Deep Convolutional Neural Network for Flood Extent Mapping Using Unmanned Aerial Vehicles Data.

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Artificial Intelligence in Disaster Management

Background:

• Flooding poses a significant threat to urban areas, necessitating rapid flood extent mapping for emergency response and damage assessment.
• Unmanned Aerial Vehicles (UAVs) offer cost-effective, high-resolution imagery acquisition for disaster-affected regions.
• Convolutional Neural Networks (CNNs) have advanced image classification and segmentation tasks, showing potential for analyzing remote sensing data.

Purpose of the Study:

• To investigate the efficacy of CNN approaches for extracting flooded areas from UAV imagery.
• To compare the performance of different CNN architectures (FCN-16s, FCN-8s, FCN-32s) against traditional methods like Support Vector Machines (SVM).

Main Methods:

• Utilized a VGG-based fully convolutional network (FCN-16s) fine-tuned for flood extraction from UAV images.
• Employed k-fold cross-validation for robust model performance estimation on a limited dataset.
• Calculated confusion matrices to quantify classification accuracy and compared results with FCN-8s, FCN-32s, and SVM.

Main Results:

• The FCN-16s model achieved a high classification accuracy of 97.52% for the water class.
• Fully Convolutional Networks (FCNs) demonstrated superior precision in extracting flooded areas compared to SVM.
• FCN-8s achieved the highest accuracy at 97.8%, followed closely by FCN-16s.

Conclusions:

• CNN-based methods, particularly FCNs, are highly effective for precise flood extent mapping using UAV imagery.
• These advanced techniques offer significant improvements over traditional classifiers for disaster management applications.
• The study highlights the potential of deep learning for rapid and accurate geospatial analysis in flood events.