Flood change detection model based on an improved U-net network and multi-head attention mechanism

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Disaster Management

Background:

• Flood disasters pose significant risks, necessitating accurate and efficient monitoring systems.
• Synthetic Aperture Radar (SAR) imagery is crucial for flood detection due to its all-weather capabilities.
• Existing monitoring methods often lack the precision required for timely disaster response.

Purpose of the Study:

• To develop an advanced deep learning model for improved accuracy and efficiency in flood disaster monitoring.
• To enhance the extraction of flood disaster change information before, during, and after flood events.
• To leverage the Transformer multi-head attention mechanism for better analysis of SAR images.

Main Methods:

• A modified U-Net network architecture was developed, integrating the Transformer multi-head attention mechanism (TM).
• The model was trained on a large dataset of annotated SAR images.
• Performance was evaluated using loss function, accuracy, and precision metrics, comparing against baseline models.

Main Results:

• The modified U-Net model with TM achieved an accuracy of 95.52%, a 3.46% improvement over the baseline U-Net.
• The model demonstrated high accuracy (90.11%) with a low loss value (approx. 0.59), outperforming other algorithms.
• Significant improvements in loss value, accuracy, and precision were observed compared to existing models.

Conclusions:

• The proposed TM-integrated U-Net model offers a novel and effective approach for SAR-based flood disaster monitoring.
• This advancement can significantly enhance disaster response, management, and risk assessment.
• The model's improved accuracy and efficiency contribute to more robust flood event analysis.