Improving Flood Prediction Using Artificial Neural Networks With Optimal Feature Selection on a Benchmark Dataset

Disasters significantly impact people's lives; among them, flooding is the worst common, and it causes sudden and secure damage to both lives and property. Addressing such real‐time crisis demands intricate and sophisticated flood prediction models with enhanced capabilities. The development of efficient flood prediction models is often hindered by the lack of available datasets and the need for optimal feature. To address the challenge of data availability, in the proposed research, we have manually prepared a novel dataset by collecting data from NASA's (National Aeronautics and Space Administration) Power Project. The proposed dataset is experimentally evaluated and verified and has been organized into a balanced benchmark dataset with 33 features using the SMOTE algorithm. To enhance the provenance of flood prediction model, we propose a novel feature selection method. This method integrates outcomes from three different feature selection techniques to identify the most prominent features. The proposed feature selection method improves the model's performance and efficiency by identifying optimal predictors. Experimental results demonstrate that the artificial neural network trained with the selected relevant features accurately predicts flood occurrences, showing enhanced accuracy compared to state‐of‐the‐art methods.