Urban Flood Hazard Mapping with Deep Learning Models in Astana, Kazakhstan

Abstract

Urban flooding has always posed significant risks to the lives, infrastructure, and economy of rapidly expanding and developing cities, which was further highlighted by the recent devastating floods in northern Kazakhstan. Two-dimensional (2D) hydrodynamic models are widely recognized for their accuracy in flood hazard and risk mapping, but these models are resource-intensive and require significant computational power, limiting their use in real-time hazard prediction. In response, machine learning, particularly convolutional neural networks (CNNs), have gained massive traction as promising alternatives for near real-time urban flood hazard mapping (UFHM). Despite this progress, most existing model implementations rely on outdated CNN architectures such as U-Net and ResNet. This study integrates the state-of-the-art ConvNeXt V2 backbone within a U-net encoder-decoder configuration (ConvNeXt-Unet) to predict flood extent and depth using 12 static and 1 dynamic predictors. Flood simulations based on the Copernicus digital elevation model (DEM) and Sentinel-2 satellite imagery are performed with HEC-RAS hydrodynamic model to create a training dataset for an urban catchment in Astana, Kazakhstan. The proposed ConvNeXt-Unet model demonstrates strong predictive performance, achieving Nash–Sutcliffe efficiency of 0.993 for regression and accuracy of 99.7% for classification. Among the predictors, the terrain ruggedness index was the most influential variable, followed by the topographic position index and the drainage direction. The results suggest that our flood hazard mapping model performs similarly to traditional hydrodynamic methods with minimal processing power and time, making it suitable for rapid flood modeling applications such as forecast and emergency response.