Integration of recurrent neural network models and flood inundation extent libraries for flood forecasting

High rainfall events have increased the frequency of floods worldwide, resulting in significant loss of life and property. Developing countries like India face severe flood situations across various states during the monsoon season. Timely and accurate flood forecasting can help disaster management authorities save lives through timely evacuation. The utilisation of deep learning models can aid in accurate flood water level prediction. Recurrent Neural Network like Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) models have potential to analyze sequential data. There is a need to compare available deep learning models, including GRU and LSTM models, to identify the most efficient model for river water level forecasting. This study applies GRU and LSTM models to satellite-derived rainfall, soil moisture, and temperature data, as well as ground-based river water level measurements in the upper river basin, to forecast river water levels in the lower region of the Bagmati river basin. The LSTM models, particularly with the Swish activation function, outperform GRU models in terms of accuracy (89 %), Mean Square Error (MSE, 0.0163), Mean Absolute Error (MAE, 0.0864), and R-squared (0.9630), demonstrating superior predictive capabilities. This work is further enhanced by integrating historical flood extent libraries, derived from remote sensing satellite data and water levels at gauge stations, to simulate probable flood inundation at specific water levels. Comparative analysis of different deep learning models and the integration of flood extent libraries significantly improves the reliability and accuracy of flood forecasting.