FloodIMG: Flood image DataBase system.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Convolutional Neural Networks (CNNs) excel at image processing by extracting hierarchical features.
• High-resolution image datasets are crucial for training effective CNN models for image segmentation.
• Existing datasets may lack sufficient diversity or specific focus for specialized tasks like flood analysis.

Purpose of the Study:

• To introduce the Flood Image (FloodIMG) database system, a novel dataset for flood-related image processing and segmentation.
• To facilitate the training and optimization of CNN models for flood detection and severity assessment.
• To provide a publicly accessible resource for researchers and developers in the field of disaster management and image analysis.

Main Methods:

• Developed Internet of Things Application Programming Interfaces (IoT APIs) to collect flood images from diverse sources like Twitter and US federal agencies (USGS, DOT).
• Collected, preprocessed, and formatted over 9200 flood-related images suitable for CNN training.
• Annotated images with bounding boxes and polygon primitives for object localization and classification, and utilized Fast Region-based CNN (R-CNN) for use-case analysis.

Main Results:

• The FloodIMG database comprises >9200 preprocessed images, with 7400 designated for training and >1800 for testing CNN models.
• Demonstrated the utility of FloodIMG using Fast R-CNN to estimate flood severity and depth in recent US flooding events.
• Established a benchmark through visualized color labels per image for flood image processing and segmentation.

Conclusions:

• The FloodIMG database provides a valuable, freely accessible resource for advancing flood image segmentation and analysis.
• The dataset enables the creation of more accurate and optimized image segmentation models for flood-related applications.
• FloodIMG supports research in disaster response and management through improved machine learning model development.