Integrating artificial neural networks with multi-criteria decision analysis for flood risk assessment and shelter suitability mapping

Floods are among the most destructive natural hazards in Sylhet, Bangladesh, determining the assessment of vulnerability and identification of mitigation strategies crucial for sustainable development. This study aims to identify flood-prone areas and suitable flood shelter zones using a GIS-based artificial neural network (GANN) model and multi-criteria decision-making (MCDM) approach. A total of 500 samples (70% for training, 30% for testing) were collected based on flooding and non-flooding characteristics. Total 13 factors, including topographic roughness index (TRI), slope, elevation, curvature, aspect, precipitation, distance from rivers (DFRi) and roads (DFRo), drainage density, topographic wetness index (TWI), normalized difference built-up index (NDBI), normalized difference vegetation index (NDVI), and land use/land cover (LULC) were considered for flood susceptibility mapping. The GANN model classified 29.9% (994.26 km2) and 27.03% (898.9 km2) of the area as very high to highly susceptible to flooding, with a 99.4% success rate and 85.5% prediction accuracy (AUC-ROC). LULC-based analysis showed that 27.92%, 12.62%, 31.66%, and 26.72% of vegetation, barren land, built-up, and dense vegetation areas, respectively, are highly susceptible. For shelter suitability, elevation, DFRo, population density, proximity to settlements and final flood map were considered, with MCDM results indicating that 0.02% (0.71 km2) and 9.09% (302.37 km2) are very highly and highly suitable for shelters. Additionally, 5.52% (42.80 km2) of barren land was prioritized for shelter establishment. These findings will help decision-makers and local communities with data-driven insights to improve flood resilience and ensure safe shelter planning in flood-prone areas.