Prediction Of Large-scale Evolution Of Photospheric Magnetic Field In Solar Active Region Based On Fusion Of Spatiotemporal Features

The ability to predict and reason about future results is a key component of an intelligent decision-making system.In recent years,machine learning prediction algorithms represented by deep learning have achieved rapid development and are widely used in weather forecasting,natural disaster warning,disease diagnosis,and other fields.However,in the research of the evolution of the solar photospheric magnetic field,due to the lack of standard data sets and comprehensive evaluation methods,and the interweaving of various complex changes in the evolution process,such as the movement,deformation,emergence,and disappearance of the magnetic structure,it is a great challenge to directly predict the evolution of the magnetic field in time and space.Therefore,there is currently no work to predict the evolution of the solar magnetic field.The image sequence formed by the evolution of the solar photospheric magnetic field contains more complicated spatiotemporal physical relationships.It is not only closely related to the various magnetic activities erupting in the solar atmosphere,such as sunspots,solar wind,and coronal mass ejection,but also affects the near-Earth space environment of our planet,climate change,and daily life.To this end,we analyze and research the prediction of the magnetic field evolution process based on spatiotemporal feature fusion,and conduct a series of studies and explorations from preprocessing data sets,building the network,comprehensively evaluating prediction results,and enhancing the visual effects of prediction results.Our main work is as follows:1.We produce a complete data set of magnetic field evolution.According to the characteristics of magnetic field evolution data provided by SDO/HMI,we select the active area covering the northern and southern hemispheres of the sun from 2011 to2015 and remove the influence of the projection effect,and preprocess them through various methods such as clipping the intensity and cropping the active area and finally produce a relatively complete magnetic field evolution sequence data set,which is published on the open-source website.2.We propose a large-scale and short-term prediction algorithm for the evolution of the solar magnetic field based on the fusion of temporal and spatial features.The algorithm is based on the existing conventional video prediction models.First,the LSTM module with spatiotemporal memory is used to fully fuse and characterize the related features of the magnetic field image sequence in time and space,and then the LSTM module with differential spatiotemporal memory is used to deal with the complex and non-deterministic changes in the evolution process,and finally the prediction of the magnetic field evolution is realized by a cyclic network structure.Besides,considering the physical properties of magnetic field data,we use the correlation coefficient,structural similarity,root mean square error and other indicators and methods to analyze and discuss the prediction results from the aspects of large-scale magnetic field structure,motion speed,magnetic field emergence,fine structure deformation and the evolution of magnetic neutral line,and verify the effectiveness of the algorithm in predicting the large-scale and short-term evolution process of the magnetic field.3.We propose a restoration algorithm for degraded images mixed with several types of degradations to enhance the prediction results of magnetic field evolution in visual effects.The algorithm directly fuses the features of different receptive field branches to enhance the structure of the restored image;uses the attention mechanism to dynamically fuse features at different hierarchies to increase model adaptability and reduce model redundancy;combines ₁L loss and perceptual loss,enhances the visual perception of the restored images.Experimental results on data sets such as DIV2K and BSD500 show that the proposed algorithm achieves better results both in terms of quantitative analysis of PSNR and SSIM and in terms of subjective visual quality.The experimental results on the magnetic field prediction results show that this algorithm can improve the visual effect of the magnetic field prediction results.