Water Mapping Using Synthetic Aperture Radar Data and Convolutional Neural Networks

Kurzfassung

Water mapping to derive flood parameters using Synthetic Aperture Radar data is an estab-lished procedure in emergency situations. In this study, the effectiveness of convolutionalneural networks (AlbuNet-34) for the purpose of water mapping using Sentinel-1 data is in-vestigated and compared to the performance of a state-of-the-art rule-based processor forwater mapping. This comparison is made using a reference dataset containing 67 globallydistributed Sentinel-1 scenes and the corresponding ground truth water masks derived fromSentinel-2 data to evaluate the performance of the classifiers on a global scale in variousenvironmental conditions. Various semi-random undersampling strategies for balancing thedataset are explored and the effect of the sample size on the performance of the models is in-vestigated. The cross entropy loss is compared to the region-based Lovász loss function andvarious data augmentation methods (flip, zoom, intensity variation, rotation, speckle simula-tion) are assessed. Furthermore, the impact of using single polarized VV or VH data and dualpolarized VV-VH data on the segmentation capabilities of AlbuNet-34 is evaluated. Finally,the concept of atrous spatial pyramid pooling used in a DeepLabV3+ model with a ResNet-50 encoder is assessed with respect to segmentation performance. The IoU scores on theglobal test set of 14 Sentinel-1 scenes vary by 0.11, depending on the sampling strategy, andthe Lovász loss increases the test IoU score by 0.01 compared to the cross entropy loss.Left-right flip and intensity augmentation improve the performance of the model, zooming androtation show only minor impact and speckle simulation decreases the performance. Themodel trained using VV-VH polarized data outperforms the rule-based flood processor andincreases accuracy by 0.01, recall by 0.03, precision by 0.04, F1 by 0.06, Kappa by 0.06 and IoU by 0.06. DeepLabV3+ yields results comparable to AlbuNet-34.