Research On Auroral Substorm Detection And Space Physical Mechanism

Aurora substorm is the primary reflection of strong disturbance in the solar-earth system.The breakup of the substorm has a close relationship with a sudden electromagnetic energy release during solar wind-magnetosphere coupling process.The research on the mechanism of substorm onset and expansion phase contributes to learning the interaction among interplanetary magnetic field,magnetosphere and ionosphere.In addition,the research on substorm is essential for understanding the process of flux transport from the solar to earth,which is significant to the space weather forecast.As we all known,plenty of substorm models have been proposed to explain the cause of substorm occurrence.However,there is no unified conclusion to explain the mechanism of substorm onset and expansion phase because of the complex and variable auroral substorm phenomena.Thus,how to analyze the substorm mechanism based on different observations is still an urgent problem to be solved.The auroral images from the ultraviolet imager(UVI)and ionospheric convection data are the most important data to record substorms.The complete auroral oval and the substorm bulge features are displayed clearly in UVI images.Specifically,an explosive increase in auroral intensity followed by zonal and poleward expansion of auroral bulge could be observed in UVI images.For ionospheric plasma flow,ionospheric convection exhibits synchronous spatial-temporal features during substorm process.Thus,this dissertation focuses on the substorm event detection and physical mechanism analysis using UVI images and ionospheric convection data.The existing substorm detection algorithms are usually based on manually designed features and rules.The dynamical ionosphere convection patterns are difficult to simulated on fewer convection velocity data.These problems lead to the limitation of substorm research.In order to solve these problems,the 3D neural networks are applied to extract spatiotemporal features of substorm events.Furthermore,some neural networks are used to the substorm detection,ionospheric convection modeling,convection velocity map completion and multimodal substorm classification.The main contributions of the dissertation are summarized as follows:1.Four relevant datasets are constructed,including VIPSL-UVISet120 K image dataset with its corresponding space physical parameters dataset named VIPSL-Space120 K,VIPSLFlow5M ionospheric convection dataset with its corresponding space physical parameters dataset named VIPSL-Space5 M.The UVI images in VIPSL-UVISet120 K dataset are collected from Polar ultraviolet imager.The VIPSL-UVISet120 K and VIPSL-Space120 K datasets are used in the substorm events detection and classification.Two-dimensional convection velocities in VIPSL-Flow5 M dataset are merged by line-of-sight velocities from high-frequency coherent scatter radars in SuperDARN.The VIPSL-Flow5 M and VIPSLSpace5 M datasets are utilized in the ionospheric convection modeling and convection velocity map completion.2.The substorm detection algorithm guided by dual-clustering is proposed for automatic substorm detection.To avoid using inefficient handcraft features,3D convolution network with subspace clustering is proposed to extract spatiotemporal features of UVI image sequences.Because of the different imaging angles among frames,the UVI images are transformed into MLAT-MLT(magnetic latitude-magnetic local time)coordinate for pixel alignment.To decrease the noise rate of UVI images,image level clustering is applied to reserve the substorm bulge and discard the unimaged areas.The experimental results show that the performance of the proposed method achieves higher recall and accuracy compared with other methods.3.The ionospheric convection velocity model based on global grid map(ICVMGM)is proposed for modeling the correlation between space physical parameters and convection velocities.And the model is applied to analyze the rules of dynamic convection patterns without the restrictions of space parameters values.Firstly,several nonlinear models are constructed in each grid cell of the global grid map based on ionospheric convection velocities and space physical parameters.Then,the global convection velocity model could be achieved by combining local models.Finally,it can be used to analyze the ionospheric convection pattern with the global convection velocity model.Subjective experimental results show that the interplanetary magnetic field has a great influence on the shape and number of convection cell.4.The weakly supervised convection velocity generative model using partial convolutional generative adversarial network is proposed for convection velocity map completion.In order to increase the proportion of observed convection velocities,global grid map is partitioned into magnetic grids based on magnetic latitude and magnetic local time.Base on this,the duty ratio of the observation data in the magnetic grids can be improved effectively.The predicted velocities from ICVMGM(Section 2)are used as the reference dataset in training phase.Because of the unbalanced distribution of convection velocities,the network cannot effectively learn the true distribution of convection velocities in the magnetic grid.Partial convolutional layers are utilized to concentrate convolution kernel on valid convection velocities for feature extraction.Experimental results show that the missing velocities in global convection velocity map could be generated well using the proposed method.5.The multimodal substorm classification using memory fusion network is proposed to classify the auroral substorms accurately.Traditional substorm classification methods are usually based on single modality data.The multimodal sequential feature based on UVI images,convection velocities and space parameters are extracted using a memory fusion network for substorm classification.The objective experimental results of different modality embedding show that the auroral particle precipitation has significant effect on substorms.The experimental results of different parameter delays demonstrate that the synchronized space physical parameters are strongly associated with substorms.Most of the onsets occurred equatorward of the convection reverse boundary.The distribution of onset locations show that te origin of substorms may be located deep within magnetosphere.Furthermore,the orientation of the interplanetary magnetic field keeps northward for a large proportion of substorms.In conclusion,this dissertation is a cross-research on computer vision and space physics.The dissertation focusses on some auroral substorm researches including auroral substorm detection,ionospheric convection modeling,convection velocity completion and multimodal substorm classification.Then,the physical mechanism of the substorm was summarized using multi-modal data.In addition,this dissertation proposes a new way for applying artificial intelligence into space physics in depth.