Data Augmentation for Building Footprint Segmentation in SAR Images: An Empirical Study

Abstract

Building footprints provide essential information for mapping, disaster management, and other large-scale studies. Synthetic Aperture Radar (SAR) provides consistent data availability over optical images owing to its unique properties, which consequently makes it more challenging to interpret. Previous studies have demonstrated the success of automated methods using Convolutional Neural Networks to detect buildings in Very High Resolution (VHR) SAR images. However, the scarcity of such datasets that are available to the public can limit research progress in this field. We explored the impact of several data augmentation (DA) methods on the performance of building detection on a limited dataset of SAR images. Our results show that geometric transformations are more effective than pixel transformations. The former improves the detection of objects with different scale and rotation variations. The latter creates textural changes that help differentiate edges better, but amplifies non-object patterns, leading to increased false positive predictions. We experimented with applying DA at different stages and concluded that applying similar DA methods in training and inference showed the best performance compared with DA applied only during training. Some DA can alter key features of a building’s representation in radar images. Among them are vertical flips and quarter circle rotations, which yielded the worst performance. DA methods should be used in moderation to prevent unwanted transformations outside the possible object variations. Error analysis, either through statistical methods or manual inspection, is recommended to understand the bias presented in the dataset, which is useful in selecting suitable DAs. The findings from this study can provide potential guidelines for future research in selecting DA methods for segmentation tasks in radar imagery.