducer encoder of a transducer encoding network module consists of four identical encoding layers, each encoding layer sequentially consists of a layer normalization LN, a multi-head self-attention module MHSA, residual connection, a layer normalization LN, a multi-layer perceptron MLP and residual connection, the multi-layer perceptron MLP consists of two layers of convolution functions and a ReLU activation function, and the transducer encoding network module can conduct global interaction on images through a multi-head attention mechanism to encode image characteristics; the transform coding network module can model a large-scale space dependency relationship between the polar light image and geomagnetic station data, keep the attention weight of each layer and obtain the image global feature;
as shown in fig. 1, the transform coding network module is composed of an embedding layer and a coding layer, the embedding layer firstly cuts an input image into a patch with a fixed size of 16×16, and then converts the patch vector into an embedding vector with a fixed length through linear mapping, so that the embedding vector accords with an input normal form; when the input image is 224×224, the embedding layer ultimately generates an embedding vector dimension of 196×256.
The self-Attention mechanism is the core of the transform coding layer, and can be described as a process of mapping a query and a set of key-value pairs (key-value) to an output, given a set of query matrix Q, key K and value matrix V, the output matrix is calculated using the Attention function as follows:
（2）
wherein Q, K and V respectively represent a query, a key and a value sequence obtained by linear transformation of an input sequence; softmax represents normalizing the degree of association;scaling the dimensions is represented, where d represents the dimensions of the query and key.
Multi-headed attention is an extension of the attention mechanism, projecting queries, keys and values into different subspaces through a learnable k-set linear transformation, running k attention operations in parallel; the k attentive outputs are then concatenated to obtain the final output by a learnable linear transformation:
（3）
（4）
wherein,，/>，/>the parameter matrix of the linear transformation of the query, the key and the value can be simply understood as the segmentation operation in practice; />Is a parameter matrix of the final linear transformation of the multi-head attention mechanism, generally, < >>Often set to +>D represents the latitude of the input sequence, k represents the number of heads in the multi-head attention;
in order to preserve the position information of each patch, before the patch is sent to the transducer encoder of the transducer encoding network module, the transducer encoder of the transducer encoding network module is composed of four identical encoding layers, and each encoding layer module is composed of a layer normalization LN, a multi-head self-attention module MHSA, a multi-layer perceptron MLP, residual connection, and the like, as shown in fig. 2. The MLP consists of two layers of convolution functions and ReLU activation functions, FIG. 2And->The output characteristics of MHSA and MLP in the current layer are respectively represented, and the calculation process is as follows:
（5）
（6）
(5) Constructing a hierarchical feature fusion module, aggregating the attention weights among the hierarchies, screening the discriminant features of the coding layers, fusing the local feature representation and the global feature representation, capturing the information lost in the deep layer of the network, and obtaining the prediction features to enter a subsequent regression prediction module; the hierarchical feature fusion module is divided into feature aggregation between front and back layers and fusion of convolution block features and coding features, wherein the feature aggregation between the front and back layers is to fuse attention weights between the front and back layers and screen marks so as to extract coding layer discrimination features after hierarchical aggregation; the fusion of the convolution block features and the coding features is to fuse the convolution block features and the coding layer discriminant features after hierarchical aggregation to form fusion features; the hierarchical feature fusion module fuses the current layer attention weight and the previous layer attention weight through matrix multiplication, extracts the code layer discriminant features after hierarchical aggregation by using a Max function, fuses the convolution block features and the code layer discriminant features after hierarchical aggregation through matrix multiplication to form fusion features, and finally inputs a transform code layer again to obtain prediction features; the hierarchical feature fusion module firstly forces to aggregate distinguishing features among layers, compensates feature information lost due to network depth, then communicates the convolution blocks and the transform coding layers layer by layer, and fuses local feature representation of each convolution block with global features of the transform coding layers;
