| dc.description.abstract | All-Sky Imagers located in the Arctic and Antarctic regions capture images of the sky at regular intervals throughout the winter season. Data from the last decades make up millions of images where auroral researchers have no way of filtering the data without time-consuming manual investigation. We implemented the convolutional neural network family called EfficientNet for automatic classification of All-Sky Imager data. We manually labeled 7,980 images from Ny-Ålesund, Svalbard, into classes based on the appearance of aurora or no visible aurora. As a goal were to classify different aurora shapes, we used the 3 classes arc, diffuse and discrete, while images without detectable aurora were classified as no aurora. We found that EfficientNet successfully detected aurora in All-Sky Imager data. Training several EfficientNet models with various hyper-parameters, the highest performing model achieved an classification accuracy of 88\% on unseen test data. By aggregating the 3 aurora classes, we archive an binary classification accuracy of 96\% on the same test data. The methods shown in this thesis can be applied to data from any auroral All-Sky Imager. We created a data set of 665,865 unlabeled Ny-Ålesund all-sky images (5577 Å and 6300 Å for the same time periods for 2014, 2016, 2018 and 2020), and matched each image to approximate solar wind parameters from NASA's OMNI data. Our model were applied to the data set, and statistical results show that variations in solar wind speed and IMF $B_z$ do not determine the observed aurora shape. Further, our classifier labeled more images as diffuse for the 6300 Å emission line (red aurora), which indicates good predictions. This was expected, as red aurora is a weaker, more diffuse form of aurora. Statistics were also made based on an hourly distribution, where we could observe dayside and nightside aurora. During polar nights, Svalbard is optimal for observing dayside aurora, but we found that the location is probably to high north to observe stronger nightside aurora events, like substorms. Our results for the hourly distributions indicate a (weak) double-peak feature for dayside aurora, which have been observed before \cite{https://doi.org/10.1029/95JA03147}\cite{https://doi.org/10.1029/97JA02638}\cite{HU2009794}. The feature is strongest for 2020, with one peak around 6-8 local time and a second peak around 14-15 local time (for discrete aurora). Our time points (in magnetic local time) do not match the previous observations, but are +3 hours shifted. | nob |