Machine Learning for Crowd-Sourcing a Social Media Data Source to Improve Response and Recovery After the Earthquake Disaster

In this paper, the methodology of detecting rescue messages extracted from social media data is presented. Rescue messages were originated after an earthquake, they are tweets that may also deliver information about position and time. A massive amount of social media data has been extracted after the two earthquake disasters of magnitude of Mw 7.7 and Mw 7.6 occurred on February 6, 2023 in Turkiye. The procedure of manual labelling and automated labelling is presented. For labeling purposes, nine BERT language models, which are based on attention and transformers, were used. The supervised learning methods were applied to assess the precision of the labels and perform classification. Furthermore, the dataset was processed with deep learning methods: Convolutional Neural Networks, Deep Neural Networks, and Long Short-Term Memory. Accuracy of data toward detection of rescue and non-rescue tweets is compared. Keywords are extracted to determine hazard situations and emergency needs toward coordination purposes including spatio-temporal information when provided by tweets. Deep learning and BERT models detect rescue and non-rescue classes assuring a level in 0.8972 and 0.9808 in recall, respectively.