as shown in fig. 3, the hierarchical feature fusion module is divided into two steps of fusion, namely feature aggregation between the front and back hierarchies and fusion of convolution block features and coding features. Feature aggregation between layers is to fuse attention weights between layers and screen markers, and the attention weights of each layer can be expressed as:
（7）
（8）
wherein: k represents the number of self-attention heads in the multi-head self-attention mechanism in the model,Nrepresenting the dimension of the embedded vector.
To fully utilize the information between each layer, the first layer islLayer and the firstlThe attention weights of the-1 layer are fused through matrix multiplication, and maximum indexes A1, A2, … and AK of k different attention heads are selected by using a Max function so as to extract the discriminant features after hierarchical aggregationThe method comprises the following steps:
（9）
（10）
secondly, fusing the convolution block features and the aggregated coding layer discriminant features to form fusion features, aligning feature sizes and channel numbers by using an up/down sampling strategy based on embedded vectors, and normalizing aligned feature values by layers, wherein the fusion features can be expressed as:
（11）
wherein,Nis the dimension of the embedded vector, C represents the number of channels, equal to the depth of the embedded vector;
as shown in fig. 1, the features after four layers are fused are used as input and connected to the last transducer coding layer, so that not only the local information among layers is reserved, but also the attention weight among layers is fully utilized, and the local information lost in the deep layer is captured.
(6) Constructing a regression prediction module, expanding the final prediction features into vectors, inputting the prediction features into the regression prediction module, and outputting 12 geomagnetic station data; the regression prediction module consists of three layers of full-connection networks, including two hidden layers and one output layer, and the number of neurons of the three layers of full-connection networks is 1024, 1024 and 12 respectively; firstly, expanding the prediction features obtained by a transform coding layer into vectors, inputting the prediction features into a regression prediction module, and finally outputting 12 geomagnetic station data; the three-layer fully-connected network structure of the regression prediction module establishes a nonlinear mapping relation between aurora image characteristics and geomagnetic station data through two hidden layers and one output layer.
(7) Defining a loss function, and training a geomagnetic station data prediction model; and (3) training and verifying a geomagnetic station data prediction model consisting of a multi-layer convolution network module, a transform coding network module, a hierarchical feature fusion module and a regression prediction module by using the training set and the verification set constructed in the step (2). The model training process firstly loads weights of the VGG-16 model and the ViT-B/16 model which are officially pre-trained on the ImageNet21k to respectively initialize parameters of the multi-layer convolution network module and the transform coding network module, and then the model automatically optimizes trainable parameters of the whole prediction model under the constraint of a loss function. The training process used a standard ADAM optimizer with a learning rate set to 0.0002 and a batch size set to 24. If usingRepresenting the real sequence->Representing the predicted sequence, the loss function is defined as the mean square error between the geomagnetic station data predicted value and the true value, as shown in equation (12):
（12）
when the Loss of the verification set and the training set is not different and does not descend any more, the model training is completed; the training set Loss drops to 0.0048 and the validation set Loss drops to 0.0054.
(8) Predicting geomagnetic station data by using a trained geomagnetic station data prediction model, and analyzing and evaluating a prediction result; inputting the geomagnetic station data prediction model trained in the step (7) by using the test set constructed in the step (2) to obtain the prediction results of 12 geomagnetic station data; and combining the true value to perform error on the prediction resultAnalysis and evaluation, wherein the evaluation criteria comprise Root Mean Square Error (RMSE), average Relative Variance (ARV) and determination coefficient (R) 2 ；
