An Artificial Neural Network-Based Geo-Spatial Model for Real-Time Flood Risk Prediction Using Multi-Source High-Resolution Data

Abstract

Flood prediction presents a pressing challenge in disaster management, especially in regions vulnerable to extreme weather events. In response, this study offers a novel approach to flood risk prediction by developing a deep learning-based Geo-Spatial Artificial Neural Network (ANN). The model actively integrates high-resolution satellite imagery, meteorological data, and topographic indicators, such as rainfall, elevation, and land use to capture complex spatial and environmental relationships that influence flood risk. This study conducted data preprocessing using Principal Component Analysis (PCA) and normalization to ensure consistency across datasets. It built the ANN with multiple hidden layers and trained it using the backpropagation algorithm on historical flood data. Furthermore, it designed the ANN model with multiple hidden layers and trained it using the backpropagation algorithm. The model achieved a notable 92% prediction accuracy, significantly outperforming traditional flood prediction methods, which typically yield 75–85% accuracy. Conventional metrics were Mean Squared Error (1.41) and R-squared (0.94). It confirmed the model’s superior ability to predict high-risk flood zones. The model also effectively captured non-linear patterns that conventional statistical or deterministic methods often failed to detect. The results showed that the model generalizes well and adapts effectively, making it suitable for real-time and data-driven flood forecasting. By integrating artificial intelligence with geo-spatial analytics, this study offers a scalable, accurate, and efficient tool for early warning systems and risk management. It recommends that future research should focus on incorporating additional data sources and refining model training techniques to further enhance scalability and performance.