# H2O-Net: Self-Supervised Flood Segmentation via Adversarial Domain Adaptation and Label Refinement

**Authors:** Peri Akiva, Matthew Purri, Kristin Dana, Beth Tellman, Tyler Anderson  
**arXiv:** [2010.05309](https://arxiv.org/abs/2010.05309)  
**Date:** 2020-10-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2010.05309v1/

## Key points

- Accurate flood detection in near real time is crucial for preventing loss of lives during natural disasters.
- Current flood alerting systems relying on low resolution, low latency satellites with long re-visit periods are unreliable.
- H2O-Net is a self-supervised deep learning method for flood segmentation using high resolution satellite imagery.
- H2O-Net bridges the domain gap between low and high latency satellite data through label refinement and domain adaptation techniques.
- H2O-Net can generate high quality ground truth data without requiring hand annotation through self-supervision mechanism during training.
- H2O-Net outperforms state-of-the-art semantic segmentation methods on satellite imagery, achieving a 10% improvement in pixel accuracy and a 12% improvement in mIoU for flood segmentation.
- The model's weights can be transferred to drone imagery, showcasing its generalizability beyond satellite-based flood detection.
- H2O-Net presents an innovative approach to accurate flood detection using deep learning and self-supervision techniques.

## Summary

Accurate flood detection in near real time is crucial for preventing loss of lives during natural disasters. Current flood alerting systems that rely on low resolution, low latency satellites with long re-visit periods are unreliable. To address this issue, the authors propose H2O-Net, a self-supervised deep learning method for flood segmentation using high resolution satellite imagery. H2O-Net bridges the domain gap between low and high latency satellite data by leveraging coarse-to-fine label refinement and adversarial domain adaptation techniques. The model learns to synthesize signals highly correlated with water presence as a domain adaptation step for semantic segmentation in high resolution satellite imagery. One key advantage of H2O-Net is its ability to generate high quality ground truth data without requiring any hand annotation. This is achieved through a self-supervision mechanism during training. The authors demonstrate the effectiveness of H2O-Net by outperforming state-of-the-art semantic segmentation methods on satellite imagery, achieving a 10% improvement in pixel accuracy and a 12% improvement in mIoU (mean Intersection over Union) for flood segmentation. Furthermore, the authors highlight the generalizability of their model by transferring the trained weights from satellite imagery to drone imagery which represents a highly different sensor and domain; showcasing the versatility and potential applicability of H2O-Net beyond just satellite based flood detection. In conclusion, H2O-Net presents an innovative approach to accurate flood detection using deep learning and self supervision techniques. By addressing the limitations of existing flood alerting systems this work has significant implications for improving disaster response efforts and ultimately saving lives.