（13）
（14）
（15）
The geomagnetic station data prediction model of the invention has good performance under each evaluation index, the root mean square error is kept within 10%, and the average relative variance is about 30%; the goodness of fit of the models is greater than 0.9.
Example 2
The invention discloses a geomagnetic station data prediction system of an aurora image, which comprises:
the original data acquisition module is used for acquiring polar ultraviolet image data of a satellite and geomagnetic station data of a ground station; acquiring regional ultraviolet aurora image data shot by an ultraviolet imager UVI from an SPDF website of a spatial physical data facility of NASA; acquiring 12 geomagnetic station data which are uniformly distributed along magnetic longitude and within the range of 60-80 degrees of magnetic latitude in the same time period from a world data center WDC, wherein the geomagnetic station data refer to minute-by-minute data of a geomagnetic field horizontal H component;
the data set constructing module is used for preprocessing the polar ultraviolet image data of the polar region to construct a data set and dividing a training set, a verification set and a test set; the method comprises the steps of preprocessing polar light image data of a polar region ultraviolet shot by UVI, reading the polar light image data by using a python script, obtaining imaging time and an image matrix of each single sample, and geographic longitude and latitude and geomagnetic longitude and latitude of the single sample, carrying out coordinate transformation on a polar light image, wherein a target coordinate system takes a geomagnetic pole as a center, the magnetic latitude ranges from 50 degrees to 90 degrees MLAT, the magnetic local time range is from 0 to 24MLT, and finally resetting negative value pixels of the image to 0 to eliminate satellite noise; the LBHL wave band 160nm-180nm image of the polar ultraviolet image data is screened out to be used as the input of a geomagnetic station data prediction model, and 12 geomagnetic station data in the same time period are output of the geomagnetic station data prediction model; the aurora image and geomagnetic station data are put into different folders to form a data set, and the data set is represented by 7:2:1 is divided into a training set, a verification set and a test set;
the multi-layer convolution network module is constructed to extract the local characteristics of aurora eggs, and the multi-layer convolution network of the multi-layer convolution network module totally comprises four convolution blocks, wherein the first two convolution blocks consist of two convolution layers and one pooling layer, and the second two convolution blocks consist of three convolution layers and one pooling layer; each convolution layer is connected with a ReLU activation function, the convolution layers uniformly adopt 3 multiplied by 3 convolution kernels, the receptive field of the convolution block is enlarged by stacking the convolution layers, and meanwhile, network parameters are reduced; the multi-layer convolution network not only reserves the position relation in the image, but also obtains the local characteristic of the image, the multi-layer convolution network module utilizes a serial network structure to extract the characteristics of the aurora image, reserves the spatial distribution relation of the aurora intensity in the aurora image and obtains the local characteristic of the image;
the method comprises the steps that a transducer coding network module is constructed, global interaction is carried out on an aurora image by the transducer coding network module to code the global characteristics of aurora eggs, a coding network of the transducer coding network module consists of an embedded layer and a coding layer, the embedded layer divides the input aurora image into patches with fixed sizes, patch vectors are converted into embedded vectors with fixed lengths through linear mapping, and position codes are embedded to conform to the input paradigm of the transducer coding layer, so that the problem of non-uniform characteristic latitude is solved; the method comprises the steps that a transducer encoder of a transducer encoding network module consists of four identical encoding layers, each encoding layer sequentially consists of a layer normalization LN, a multi-head self-attention module MHSA, residual connection, a layer normalization LN, a multi-layer perceptron MLP and residual connection, the multi-layer perceptron MLP consists of two layers of convolution functions and a ReLU activation function, and the transducer encoding network module can conduct global interaction on images through a multi-head attention mechanism to encode image characteristics; the transform coding network module can model a large-scale space dependency relationship between the polar light image and geomagnetic station data, keep the attention weight of each layer and obtain the image global feature;
