Rapid and large-scale mapping of flood inundation via integrating spaceborne synthetic aperture radar imagery with unsupervised deep learning
Xin Jiang,Shijing Liang,Xinyue He,Alan D. Ziegler,Peirong Lin,Ming Pan,Dashan Wang,Junyu Zou,Dalei Hao,Ganquan Mao,Yelu Zeng,Jie Yin,Lian Feng,Chiyuan Miao,Eric F. Wood,Zhenzhong Zeng
DOI: https://doi.org/10.1016/j.isprsjprs.2021.05.019
IF: 12.7
2021-08-01
ISPRS Journal of Photogrammetry and Remote Sensing
Abstract:<p>Synthetic aperture radar (SAR) has great potential for timely monitoring of flood information as it penetrates the clouds during flood events. Moreover, the proliferation of SAR satellites with high spatial and temporal resolution provides a tremendous opportunity to understand the flood risk and its quick response. However, traditional algorithms to extract flood inundation using SAR often require manual parameter tuning or data annotation, which presents a challenge for the rapid automated mapping of large and complex flooded scenarios. To address this issue, we proposed a segmentation algorithm for automatic flood mapping in near-real-time over vast areas and for all-weather conditions by integrating Sentinel-1 SAR imagery with an unsupervised machine learning approach named Felz-CNN. The algorithm consists of three phases: (i) super-pixel generation; (ii) convolutional neural network-based featurization; (iii) super-pixel aggregation. We evaluated the Felz-CNN algorithm by mapping flood inundation during the Yangtze River flood in 2020, covering a total study area of 1,140,300 km<sup>2</sup>. When validated on fine-resolution Planet satellite imagery, the algorithm accurately identified flood extent with producer and user accuracy of 93% and 94%, respectively. The results are indicative of the usefulness of our unsupervised approach for the application of flood mapping. Meanwhile, we overlapped the post-disaster inundation map with a 10-m resolution global land cover map (FROM-GLC10) to assess the damages to different land cover types. Of these types, cropland and residential settlements were most severely affected, with inundation areas of 9,430.36 km<sup>2</sup> and 1,397.50 km<sup>2</sup>, respectively, results that are in agreement with statistics from relevant agencies. Compared with traditional supervised classification algorithms that require time-consuming data annotation, our unsupervised algorithm can be deployed directly to high-performance computing platforms such as Google Earth Engine and PIE-Engine to generate a large-spatial map of flood-affected areas within minutes, without time-consuming data downloading and processing. Importantly, this efficiency enables the fast and effective monitoring of flood conditions to aid in disaster governance and mitigation globally.</p>
imaging science & photographic technology,remote sensing,geography, physical,geosciences, multidisciplinary