Progressive Attention-Enhanced EfficientNet-UNet for Robust Water-Body Mapping from Satellite Imagery.

Area of Science:

• Remote Sensing
• Computer Vision
• Environmental Science

Background:

• Accurate identification of water bodies in satellite imagery is crucial for sustainable water resource management and climate-resilient infrastructure.
• Existing segmentation methods often struggle with complex water-body patterns and boundaries.

Purpose of the Study:

• To develop a novel deep learning architecture for high-fidelity water body extraction from satellite data.
• To improve the precision, sensitivity, and overall accuracy of water body segmentation using attention mechanisms.

Main Methods:

• Integration of a Convolutional Block Attention Module (CBAM) into a modified EfficientNet-UNet backbone.
• Rigorous training using five-fold cross-validation, dynamic test-time augmentation, and Lovász loss optimization.
• Evaluation on an independent test set using metrics such as precision, sensitivity, specificity, accuracy, Dice score, and IoU.

Main Results:

• The proposed model achieved high performance metrics: precision (90.67%), sensitivity (86.96%), specificity (96.18%), accuracy (93.42%), Dice score (88.78%), and IoU (79.82%).
• Demonstrated significant improvement over conventional segmentation pipelines in extracting complex water-body features.
• Validated the effectiveness of attention mechanisms for detailed water body boundary delineation.

Conclusions:

• Attention-guided deep learning networks offer a robust and efficient pathway for high-fidelity water body mapping.
• The developed model is computationally efficient and suitable for large-scale water resource and ecosystem monitoring.
• This research contributes a tailored UNet-style architecture with CBAM for remote sensing applications.