the hierarchical feature fusion module is used for constructing a hierarchical feature fusion module, aggregating the attention weights among the hierarchies, screening the discriminant features of the coding layer, and fusing the local feature representation and the global feature representation to obtain the prediction features; the hierarchical feature fusion module is divided into feature aggregation between front and back layers and fusion of convolution block features and coding features, wherein the feature aggregation between the front and back layers is to fuse attention weights between the front and back layers and screen marks so as to extract coding layer discrimination features after hierarchical aggregation; the fusion of the convolution block features and the coding features is to fuse the convolution block features and the coding layer discriminant features after hierarchical aggregation to form fusion features; the hierarchical feature fusion module fuses the current layer attention weight and the previous layer attention weight through matrix multiplication, extracts the code layer discriminant features after hierarchical aggregation by using a Max function, fuses the convolution block features and the code layer discriminant features after hierarchical aggregation through matrix multiplication to form fusion features, and finally inputs a transform code layer again to obtain prediction features; the hierarchical feature fusion module firstly forces to aggregate distinguishing features among layers, compensates feature information lost due to network depth, then communicates the convolution blocks and the transform coding layers layer by layer, and fuses local feature representation of each convolution block with global features of the transform coding layers;
the regression prediction module is used for constructing a regression prediction module, expanding the final prediction characteristics into vectors, inputting the prediction characteristics into the regression prediction module, and outputting 12 geomagnetic station data; the regression prediction module consists of three layers of full-connection networks, including two hidden layers and one output layer, and the number of neurons of the three layers of full-connection networks is 1024, 1024 and 12 respectively; firstly, expanding the prediction features obtained by a transform coding layer into vectors, inputting the prediction features into a regression prediction module, and finally outputting 12 geomagnetic station data; the three-layer fully-connected network structure of the regression prediction module establishes a nonlinear mapping relation between aurora image characteristics and geomagnetic station data through two hidden layers and one output layer;
training a geomagnetic station data prediction model module, defining a loss function, training and verifying a geomagnetic station data prediction model consisting of a multi-layer convolution network module, a Transformer coding network module, a hierarchical feature fusion module and a regression prediction module by utilizing a training set and a verification set constructed in a data set constructing module, and if the geomagnetic station data prediction model is used for training and verifying the geomagnetic station data prediction model consisting of the multi-layer convolution network module, the Transformer coding network module, the hierarchical feature fusion module and the regression prediction moduleRepresenting the real sequence->Representing the predicted sequence, the loss function is defined as the mean square error between the geomagnetic station data predicted value and the true value, as shown in equation (12):
（12）
adjusting the super parameters for training to obtain a trained geomagnetic station data prediction model;
the geomagnetic station data prediction module predicts geomagnetic station data by using a trained geomagnetic station data prediction model, and analyzes and evaluates a prediction result; inputting a trained geomagnetic station data prediction model by using a test set constructed in a data set constructing module to obtain prediction results of 12 geomagnetic station data; and combining the true value to perform error analysis on the prediction result, wherein the evaluation criteria comprise Root Mean Square Error (RMSE), average Relative Variance (ARV) and decision coefficient (R) 2 ；
（13）
（14）
（15）
The geomagnetic station data prediction model of the invention has good performance under each evaluation index, the root mean square error is kept within 10%, and the average relative variance is about 30%; the goodness of fit of the models is greater than 0.9.

