A novel spatially enabled neural model for flood susceptibility in Northern Morocco

Flooding remains a major hazard in northern Morocco, where rapid urban growth and limited monitoring systems heighten the need for reliable flood susceptibility assessment. This study addresses this challenge by developing a Long Short-Term Memory (LSTM) Deep Learning (DL) model capable of predicting flood‐prone areas using satellite imagery and Geographic Information System (GIS) data. Eleven flood conditioning factors were incorporated, 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). Unlike previous studies relying on static GIS factors or traditional Machine Learning (ML) methods, this work evaluates how the influence of 11 conditioning factors varies across regions and tests the cross-regional transferability of the LSTM model. A balanced dataset of 1,946 samples was generated through data augmentation, and optimization techniques were implemented to enhance model performance. The proposed model achieved an accuracy of 96.06%, a precision of 94.56%, a recall of 97.54%, and an area under the Receiver Operating Characteristic (ROC) curve (AUC) of 96.10%, confirming its strong predictive capability. These findings show that LSTM can accurately capture spatial environmental patterns linked to flood occurrence, offering valuable support for urban planning, early warning systems, and improved flood risk management in data-scarce regions.