Automatic recognition of poleward moving auroras from all-sky image sequences based on HMM and SVM

We present an automatic method to recognize the poleward moving auroras (PMAs) from all-sky image sequences. A simplified block matching algorithm combined with an orientation coding scheme and histogram statistics strategy was utilized to estimate the auroral motion between interlaced images. An all-sky image sequence was first modeled by hidden Markov models (HMMs) and then represented by HMM similarities. The imbalanced classification problem, i.e., non-PMA events far outnumbering PMA events, was addressed by the metric-driven biased support vector machine (SVM). The proposed method was evaluated using auroral observations in 2003 at the Chinese Yellow River Station. Five days observations were manually labeled as PMA or non-PMA events considering both the keogram and all-sky image information. The supervised classification experiments were carried out and achieved satisfactory results. We further detected PMAs from auroral observations in the remaining days and the resultant double-peak occurrence distribution was compared with that of the well-known poleward moving auroral forms (PMAFs).