Claims (10)

1. A geomagnetic station data prediction method of an aurora image, comprising the steps of:
(1) Acquiring polar ultraviolet image data of a satellite region and geomagnetic station data of a ground station;
(2) Preprocessing the polar ultraviolet image data to construct a data set, and dividing a training set, a verification set and a test set;
(3) Constructing a multi-layer convolution network module to extract local characteristics of aurora eggs, wherein the multi-layer convolution network of the multi-layer convolution network module totally comprises four convolution blocks, the first two convolution blocks consist of two convolution layers and one pooling layer, and the second two convolution blocks consist of three convolution layers and one pooling layer;
(4) Constructing a transducer coding network module to carry out global interaction on an aurora image so as to code the global characteristics of aurora eggs, wherein a coding network of the transducer coding network module consists of an embedded layer and a coding layer, the embedded layer divides the input aurora image into patches with fixed sizes, then the patch vectors are converted into embedded vectors with fixed lengths through linear mapping, and the embedded position codes are used for conforming to the input normal form of the transducer coding layer;
(5) Constructing a hierarchical feature fusion module, aggregating the attention weights among the hierarchies, screening the discriminant features of the coding layers, and fusing the local feature representation and the global feature representation to obtain predicted features; the hierarchical feature fusion module is divided into feature aggregation between front and back layers and fusion of convolution block features and coding features, wherein the feature aggregation between the front and back layers is realized by fusing attention weights between the front and back layers and screening marks so as to extract coding layer discriminant features after hierarchical aggregation; the fusion of the convolution block features and the coding features is to fuse the convolution block features and the coding layer discriminant features after hierarchical aggregation to form fusion features;
(6) Constructing a regression prediction module, expanding the final prediction features into vectors, inputting the prediction features into the regression prediction module, and outputting geomagnetic station data;
(7) Defining a loss function, training and verifying a geomagnetic station data prediction model consisting of a multi-layer convolution network module, a transform coding network module, a hierarchical feature fusion module and a regression prediction module by using the training set and the verification set constructed in the step (2), and adjusting super parameters for training to obtain a trained geomagnetic station data prediction model;
(8) Predicting geomagnetic station data by using a trained geomagnetic station data prediction model, and analyzing and evaluating a prediction result; inputting the geomagnetic station data prediction model trained in the step (7) by using the test set constructed in the step (2) to obtain a geomagnetic station data prediction result; and combining the true value to perform error analysis and evaluation on the prediction result.
2. The geomagnetic station data prediction method of an aurora image of claim 1, wherein: in the step (1), region ultraviolet aurora image data shot by an ultraviolet imager UVI are acquired from an SPDF website of a spatial physical data facility of NASA; geomagnetic station data which are in the same time period, have magnetic latitude of 60-80 degrees and are uniformly distributed along magnetic longitude are obtained from a world data center WDC, and the geomagnetic station data refer to minute-by-minute data of a geomagnetic field horizontal H component.
3. The geomagnetic station data prediction method of an aurora image of claim 2, wherein: the method comprises the steps of (1) preprocessing polar ultraviolet image data of a region shot by UVI, reading the polar ultraviolet image data by using a python script, obtaining imaging time, an image matrix and geographic longitude and latitude and geomagnetic longitude and latitude of each single sample by using a cdf data file, carrying out coordinate transformation on a polar light image, wherein a target coordinate system takes a geomagnetic pole as a center, the magnetic latitude range is 50-90 degrees MLAT, the magnetic local time range is 0-24MLT, and finally resetting negative pixels of the image to be 0; the LBHL band 160nm-180nm image of the polar ultraviolet image data in the polar ultraviolet region is screened out and used as the input of a geomagnetic station data prediction model, and geomagnetic station data in the same period is output of the geomagnetic station data prediction model; the aurora image and geomagnetic station data are put into different folders to form a data set, and the data set is represented by 7:2: the scale of 1 is divided into a training set, a validation set and a test set.
4. A geomagnetic station data prediction method for an aurora image according to claim 3, wherein: and (3) each convolution layer in the step (3) is followed by a ReLU activation function, and the convolution layers uniformly adopt 3 multiplied by 3 convolution kernels.
5. The geomagnetic station data prediction method of an aurora image of claim 4, wherein: the transform encoder of the transform encoding network module in the step (4) consists of four identical encoding layers, each encoding layer sequentially consists of a layer normalization LN, a multi-head self-attention module MHSA, residual error connection, a layer normalization LN, a multi-layer perceptron MLP and residual error connection, and the multi-layer perceptron MLP consists of two layers of convolution functions and a ReLU activation function.
