Document Type : Original Article
Authors
1
GIS Department, School of Surveying and Geospatial Eng., College of Engineering, University of Tehran, Tehran, Iran
2
Centre of Excellence in Geomatic Eng. in Disaster Management and Land Administration in Smart City Lab., School of Surveying and Geospatial Eng., College of Engineering, University of Tehran, Tehran, Iran
10.22059/eoge.2026.409803.1208
Abstract
Floods are among the most destructive natural hazards worldwide, demanding accurate and timely flood extent mapping for effective disaster management. With the increasing availability of satellite observations, deep learning methods have become vital approaches to improve flood detection. This study systematically compares convolutional neural network (CNN) architectures (U-Net, DeepLabV3 and HRNet) with a conditional generative adversarial network (GAN) (Pix2Pix) for flood extent mapping using the Sen1Floods11 dataset, which provides paired Sentinel-1 synthetic aperture radar (SAR) and Sentinel-2 optical imagery with annotated flood masks. Models were trained separately on SAR and optical data to assess modality-specific performance and evaluated using Overall accuracy, Intersection over Union (IoU), Precision, Recall and F1-score show that SAR-based models consistently outperform optical-based counterparts, highlighting the robustness of radar data under adverse weather conditions. Among CNNs, DeepLabV3 achieved the highest IoU (84.56%) and F1-score (92.35%), offering precise delineation of flooded areas. In contrast, Pix2Pix achieved the highest overall accuracy (97.45%) and Recall (93.79%), capturing broader flood extents but with a tendency to overestimate flood extent. These findings indicate that CNNs are well-suited for high-precision applications such as risk assessment, while GAN-based models may be advantageous for rapid emergency response where minimizing missed detections is critical. The study underscores the complementary strengths of CNNs and GANs and points toward multimodal fusion and real-time deployment as promising future directions for operational flood monitoring.
Keywords
Main Subjects
)