Abstract
Disaster-related misinformation on social media can significantly disrupt public understanding, weaken trust in official information, and hinder emergency response operations, creating an urgent need for reliable automated detection systems. This review examines machine learning approaches for detecting misinformation in crisis environments and analyzes trends in datasets, modeling strategies, and evaluation methodologies. The surveyed studies are categorized into four major groups: traditional machine learning models using handcrafted lexical and temporal features, deep learning architectures including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and transformers, graph-based approaches exploiting propagation and social interaction patterns, and multimodal frameworks integrating textual, visual, and metadata. The review additionally evaluates publicly available disaster-related and misinformation datasets, emphasizing the influence of annotation quality, label granularity, and class imbalance on model performance. Comparative analysis indicates that transformer and multimodal methods often achieve superior results, although performance commonly declines under cross-disaster and cross-platform conditions. Finally, the paper discusses major challenges including explainability and real-time deployment, while outlining promising future research directions.