U-Net-based Semantic Classification for Flood Extent Extraction using SAR Imagery and GEE Platform: A Case Study for 2019 Central US Flooding

Abstract

Data-driven models for water body extraction have experienced accelerated growth in recent years, thanks to advances in processing techniques and computational resources, as well as improved data availability. In this study, we modified the standard U-Net, a convolutional neural network (CNN) method, to extract water bodies from scenes captured from Sentinel-1 satellites of selected areas during the 2019 Central US flooding. We compared the results to several benchmark models, including the standard U-Net and ResNet50, an advanced thresholding method, Bmax Otsu, and a recently introduced flood inundation map archive. Then, we looked at how data input types, input resolution, and using pre-trained weights affect the model performance. We adopted a three-category classification frame to test whether and how permanent water and flood pixels behave differently. Most of the data in this study were gathered and pre-processed utilizing the open access Google Earth Engine (GEE) cloud platform. According to the results, the adjusted U-Net outperformed all other benchmark models and datasets. Adding a slope layer enhances model performance with the 30m input data compared to training the model on only VV and VH bands of SAR images. Adding DEM and Height Above Nearest Drainage (HAND) model data layer improved performance for models trained on 10-m datasets. The results also suggested that CNN-based semantic segmentation may fail to correctly classify pixels around narrow river channels. Furthermore, our findings revealed that it is necessary to differentiate permanent water and flood pixels because they behave differently. Finally, the results indicated that using pre-trained weights from a coarse dataset can significantly minimize initial training loss on finer datasets and speed up convergence.