6. The geomagnetic station data prediction method of an aurora image of claim 5, wherein: in the step (5), the hierarchical feature fusion module fuses the current layer attention weight and the previous layer attention weight through matrix multiplication, extracts the coding layer discriminant features after hierarchical aggregation by using a Max function, fuses the convolution block features and the coding layer discriminant features after hierarchical aggregation to form fusion features through matrix multiplication, and finally inputs a transform coding layer again to obtain prediction features.
7. The geomagnetic station data prediction method of an aurora image of claim 6, wherein: the regression prediction module in the step (6) consists of three layers of fully-connected networks, wherein the three layers of fully-connected networks comprise two hidden layers and an output layer, and the number of neurons of the three layers of fully-connected networks is 1024, 1024 and 12 respectively; firstly, expanding the prediction features obtained by the transform coding layer into vectors, inputting the prediction features into a regression prediction module, and finally outputting geomagnetic station data.
8. The geomagnetic station data prediction method of an aurora image of claim 7, wherein: if in the step (7)Representing the real sequence->Representing the predicted sequence, the loss function is defined as the mean square error between the geomagnetic station data predicted value and the true value:
。
9. the geomagnetic station data prediction method of an aurora image of claim 8, wherein: the evaluation criteria in the step (8) include a root mean square error RMSE, an average relative variance ARV and a decision coefficient R 2 。
10. A geomagnetic station data prediction system for an aurora image, comprising:
the original data acquisition module is used for acquiring polar ultraviolet image data of a satellite and geomagnetic station data of a ground station;
the data set constructing module is used for preprocessing the polar ultraviolet image data of the polar region to construct a data set and dividing a training set, a verification set and a test set;
the multi-layer convolution network module is constructed to extract the local characteristics of aurora eggs, and the multi-layer convolution network of the multi-layer convolution network module totally comprises four convolution blocks, wherein the first two convolution blocks consist of two convolution layers and one pooling layer, and the second two convolution blocks consist of three convolution layers and one pooling layer;
the method comprises the steps of constructing a transducer coding network module, performing global interaction on an aurora image to code aurora ovum global characteristics, wherein a coding network of the transducer coding network module consists of an embedded layer and a coding layer, the embedded layer divides the input aurora image into patches with fixed sizes, then converts patch vectors into embedded vectors with fixed lengths through linear mapping, and embeds position codes to accord with an input paradigm of the transducer coding layer;
the hierarchical feature fusion module is used for constructing a hierarchical feature fusion module, aggregating the attention weights among the hierarchies, screening the discriminant features of the coding layer, and fusing the local feature representation and the global feature representation to obtain the prediction features; the hierarchical feature fusion module is divided into feature aggregation between front and back layers and fusion of convolution block features and coding features, wherein the feature aggregation between the front and back layers is realized by fusing attention weights between the front and back layers and screening marks so as to extract coding layer discriminant features after hierarchical aggregation; the fusion of the convolution block features and the coding features is to fuse the convolution block features and the coding layer discriminant features after hierarchical aggregation to form fusion features;
the regression prediction module is used for constructing a regression prediction module, expanding the final prediction characteristics into vectors, inputting the prediction characteristics into the regression prediction module and outputting geomagnetic station data;
training a geomagnetic station data prediction model module, defining a loss function, training and verifying a geomagnetic station data prediction model formed by a multi-layer convolution network module, a Transformer coding network module, a hierarchical feature fusion module and a regression prediction module by utilizing a training set and a verification set constructed in a construction data set module, and adjusting super parameters for training to obtain a trained geomagnetic station data prediction model;
the geomagnetic station data prediction module predicts geomagnetic station data by using a trained geomagnetic station data prediction model, and analyzes and evaluates a prediction result; inputting a trained geomagnetic station data prediction model by using a test set constructed in a data set constructing module to obtain a geomagnetic station data prediction result; and combine with realityError analysis is carried out on the prediction result by the value, and the evaluation criteria comprise Root Mean Square Error (RMSE), average Relative Variance (ARV) and decision coefficient (R) 2 。
CN202410081481.7A 2024-01-19 2024-01-19 Geomagnetic station data prediction method and system for aurora image Active CN117593666B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202410081481.7A CN117593666B (en) 2024-01-19 2024-01-19 Geomagnetic station data prediction method and system for aurora image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202410081481.7A CN117593666B (en) 2024-01-19 2024-01-19 Geomagnetic station data prediction method and system for aurora image

