Deep neural networks for detection of solar corona mass ejections

Deep neural networks for detection of solar corona mass ejections is a work in which you tried to classify the Coronal Mass Ejections. The CME is a significant release of sun matter and electromagnetic radiation during solar eruption. Because of the high energy that is sometimes carry to the Earth by the solar wind it is useful to classify these events in order to take precautions to safeguard astronauts health and satellites integrity around the Earth. Today this type of classification is done manually: the purpose of this thesis is to try to give a tool able to detect these events automatically by analyzing the solar photos taken by coronographs orbiting the Earth. In order to achieve this aim, Deep Learning techniques have been used; in particular you tried to understand if you could handle that data with a space-time approach. The data processing involved several selection processes in which many of these images were deleted as they did not respect the temporal and dimensional parameters chosen as inputs for the neural network. At this point sequential analysis and choice of the best trade-off between elements in sequence and number of sequences obtained is done. By doing this you were able to mark the sequences that produce a CME and those that do not produce a CME. Due to the continuous resizing of the data, different datasets have been created based on the quality of the events themselves. Once the final data were obtained, you proceeded with the definition of neural network models: a first approach was to create a network able of extracting spatial features from image sequences, which they are handle by the temporal part of the network that produces the classification signal. This type of network did not produce interesting results on the different datasets created. To understand if the space-time approach was right, a new model was defined that simultaneously extracted the two types of information. This model produced results on the dataset created with the best quality data; it was able to correctly classify about two-thirds of the testing set. In conclusion, it is therefore possible in the future to improve this result if the quality and amount of data grows.