A Comparison of three deep learning-based methods for large-scale urban flood mapping using SAR data

<p>Synthetic Aperture Radar (SAR)-based floodwater detection in urban areas remains challenging because of the complex urban environment. Generally, open water appears as dark in SAR intensity images due to low values of backscattering caused by specular reflections, while standing water in built-up areas may lead to an increase of the backscattering value depending on the strength of the double bounce effect between the floodwater and the building&#8217;s facades. According to previous studies, it is known that the multitemporal interferometric SAR coherence is valuable for improving flood detection in urbanized areas while SAR intensity is more suited to map floods over bare soil. Deep convolutional neural networks approaches have also shown promising results in remote sensing applications, such as land cover classification, object detection and floodwater mapping. For the latter case and with particular attention to urban areas, there is not yet a well-established and unique method neither a privileged dataset to perform the detection of floodwater. In order to have a better understanding of the ability of different deep learning models for urban flood mapping, we compared the performance of three different deep learning-based methods, i.e. U-Net, U-Net with convolutional block attention module (CBAM) and U-Net with an Urban-aware module developed by us, for large-scale urban flood mapping. Here, we used as input multi-temporal intensity and interferometric SAR coherence data and the classification differentiates between flooded bare soils/sparely vegetated areas and flooded urban areas. To learn how to focus on different inputs, the urban-aware U-Net considers prior information provided by a SAR-derived probabilistic urban mask while CBAM U-Net only uses annotated data.<br>The annotated training dataset is composed of a small subset of Sentinel-1 data acquired during the Houston (US) flood, caused by Hurricane Harvey in 2017, and the Iwaki (Japan) flood, caused by Typhoon Hagibis in 2016, where ground truth data are available. Three independent datasets (i.e. Houston (US) flood in 2017, Koriyama (Japan) flood in 2016 and Beira (Mozambique) flood in 2019) were considered as test cases in order to evaluate the generalizability capabilities of the proposed approach with respect to different scenarios. To evaluate the accuracy of flood mapping in urban areas, we adopted the F1 score. The urban-aware U-Net improves the F1 score to 0.63 in the Houston case and 0.66 in the Beira case while the other two models&#8217; results have quite low F1 values (0.04 ~ 0.38) in Houston case and Beira case. Moreover, a visual inspection of the resulting floodwater maps over the entire Sentinel-1 frame suggests that urban-aware U-Net has less over-detection compared with U-Net and CBAM U-Net. These results indicate that the prior information helps in the proper use of multi-temporal SAR data in large-scale flood mapping. Moreover, considering that the models were trained using a very small and independent dataset and given the agreement of the results with the available ground truth, we consider urban-aware U-Net as a promising approach, having the potential to be used for near real-time urban flood mapping in case of future flood events.</p>