DeepFlood for Inundated Vegetation High-Resolution Dataset for Accurate Flood Mapping and Segmentation

Area of Science:

• Earth and Environmental Sciences
• Computer Science
• Remote Sensing

Background:

• Accurate flood extent mapping is crucial for disaster management but faces scalability and transferability issues with traditional methods.
• Deep learning, especially Convolutional Neural Networks (CNNs), offers a promising approach for automated flood mapping by learning spatial patterns.
• Existing datasets often lack the comprehensive annotations and diverse coverage needed for robust deep learning models.

Purpose of the Study:

• Introduce DeepFlood, a novel, high-resolution dataset for training deep learning models for flood mapping.
• Enable multi-modal flood mapping approaches using diverse imagery sources.
• Address the limitations of current datasets by providing detailed labels, including inundated vegetation.

Main Methods:

• Developed DeepFlood dataset with high-resolution manned/unmanned aerial and Synthetic Aperture Radar (SAR) imagery.
• Annotated imagery with detailed labels, focusing on challenging features like inundated vegetation.
• Evaluated various semantic segmentation architectures on the DeepFlood dataset.

Main Results:

• Demonstrated the usability and efficacy of the DeepFlood dataset in post-disaster flood mapping scenarios.
• Showcased the potential of deep learning models trained on DeepFlood for accurate flood extent assessment.
• Highlighted the dataset's capability to support multi-modal approaches.

Conclusions:

• DeepFlood provides a valuable resource for advancing deep learning-based flood mapping.
• The dataset facilitates improved accuracy and efficiency in disaster response and mitigation planning.
• Future research can leverage DeepFlood for developing more sophisticated flood prediction and management tools.