Techniques for automated classification of nighttime ionospheric images

Ionosphere;classification;airglow;dimensional reduction;equatorial plasma bubble (EPB);medium-scale traveling ionospheric disturbance (MSTID);conditional random field (CRF);Fisher discriminant analysis (FDA);Local Fisher discriminant analysis (LFDA)

Gehrels, Thomas ; Makela ; Jonathan J.

关键词: Ionosphere;    classification;    airglow;    dimensional reduction;    equatorial plasma bubble (EPB);    medium-scale traveling ionospheric disturbance (MSTID);    conditional random field (CRF);    Fisher discriminant analysis (FDA);    Local Fisher discriminant analysis (LFDA);   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/44234/Thomas_Gehrels.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

A large database of images of the night sky above Hawaii and Chile is being collected in order to study ionosphericstructures. The images of interest are those that contain equatorial plasma bubbles (EPB) or medium-scale traveling ionospheric disturbances (MSTID). However the majority of the images collected contain neither EPBs nor MSTIDs, or are contaminated by other light sources or clouds. In order to identify the images of interest, discriminative classification models are considered for determining the relationship between measured image features and labels for a sequence of images. To provide features that enable this modeling, the use of texture, difference, object motion, correlation, and long-term variation measurements are explored. Additionally, the Local Fisher Discriminant Analysis (LFDA) algorithmis considered as a means to reduce the computational complexity of the classification process through dimensional reduction.It was found that a conditional random field (CRF) model provides the best classification accuracy.Accuracies of 80% - 90% were achieved for classification of EPBs, clear images and cloudy images. Classification of MSTIDs had accuracy of 65%, possibly due to the limited size of the test and training sets. The LFDA technique for dimensional reduction of the feature vector proved effective. When the classification accuracy using this data was compared to that which was dimensionally reduced using principal component analysis (PCA), the LFDA performed equally well or better for all feature vector lengths tested.