DEEP LEARNING-BASED FLOOD INUNDATION PREDICTION IN THE PATTANI RIVER BASIN

Abstract

Accurate flood prediction is critical for effective disaster management and mitigation. This study employs deep learning models, including Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM) networks, and Convolutional Neural Networks (CNNs), to enhance flood forecasting for both water level and flood inundation predictions. By integrating upstream river flow, river water level, and tidal level data, the models aim to improve prediction accuracy. Water level forecasting involved evaluating GRU and LSTM models across four scenarios over lead times up to 24 hours, using Root Mean Squared Error (RMSE) and Nash-Sutcliffe Efficiency (NSE) as performance metrics. The results showed that GRU models consistently outperformed LSTM models when using all three parameters, while LSTM exhibited the worst performance, with higher RMSE and lower NSE values. For flood inundation prediction, CNNs were employed using Sentinel-1 GRD images as target data. Scenarios incorporating all three parameters achieved the highest average True Positive Rate (TPR) for both non-flooded and flooded areas, underscoring the value of integrating diverse data sources for accurate flood predictions. This research presents a sustainable, real-time flood prediction solution that reduces computational time while maintaining high accuracy. The findings support smarter water management strategies, aiding authorities in minimizing flood impacts on communities and infrastructure.