Abstract
In this study, the effectiveness of several convolutional neural network architectures (AlbuNet-34/FCN/DeepLabV3+/U-Net/U-Net++) for water and flood mapping using Sentinel-1 amplitude data is compared to an operational rule-based processor (S-1FS). This comparison is made using a globally distributed dataset of Sentinel-1 scenes and the corresponding ground truth water masks derived from Sentinel-2 data to evaluate the performance of the classifiers on a global scale in various environmental conditions. The impact of using single versus dual-polarized input data on the segmentation capabilities of AlbuNet-34 is evaluated. The weighted cross entropy loss is combined with the Lovász loss and various data augmentation methods are investigated. Furthermore, the concept of atrous spatial pyramid pooling used in DeepLabV3+ and the multiscale feature fusion inherent in U-Net++ are assessed. Finally, the generalization capacity of AlbuNet-34 is tested in a realistic flood mapping scenario by using additional data from two flood events and the Sen1Floods11 dataset. The model trained using dual polarized data outperforms the S-1FS significantly and increases the intersection over union (IoU) score by 5%. Using a weighted combination of the cross entropy and the Lovász loss increases the IoU score by another 2%. Geometric data augmentation degrades the performance while radiometric data augmentation leads to better testing results. FCN/DeepLabV3+/U-Net/U-Net++ perform not significantly different to AlbuNet-34. Models trained on data showing no distinct inundation perform very well in mapping the water extent during two flood events, reaching IoU scores of 0.96 and 0.94, respectively, and perform comparatively well on the Sen1Floods11 dataset.
Highlights
T HE demand for reliable and robust crisis information after catastrophic disasters has substantially grown in the past decades [1]
We compared the performance in water and flood mapping of a state-of-the-art rule-based Sentinel-1 flood processor (S-1FS) with five convolutional neural networks (CNNs) architectures and assessed the impact of various hyperparameters on the performance of the trained CNN models. (i) We confirmed the observation by Liu et al [31] and Katiyar et al [23] that VH or VV-VH polarized data is the preferred input feature for the purpose of water mapping using CNNs and Sentinel-1 data. (ii) We further showed that a linearly weighted combination of the weighted cross entropy loss function and the Lovász loss function yields better testing results than either of the loss functions on their own
Considering other studies which reported similar results using either different loss functions [24] or data from a different domain [52], our results strengthen the assumption that combining distribution-based and region-based loss functions is beneficial for many segmentation tasks. (iii) Our results further indicate that geometric data augmentation methods should be treated with care when working with Synthetic Aperture Radar (SAR) data for water mapping, whereas radiometric data augmentation in the form of intensity augmentation or speckle noise simulation leads to better testing results. (iv) All examined CNN architectures outperform the rule-based S-1FS in the task of water mapping, problems can arise in arid or mountainous environments using either method
Summary
T HE demand for reliable and robust crisis information after catastrophic disasters has substantially grown in the past decades [1]. Flood events make up one third of all recorded natural disasters in the past century [2] and were related to approximately 52 % of all activations of the International Charter “Space and Major Disasters” between the years 1999 – 2013 [1] They are usually not localized but affect large regions simultaneously and the atmospheric conditions often prevent observation using optical or multi-spectral sensor systems. SAR is a side-looking imaging radar system that utilizes microwave radiation to create images of the surface of the earth It is an active instrument, that requires no illumination from the sun and can penetrate cloud cover, it is often used for rapid mapping purposes during flood events [1]
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
