Self-Supervised Contrastive Model for Flood Mapping and Monitoring on SAR Time-Series

Flooding is a natural disaster that has been increasing in recent years due to climate and land-use changes. Earth observations, such as Synthetic Aperture Radar (SAR) data, are valuable for assessing and mitigating the negative impacts of flooding. Cloud cover is highly correlated with flooding events, making SAR a preferable choice over optical data for flood mapping and monitoring.

Traditional methods for flood mapping and monitoring using SAR data, such as otsu and CVA, can be affected by noise, false detections due to shadows and occlusions, and geometric distortions. While automatic thresholding can be effective with these methods, manual adjustment of the threshold is often required to produce an accurate change map. 

Supervised deep learning methods using large amounts of labeled data could potentially improve the accuracy of flood mapping and monitoring. We have a large amount of earth observation data, but the availability of labeled data is limited and labeling data is time-consuming and requires domain expertise. On the other hand, Supervised model training on small datasets causes severe generalizability issues when inference is taken on a new site. 

To address these challenges, we propose a novel self-supervised method for mapping and monitoring floods on Sentinel-1 SAR time-series data. We propose a probabilistic model trained on unlabeled data using self-supervised techniques, such as reconstruction and contrastive learning. The model is trained to learn the spatiotemporal features of the area. It monitors the changes by comparing the latent feature distribution at each time stamp and generates change maps to reflect the changes in the area. 

We also propose a framework for flood monitoring that continuously monitor the area using time series data. This framework automatically detects the change point i.e. when the major change started reflecting on available SAR data. Our continuous monitoring framework combined with a better temporal resolution (better than Sentinel-1) can potentially detect flood events in an early stage, allowing for more time for evacuation planning. 

The model is evaluated on nine recent flood events from ‘Mekong’, ‘Somalia’, ‘Scotland’, ‘Australia’, ‘Bosnia’, ‘Germany’, ‘Spain’, ‘Bolivia’, and ‘Slovakia’ sites. We compared our results with traditional methods, and existing supervised and unsupervised methods. Our detailed evaluation indicates that our model is more accurate and generalizable to new sites. The model achieves an average Intersection Over Union (IoU) value of 70% and an F1 score of 81.14%, which are both higher than the scores of the previous best-performing method. Overall, our proposed model’s improvement range from 7-26% in terms of F1 and 8-31% in terms of IoU score.