U-Net-Based Semantic Classification for Flood Extent Extraction Using SAR Images: A Case Study for 2019 Central US Flooding
Abstract
SAR (Synthetic Aperture Radar) is a critical source of earth observations that are unaffected by weather or time of day. Due to rapid development of data-driven models, particularly deep learning models, we can investigate and exploit SAR images in a more data-driven manner. Many studies have demonstrated the efficacy of using deep learning models to extract flood extent from SAR images. However, most of those studies compared advanced deep learning models to results generated by less advanced benchmark methodologies, such as traditional machine learning models and/or older versions of threshold-based approaches. Furthermore, few studies used other researchers' datasets to validate their findings. Those two reasons may result in insufficient comparison and validation of the results, leading to misleading conclusions. In this study, we first extracted water bodies from several SAR images captured during the 2019 Central US Flooding using a modified U-Net, and then compared our results to two other deep learning models, standard U-Net and ResNet50, an advanced thresholding method, Bmax Otsu, and a recently introduced flood inundation map archive. We also investigated how data input types, input resolution, and the use of pre-trained weights affect U-Net performance. We used a three-category classification frame (permanent water, flood, and dry area) to investigate if and how permanent water and flood pixels behave differently. The Google Earth Engine (GEE) cloud platform was used for data collection and pre-processing. The adjusted U-Net outperformed all other benchmark models and datasets, according to the results. Adding a slope layer to input data with a resolution of 30m improves model performance when compared to results trained on only VV and VH bands of SAR images. Adding Digital Elevation Model (DEM) and Height Above Nearest Drainage (HAND) layers, on the other hand, had limited effect on 30-m input performance but did make a difference for models trained on 10-m datasets. The findings also suggested that narrow river channels, where CNN-based semantic segmentation may fail to correctly classify pixels, should be given more attention. Finally, our findings revealed that distinguishing between permanent water and flood pixels is necessary because they behave differently in semantic segmentation.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFMNH45B0463L