Extracting disaster location identification from social media images using deep learning

Social media platforms have recently played a predominant role in collecting and sharing reliable, timely information for disaster assessment and management. Social media information is available in either image or text format associated with their geographic location (geotagged), and meaningful information can be mined from these multimodal data to allow situational awareness and enhance decision-making during disasters. In order to extract geographic location from an image in which a disaster happens while geotagging information in conjunction with images on social media is not obtainable generally, our proposed framework is designed to extract geographic location from an image by using phone numbers and includes several key modules: image collection, phone number detection, and Google Maps API for location extraction. In this study, manually annotated multi-digit phone numbers dataset along with the Street View House Numbers (SVHN) dataset were used to train a convolutional neural network (CNN) based detection model (i.e., the RetinaNet object detection algorithm) to locate and detect the multidigit phone numbers from an image. Experimental results indicated that the detection model for phone number detection achieved more than 79% (0.79) Average Precision (AP) of all digits and a reasonable mean Average Precision (mAP) of 82% (0.8248) with an IoU (Intersection over Union) threshold of 0.5. Google Maps API can provide location information based on the phone numbers extracted from object detectors with less distortion in the distance. These results demonstrated the effectiveness of our proposed approach, and it can be utilized in any type of disaster event, such as earthquakes, flooding, wildfires, etc., for improving situation awareness, disaster assessment, and management.