Enhancing Natural Disaster Monitoring: A Deep Learning Approach to Social Media Analysis Using Indonesian BERT Variants

Karlina Elreine Fitriani (Lambung Mangkurat University)
Mohammad Reza Faisal (Lambung Mangkurat University)
Muhammad Itqan Mazdadi (Lambung Mangkurat University)
Fatma Indriani (Lambung Mangkurat University)
Dodon Turianto Nugrahadi (Lambung Mangkurat University)
Septyan Eka Prastya (Sari Mulia University)

Indonesian Journal of Electronics, Electromedical Engineering, and Medical Informatics
Vol. 7 No. 1 (2025): February

Social media has become a primary source of real-time information that can be leveraged by artificial intelligence to identify relevant messages, thereby enhancing disaster management. The rapid dissemination of disaster-related information through social media allows authorities to respond to emergencies more effectively. However, filtering and accurately categorizing these messages remains a challenge due to the vast amount of unstructured data that must be processed efficiently. This study compares the performance of IndoRoBERTa, IndoRoBERTa MLM, IndoDistilBERT, and IndoDistilBERT MLM in classifying social media messages about natural disasters into three categories: eyewitness, non-eyewitness, and don’t know. Additionally, this study analyzes the impact of batch size on model performance to determine the optimal batch size for each type of disaster dataset. The dataset used in this study consists of 1000 messages per category related to natural disasters in the Indonesian language, ensuring sufficient data diversity. The results show that IndoDistilBERT achieved the highest accuracy of 81.22%, followed by IndoDistilBERT MLM at 80.83%, IndoRoBERTa at 79.17%, and IndoRoBERTa MLM at 78.72%. Compared to previous studies, this study demonstrates a significant improvement in classification accuracy and model efficiency, making it more reliable for real-world disaster monitoring. Pre-training with MLM enhances IndoRoBERTa’s sensitivity and IndoDistilBERT’s specificity, allowing both models to better understand context and optimize classification results. Additionally, this study identifies the optimal batch sizes for each disaster dataset: 32 for floods, 128 for earthquakes, and 256 for forest fires, contributing to improved model performance. These findings confirm that this approach significantly improves classification accuracy, supporting the development of machine learning-based early warning systems for disaster management. This study highlights the potential for further model optimization to enhance real-time disaster response and improve public safety measures more effectively and efficiently.