Global flood extent segmentation in optical satellite images

Area of Science:

• Earth and Environmental Sciences
• Remote Sensing
• Artificial Intelligence

Background:

• Floods are destructive natural disasters, with increasing intensity and frequency projected.
• Current satellite flood mapping methods struggle with cloud cover, accuracy, and geographic generalization.
• Accurate, timely flood maps are crucial for disaster response and damage assessment.

Purpose of the Study:

• To develop a machine learning framework for operational flood mapping using optical satellite imagery.
• To address limitations of current methods, including cloud coverage and generalization.
• To provide accurate and fast water segmentation masks for improved emergency response.

Main Methods:

• A clouds-aware segmentation model was trained on an extended WorldFloods dataset.
• The framework utilizes optical satellite data from Sentinel-2 and Landsat 8/9.
• Performance was evaluated using a novel global dataset of manually labeled flood maps.

Main Results:

• The model generates accurate water segmentation masks, even under semitransparent clouds.
• The approach demonstrates superior performance compared to traditional spectral index thresholding methods.
• Flood maps were successfully generated for recent major floods in Pakistan and Australia.

Conclusions:

• The proposed machine learning framework enhances operational flood mapping capabilities.
• It offers improved accuracy, speed, and geographic applicability for flood detection.
• The high-resolution flood extent data aids in assessing impacts on populations and infrastructure.