Applying GNN Models for Diverse Disaster Detection using Temporal Knowledge Graphs

As unexpected disasters increase, the number of casualties and economic damages are increasing. Accordingly, efforts to collect and process data have been made to predict and respond to disasters. However, because the data collected from a certain disaster is enormous and diverse, it is difficult to identify an exact disaster type and its situations at the early stage of a disaster. To that end, in this paper, we first classify disasters into six categories according to their characteristics and extend our ontology-based temporal knowledge graphs to contain these characteristics. Finally, to detect a disaster from temporal knowledge graphs, Graph Neural Networks (GNN) or other deep learning techniques can be useful. For the evaluation, we selected four disasters belonging to six categories and constructed temporal knowledge graphs for each disaster. Then, to see how quickly a disaster can be detected from the constructed graphs, we tested three GNN models, including Graph Convolutional Network (GCN), SageConv, and Graph Attention Network (GAT). Our experimental results show that temporal disaster knowledge graphs can accurately represent the characteristics of various disasters, enabling the detection of disasters from heterogeneous data collected at disaster sites.