Performance Evaluation of 3-D Convolutional Neural Network for Multitemporal Flood Classification Framework With Synthetic Aperture Radar Image Data

Abstract

Urban flooding significantly impacts populations and often coincides with heavy rainfall, making optical satellite observation challenging due to cloud cover. This study proposes a novel approach using synthetic aperture radar (SAR) sensors, which can penetrate clouds, to classify flooded urban areas. The framework employs a 3-D convolutional neural network (3-D CNN) to process multitemporal SAR data from Sentinel-1 (S-1). The dataset included 24 S-1 scenes with Dual VV and VH polarization from March 2019 to February 2020, divided into two co-event images, 18 preevent images, and four postevent images. The 3-D CNN achieved an average overall accuracy of 70.3% and a peak accuracy of 71.8%. These results demonstrate the 3-D CNN's potential to accurately estimate flood extent and identify flood-prone areas, supporting early detection and flood prevention in other cities.

Keywords

• 3-D convolutional neural network (3-D CNN)
• machine learning (ML)
• multitemporal data
• synthetic aperture radar (SAR)
• temporal variation
• urban flood