Transfer Learning Approaches for Emergency Event Detection Through Crowd sourced Social Media Imagery

Abstract

Social media is platform where people share their day today activities and information flows in social networks like whirlwind. Emergency events are very critical information which needs to be passed to required people so that they can take necessary actions. Data that flows in social media is crowdsourced using different people, most of the time the disaster or emergency event information flows not only in the closed group but it disseminates in open space. Data that flows here can be used to extract important information like what is the emergency situation and how this information can be disseminated to help the people and send it to people who uses it to provide the response activity. Most of the works are related to identifying the emergency using tweets posted, satellite images but all these techniques have many challenges which can be overcome by using crowdsourced images by applying transfer learning techniques. Work has used the image dataset which has been collected from social networks by crowdsourcing. Transfer learning methods used for conduction of this experiment is VGG16, Resnet50, Inception V3 out of all the models the Resnet50 out performed all the other model with very good accuracy of 94