U-net-based semantic classification for flood extent extraction using SAR imagery and GEE platform: A case study for...

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Artificial Intelligence

Background:

• Water body extraction is crucial for environmental monitoring and disaster management.
• Advancements in satellite data and computational power have driven progress in data-driven extraction models.
• Sentinel-1 satellite data offers valuable high-resolution information for hydrological studies.

Purpose of the Study:

• To develop and evaluate an improved U-Net model for water body extraction using Sentinel-1 data.
• To compare the modified U-Net model against benchmark methods for flood mapping accuracy.
• To investigate the impact of different data inputs, resolutions, and pre-trained weights on model performance.

Main Methods:

• Modified U-Net convolutional neural network (CNN) architecture for semantic segmentation.
• Utilized Sentinel-1 Synthetic Aperture Radar (SAR) data (VV and VH bands) from the 2019 Central US flooding.
• Incorporated auxiliary data layers including elevation (DEM), slope, and Height Above Nearest Drainage (HAND).
• Trained and tested models using Google Earth Engine (GEE) cloud platform.

Main Results:

• The adjusted U-Net model significantly outperformed standard U-Net, ResNet50, and Bmax Otsu methods.
• Adding a slope layer improved 30m resolution data performance; DEM and HAND improved 10m resolution data performance.
• Distinguishing between permanent water and flood pixels enhanced classification accuracy.
• Pre-trained weights from coarser datasets reduced initial training loss and accelerated convergence.

Conclusions:

• The modified U-Net model provides a robust and accurate method for water body and flood extraction from Sentinel-1 data.
• Input data type, resolution, and the use of pre-trained weights are critical factors influencing model performance.
• Differentiating permanent water from floodwater is essential for precise hydrological analysis and flood mapping.