MMBDisasterSeverity: A Multimodal Multitask Dataset for Disaster Severity Detection in Bangla Language
Description
The MMBDisasterSeverity dataset is a large-scale multimodal multitask benchmark developed for disaster severity assessment in Bangla, addressing a critical gap in disaster informatics. While existing disaster-related datasets primarily focus on disaster detection/classification, they largely overlook the assessment of disaster severity, despite its crucial role in emergency response, resource allocation, and humanitarian decision-making. To bridge this gap, we extend the BanglaCalamityMMD dataset by introducing a fine-grained severity annotation for multimodal disaster analysis. The original BanglaCalamityMMD dataset consists of 7,903 image-text pairs collected from social media platforms, covering eight disaster-related categories: Landslides, Wildfire, Tropical Storm, Drought, Flood, Earthquake, Human Damage, and Non-Disaster. Building upon this resource, we manually annotate each disaster-related instance with one of five severity levels: Minimal, Mild, Moderate, Severe, and Catastrophic. To the best of our knowledge, this is the first Bangla multimodal dataset to incorporate disaster severity annotations, enabling research beyond disaster identification toward fine-grained impact assessment. The dataset is divided into three standard splits to facilitate fair model development and evaluation. The training set contains 6,323 instances, the validation set contains 790 instances, and the test set contains 790 instances. The severity label distribution is intentionally preserved across the splits while reflecting the natural imbalance of real-world disaster events. In the training set, there are 2,108 Moderate, 1,652 Severe, 1,081 Mild, 1,017 Minimal, and 465 Catastrophic instances. The test set contains 240 Severe, 213 Moderate, 149 Mild, 107 Minimal, and 81 Catastrophic samples. The validation set includes 243 Severe, 206 Moderate, 165 Minimal, 100 Mild, and 76 Catastrophic instances. Unlike conventional disaster detection datasets that answer only whether a disaster has occurred, MMBDisasterSeverity enables models to estimate the degree of disaster impact by jointly leveraging textual and visual information. The dataset supports a wide range of multimodal research tasks, including disaster severity prediction, multimodal representation learning, impact-aware emergency response systems, humanitarian AI, and multimodal disaster intelligence for low-resource languages. By introducing a novel severity annotation layer for Bangla social media data, the dataset provides a valuable benchmark for developing robust multimodal multitask models capable of understanding both the presence and the intensity of disaster events, ultimately contributing to more informed and timely disaster management.
Files
Institutions
- International Islamic University ChittagongChittagong, Chittagong