Publications (2)

Publication Number Publication Date
CN117593666A true CN117593666A (en) 2024-02-23
CN117593666B CN117593666B (en) 2024-05-17

Family

ID=89918833

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202410081481.7A Active CN117593666B (en) 2024-01-19 2024-01-19 Geomagnetic station data prediction method and system for aurora image

Country Status (1)

Country Link
CN (1) CN117593666B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118133888A (en) * 2024-04-29 2024-06-04 南京航空航天大学 Polar photoelectric current collecting index prediction method and system based on ultraviolet polar light image

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2022037575A (en) * 2020-08-25 2022-03-09 株式会社日立製作所 Image evaluation predicting device and method
CN116229295A (en) * 2023-02-28 2023-06-06 西安电子科技大学 Remote sensing image target detection method based on fusion convolution attention mechanism
CN116469100A (en) * 2022-12-27 2023-07-21 长春理工大学 Dual-band image semantic segmentation method based on Transformer
CN116630387A (en) * 2023-06-20 2023-08-22 西安电子科技大学 Monocular image depth estimation method based on attention mechanism

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2022037575A (en) * 2020-08-25 2022-03-09 株式会社日立製作所 Image evaluation predicting device and method
CN116469100A (en) * 2022-12-27 2023-07-21 长春理工大学 Dual-band image semantic segmentation method based on Transformer
CN116229295A (en) * 2023-02-28 2023-06-06 西安电子科技大学 Remote sensing image target detection method based on fusion convolution attention mechanism
CN116630387A (en) * 2023-06-20 2023-08-22 西安电子科技大学 Monocular image depth estimation method based on attention mechanism

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118133888A (en) * 2024-04-29 2024-06-04 南京航空航天大学 Polar photoelectric current collecting index prediction method and system based on ultraviolet polar light image

Also Published As

Publication number Publication date
CN117593666B (en) 2024-05-17

Similar Documents

Publication Publication Date Title
CN110443143B (en) Multi-branch convolutional neural network fused remote sensing image scene classification method
CN112991354B (en) High-resolution remote sensing image semantic segmentation method based on deep learning
CN112347888B (en) Remote sensing image scene classification method based on bi-directional feature iterative fusion
CN112329760B (en) Method for recognizing and translating Mongolian in printed form from end to end based on space transformation network
CN112507793A (en) Ultra-short-term photovoltaic power prediction method
CN113221641B (en) Video pedestrian re-identification method based on generation of antagonism network and attention mechanism
CN114092832B (en) High-resolution remote sensing image classification method based on parallel hybrid convolutional network
CN108090472B (en) Pedestrian re-identification method and system based on multi-channel consistency characteristics
CN111950453A (en) Optional-shape text recognition method based on selective attention mechanism
CN117593666B (en) Geomagnetic station data prediction method and system for aurora image
CN112950780B (en) Intelligent network map generation method and system based on remote sensing image
CN110751209A (en) Intelligent typhoon intensity determination method integrating depth image classification and retrieval
CN112488963A (en) Method for enhancing crop disease data
CN110599502A (en) Skin lesion segmentation method based on deep learning
CN116912708A (en) Remote sensing image building extraction method based on deep learning
CN115131313A (en) Hyperspectral image change detection method and device based on Transformer
CN117690178B (en) Face image recognition method and system based on computer vision
CN111242028A (en) Remote sensing image ground object segmentation method based on U-Net
CN115032602A (en) Radar target identification method based on multi-scale convolution capsule network
CN117351363A (en) Remote sensing image building extraction method based on transducer
CN116958700A (en) Image classification method based on prompt engineering and contrast learning
CN117173595A (en) Unmanned aerial vehicle aerial image target detection method based on improved YOLOv7
CN113313185B (en) Hyperspectral image classification method based on self-adaptive spatial spectrum feature extraction
CN114926691A (en) Insect pest intelligent identification method and system based on convolutional neural network
CN114529939A (en) Pedestrian identification method based on millimeter wave radar point cloud clustering and deep learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant