← Back to Publications List

Deep Learning-Based Early Warning System for Natural Disaster Prediction in Bangladesh

Students & Supervisors

Student Authors
Arizit Chaki Artha
Bachelor of Science in Computer Science & Engineering, FACULTY OF SCIENCE & TECHNOLOGY
Ekramul Hasib
Bachelor of Science in Computer Science & Engineering, FACULTY OF SCIENCE & TECHNOLOGY
Sazid Mohsin
Bachelor of Science in Computer Science & Engineering, FACULTY OF SCIENCE & TECHNOLOGY
S.f Sheikh Saadi
Bachelor of Science in Computer Science & Engineering, FACULTY OF SCIENCE & TECHNOLOGY
Supervisors
Md Mortuza Ahmmed
Associate Professor, Faculty, FACULTY OF SCIENCE & TECHNOLOGY

Abstract

Bangladesh is highly exposed to natural disasters, and improving early warning systems is vital for reducing disaster risk and strengthening national preparedness. With the increasing availability of digital data, deep learning methods have become promising tools for forecasting disaster-related events. This study analyzed long-term trends in the performance of deep learning models by analysing yearly changes in model accuracy, prediction error, and training data size. The goal was to understand how data quantity and model behaviour develop over time and how these factors influence the overall reliability of AI-based early warning systems for Bangladesh. This study used a quantitative, data-driven approach based on yearly records of model accuracy, prediction error, and training data size. The dataset was pre-processed, normalized, and divided into training and testing subsets. An LSTM-based deep learning architecture was used due to its strength in modeling sequential data. The model was trained using the Adam optimizer and Mean Squared Error loss function. Correlation and performance analyses were also conducted to evaluate the influence of data size and error trends on overall accuracy. The results show that AI model accuracy increased steadily over the examined years, indicating better improvement in deep learning-based disaster prediction. Larger training data volumes strongly correlated with higher model accuracy. However, prediction error values fluctuated significantly, suggesting instability and possible overfitting in certain years. Overall findings confirm that deep learning improves early disaster prediction performance in Bangladesh, although further refinement is needed to reduce uncertainty. Overall, the study proves that deep learning is much better adapted to the task of predicting disasters in the context of Bangladesh. The results indicate that the more data used in the process of training the model, the better the precision achieved. However, the discrepancies in the rate of error point towards the requirement for more stable methods.

Keywords

Deep Learning Early Warning Systems Natural Disaster Prediction LSTN Model Bangladesh AI

Publication Details

  • Type of Publication:
  • Conference Name: 2nd International Conference on Frontiers in Science (2nd ICFS:ITGI)
  • Date of Conference: 15/01/2026 - 15/01/2026
  • Venue: Bangladesh University of Engineering and Technology (BUET), Dhaka, Bangladesh
  • Organizer: Faculty of Science, BUET