[A-CG36] Satellite Earth Environment Observation

[ACG36-P23] A Self-Supervised Learning framework for Rapid Estimation of Flood Inundation Area Using Multi-Source Remote Sensing Data

*Tomohiro Tanaka¹, Narumasa Tsutsumida¹ (1.Saitama University)

Keywords:Disaster, Deep Learning, Sentinel

Flooding is one of the most prevalent natural hazards, leading to significant human and economic losses. These risks have increased recently due to climate change and rapid urbanization, highlighting the need for effective and immediate disaster management strategies. Rapid and accurate assessments during disasters are essential for conducting successful rescue operations and minimizing damage. Earth observation is useful for such assessments. However, satellite monitoring has been limited, as it typically relies on a single data source from either Synthetic Aperture Radar (SAR) or Multispectral (MS) sensors. This limitation in frequency and land surface monitoring ability makes rapid flood monitoring difficult. Therefore, our study aims to address this issue by developing a new method to estimate flood inundation areas. This method combines multi-source remote sensing imagery with deep learning (DL). We have designed a DL architecture that integrates SAR and MS data in a Self-Supervised Learning (SSL) framework. This allows us to estimate flood inundation areas using any data immediately available during or after a disaster. To demonstrate its feasibility, we trained and tested our model using flood event data from Spain in 2019, derived from Sentinel-1 and Sentinel-2 images. We were able to successfully estimate inundation areas using only Sentinel-1 (IoU: 0.60) and only Sentinel-2 (IoU: 0.81). Our approach is not limited to using Sentinel imagery. We plan to extend its use to higher-frequency observation constellation satellite data. We expect this to allow even faster and more accurate inundation estimations, further aiding disaster assessment efforts.