Auroral Events Detection And Analysis Based On ASI And UVI Images

An aurora is a natural light display in the sky particularly in the high latituderegions, caused by the collision of energetic charged particles with atoms in the highlatitude atmosphere. The spatial structure and temporal evolution of auroral luminosityare ascribed to the cumulative effects of the solar wind–magnetosphere interaction andthe physics of the magnetosphere–ionosphere interaction. Therefore, much informationof the magnetosphere and the ionosphere in solar-terrestrial space can be obtained aftersystematic observation and deep analysis to auroras, which helps us to study the wayand extent of the sun activities affecting the earth and it is very important to learnabout the behave of space weather.Auroral research has been one of the important topics in space physics worldwide,and the comprehensive observation to aurora has become significant in polar adventure.Optical imaging is the earliest applied and most popular auroral observation tools, withthe merit that it can obtain the2-D auroral images and can continually observe auroralspacial motion characteristics. At present, the frequently-used optical observationsinclude auroral images obtained by all-sky imagers (ASI) and ultra-violet imagers(UVI). With the systematic observation to aurora, massive auroral images have beenincreasing each year, so how to efficiently utilize the huge auroral data is an urgentproblem to be solved for auroral researchers in different countries.Traditional auroral image analysis has always been made manually, which is notonly difficult and time-consuming, but also the conclusions drawn from the normalcase study are lack of universality and block the popularity. To address this problem,we try to automatically analyze the massive auroral data. Based on the high resolutionASI images observed in Yellow River Station, we focus on auroral morphologyclassification and poleward moving auroral events detection, and based on the globalUVI images observed by Polar satellite, we aim at substorm onset detection andmodeling auroral oval boundary locations. These are significant topics in auroralresearch and we try to analyze them from a perspective of computer learning, which isa fresh attempt in auroral research. A few methods suited for the representation,classification and detection of auroral image data have been proposed in this paper. Insummary, the author’s major contributions are outlined as follows:1) Because the naturally occurring auroras are a dynamically evolving process,we propose to study auroral morphology classification based on ASI imagesequence. The uniform local binary patterns (uLBP) are employed to describe the2-D space structures of ASI images, and then the extracted uLBPsequences are modeled by hidden Markov models (HMM). Thus this methodcan characterize both temporal and spatial information of auroras meanwhile.The velocity of auroral evolution is various, and the length of auroralsequences is different. We present an affine log-likelihood normalizationtechnique to manage the sequences with different lengths. The proposedmethod is used in the automatic recognition of four primary categories of ASIimages. Compared to the frame-based method, the supervised classificationresults of our method achieve higher accuracies and lower rejection rates, andthe occurrence distributions of the four auroral categories further illustrate thevalidity of the proposed method on auroral representation and classification.2) We present an automatic method to recognize the poleward moving aurorasfrom all-sky image sequences. A simplified block matching algorithmcombined with an orientation coding scheme and histogram statistics strategywas utilized to estimate the auroral motion between interlaced images. Anall-sky image sequence was first modeled by HMM models and thenrepresented by HMM similarities. The imbalanced classification problem, i.e.,non-poleward moving auroral events far outnumbering poleward movingauroral events, was addressed by the metric-driven biased Support VectorMachine (SVM). The experimental results show that the occurrence rules ofthe poleward moving auroral events calculated by the proposed automaticmethod are the same with the existing manual statistical results.3) Substorm research largely depends on the precise definition and the timingaccuracy of the substorm onset used in various observations. At present, theglobal UVI images are widely accepted as the best tool from which to learnsubstorm. However, the existing studies are all based on the manualrecognition of substorm onsets. We propose to automatically identify auroralsubstorm onset timing from UVI images. We first transformed the originalUVI images into the MLT-MLAT rectangular coordinate system, and thebright bulge was determined with the spatial fuzzy c-means method. Theproposed technique was tested using Polar UVI observations acquired fromthree months of the winter season in1996-1997. The identified substorm onsetresults were compared with the available manual statistical report, and theexperimental results demonstrate the proposed technique can efficientlyrecognize those auroral events with substorm features. 4) The location of auroral oval boundary gives us valuable information aboutmagnetosphere processes. Different from traditional study to investigate thevariation of auroral oval boundaries with a certain magnetic index, we proposeto consider their variation from the source, i.e., the relationship betweenauroral boundaries and solar wind plasma and the interplanetary magneticfield. Auroral oval boundaries were automatically identified from UVI imagesof three winters. Based on the big data, at each one-hour MLT sector, westatistically analyze the response of oval boundaries to various solar-windmagnetic field conditions by using multivariate regression technique. Thepredictions were compared to the actual magnetic latitude with the MeanAbsolute Deviation (MAD) as an evaluator. The average MAD is about1.65magnetic degrees. The high fitness between predictions and observationsdemonstrates that the presented models can be used for prediction of auroraloval.