Accurate Flood Inundation Mapping (FIM) is of paramount importance for effective disaster risk reduction in the context of climate action. This paper presents an innovative approach to FIM using optimized deep learning models. Leveraging a vast historical dataset of flood events obtained from Radiant MLHub, which includes multispectral remote sensing data consisting of 12 spectral bands categorized as “Flood” and “No Flood,” a deep learning model is trained to predict flood extents with high accuracy. Bayesian Optimization (BO) and the Hyperband algorithm are employed during the hyperparameter tuning process of the deep learning models. Two base models are used: a Convolutional Neural Network (CNN) and the pre-trained Visual Geometry Group 16 (VGG16) model. To achieve the best performance, four scenarios-CNN-BO, CNN-Hyperband, VGG16-BO, and VGG16-Hyperband-are investigated. The results show that VGG16-Hyperband outperforms the other scenarios, achieving 87.3% validation accuracy.
Development of Flood Inundation Mapping by Using Optimized Deep Learning Model
2024-11-08
327298 byte
Conference paper
Electronic Resource
English
Improving flood inundation models using remotely sensed data
British Library Conference Proceedings | 2009
|Mapping Maryland's Inundation Zones
British Library Conference Proceedings | 2022
|Multi-Block Computation for Flood Inundation Studies
HENRY – Federal Waterways Engineering and Research Institute (BAW) | 2008
|Dynamic Wave Model for Flood Inundation Simulation in a River
British Library Conference Proceedings | 1999
|