Flood Detection Using Variational Autoencoder in SAR Images

Monitoring large areas to detect unexpected changes is beneficial in disaster response preparations, particularly when using a persistent information source such as Synthetic Aperture Radar (SAR) images. However, the complexity of radar images makes it difficult to select features that differentiate relevant and irrelevant changes. In this research, a Variational Autoencoder (VAE) was trained in a self-supervised way to learn meaningful representations of the multitemporal SAR dataset. The embeddings from adjacent timeframes were then compared using a distance function. Direct pixel comparisons using the Cosine Similarity and embedding comparisons with the Wasserstein divergence were both able to highlight the flood event in the final SAR image pair; however, the latter managed to ignore changes in the previous pairs better than the former.