A Novel LSTM Approach for Reliable and Real-Time Flood Prediction in Complex Watersheds

In the context of global climate change, the world is increasingly experiencing abnormal phenomena, with natural disasters being among the most critical challenges. Adapting to these changes and mitigating their risks has become imperative. Floods, as one of the most devastating natural threats, are a crucial subject of study, particularly in understanding and predicting their dynamic behavior. This research highlights the importance of flood mapping and assessment using satellite imagery and advanced technologies such as Geographical Information System (GIS) and Deep Learning (DL). The study focuses on Tetouan city, located in northern Morocco, which provides ideal conditions for this research. Eleven flood conditioning factors were analyzed, including elevation, slope, aspect, Stream Power Index (SPI), Topographic Position Index (TPI), Topographic Wetness Index (TWI), curvature, drainage density (DD), distance to rivers (DR), Normalized Difference Vegetation Index (NDVI), and land use (LU). To identify the most relevant factors influencing flood occurrence, Information Gain Ratio (IGR) and Frequency Ratio (FR) methods were applied, allowing for the exclusion of non-impactful factors. The Long Short-Term Memory (LSTM) deep learning technique was utilized on a balanced dataset of 1946 samples generated through data augmentation. Additional optimization techniques were implemented to enhance the model’s performance. The findings demonstrate a high prediction accuracy of 96.06%, underscoring the model's effectiveness in flood risk assessment.