Enhanced machine learning models development for flash flood mapping using geospatial data

Flash floods are dangerous and unpredictable. This study aimed to improve flash flood prediction in Algeria's Hodna Basin using advanced AI models and GIS (GeoAI). Each watershed exhibits unique characteristics that contribute to flooding, primarily driven by hydrological and topographic factors. To capture and incorporate these distinctive attributes, a wide range of data sources were integrated, including topographic features, hydrological parameters, and remote sensing data. These data encompassed slope, rainfall, aspect, elevation, land use/land cover (LULC), topographic wetness index, distance from rivers, stream power index, curvature, hill shade, and geology. These diverse factors served as input variables for the present models. The data sources employed were Landsat 8, Sentinel-2 imagery, climate hazards group infrared precipitation with station data (CHIRPS) data and USGS data, which were integrated within into a Geographic Information System (GIS) framework. The research was applied a stacking clustering technique, combining three models: categorical boosting-convolutional neural networks (Catboost-CNN), categorical boosting-deep belief networks (CatBoost-DBNs), and categorical boosting-long short-term memories (CatBoost-LSTMs). To assess the performance of the models, the dataset underwent random partitioning into two subsets: 70% for training and calibration, and 30% for testing. Various statistical metrics, including sensitivity, specificity, accuracy, F1 score, precision and recall, and the area under the receiver operating characteristic curve (AUC-ROC), were employed to evaluate model effectiveness. The study's findings showcase the stacked CatBoost-CNNs algorithm achieving exceptional prediction accuracy at 92%. Furthermore, CatBoost-DBNs demonstrated a commendable accuracy of 81%, while CatBoost-LSTMs achieved an accuracy of 89%. Leveraging the capabilities of GIS, a flash flood susceptibility map was generated. These results compellingly indicated that the stacking methodology substantially improves the accuracy of flash flood forecasting, leading to practical outcomes. The findings of the study offer valuable insights to inform future research and decision-making.