Advanced Hydro-Informatic Modeling through Feedforward Neural Network, Federated Learning, and Explainable AI for Enhancing Flood Prediction

Citation

Mahir, Shahariar Hossain and Tashrif, Md Tanjum An and Karim, Md Ahsan and Kundu, Dipanjali and Rahman, Anichur and Hamza, Md. Amir and Farid, Fahim Al and Miah, Abu Saleh Musa and Mansor, Sarina (2025) Advanced Hydro-Informatic Modeling through Feedforward Neural Network, Federated Learning, and Explainable AI for Enhancing Flood Prediction. IEEE Open Journal of the Computer Society. pp. 1-12. ISSN 2644-1268

[img] Text
4.pdf - Published Version
Restricted to Repository staff only

Download (6MB)

Abstract

Flood prediction is one of the most critical challenges facing today’s world. Predicting the probable time of a flood and the area that might get affected is the main goal of it, and more so for a region like Sylhet, Bangladesh where transboundary water flows and climate change have increased the risk of disasters. Accurate flood detection plays a vital role in mitigating these impacts by allowing timely early warnings and strategic planning. Recent advancements in flood prediction research include the development of robust, accurate, and low-cost flood models designed for urban deployment. By applying and utilizing powerful deep learning models show promise in improving the accuracy of prediction and prevention. But those models faced significant issues related to scalability, data privacy concerns and limitations of cross-border data sharing including the inaccuracies in prediction models due to changing climate patterns. To address this, our research adopts the Federated Learning (FL) framework in an effort to train state�of-the-art deep learning models like Long Short-Term Memory Recurrent Neural Network (LSTM-RNN), Feed-Forward Neural Network (FNN) and Temporal Fusion Transformer-Convolutional Neural Network (TFT -CNN) on a 78-year dataset of rainfall, river flow, and meteorological variables from Sylhet and its upstream regions in Meghalaya and Assam, India. This approach promotes data privacy and allows collaborative learning while working under cross-border data-sharing constraints, therefore improving the accuracy of prediction. The results showed that the best-performing FNN model achieved an R-squared value of 0.96, a Mean Absolute Error (MAE) value of 0.02, Percent bias (PBIAS) value of 0.4185 and lower Root Mean Square Error (RMSE) in the FL environment. Explainable AI techniques, such as SHAP, sheds light on the most significant role played by upstream rainfall and river dynamics, particularly from Cherrapunji and the Surma-Kushiyara river system, in driving flood events in Sylhet. These results demonstrate the effectiveness of privacy-preserving and AI-driven methodology implemented. These are being used in improving flood prediction and provide actionable insights for policymakers and disaster management authorities to pave the way toward scalable, transnational strategies that can be applied to mitigate the effects of flooding in vulnerable regions.

Item Type: Article
Uncontrolled Keywords: Federated Learning, Machine Learning, Deep Learning, LSTM-RNN, FNN, Flood Prediction, Data Analysis, Explainable AI, Surma River.
Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5101-6720 Telecommunication. Including telegraphy, telephone, radio, radar, television
Divisions: Faculty of Engineering (FOE)
Depositing User: Ms Suzilawati Abu Samah
Date Deposited: 29 Apr 2025 08:04
Last Modified: 29 Apr 2025 08:04
URII: http://shdl.mmu.edu.my/id/eprint/13687

Downloads

Downloads per month over past year

View ItemEdit (login required)