Natural disasters frequently occur unexpectedly and seriously threaten human safety and infrastructure. Traditional detection systems rely heavily on IoT sensors and satellite monitoring, which are often costly and less accessible in resource-limited or remote areas. In contrast, social media provides a rich and real-time source of information, as users frequently post eyewitness reports during disaster events. However, automatically classifying these posts into relevant disaster categories remains challenging due to the short and informal nature of the text. The research aims to develop a high-performing classification model for disaster-related tweets using graph-based neural architectures and structured word embedding representations. The method used is a comparative implementation of Graph Convolutional Network (GCN) and Graph Attention Network (GAT) models, with input constructed by concatenating vectors from three word embedding techniques—Word2Vec, FastText, and GloVe—across seven multilingual datasets. The result of this study is that GAT outperformed GCN in all scenarios, with FastText embeddings yielding the highest individual performance. In contrast, combined embeddings sometimes led to performance degradation due to redundancy. The average F1-score for GCN is 0.749, while GAT achieves 0.915. The research conclusions indicate that GAT with word embedding input provides a novel and effective multilingual disaster tweet classification framework, offering valuable insights for future AI-based natural disaster monitoring systems.
Faisal et al. (2026) studied this question.