RS-Dseg: semantic segmentation of high-resolution remote sensing images based on a diffusion model component with...

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Semantic segmentation is vital for remote sensing image interpretation, particularly for high-resolution data with intricate details.
• Challenges include extracting fine semantic information and managing class imbalance in multiclass segmentation tasks.

Purpose of the Study:

• To propose a novel approach using diffusion models for enhanced semantic segmentation of remote sensing images.
• To address challenges in feature extraction and class imbalance in high-resolution remote sensing data.

Main Methods:

• Utilized diffusion models built on the UNet architecture for capturing multiscale features and contextual information.
• Incorporated a lightweight classification module with spatial-channel attention to focus on significant feature regions.
• Employed unsupervised pre-trained components to accelerate model convergence.

Main Results:

• Achieved state-of-the-art performance on Postdam, GID, and Five Billion Pixels datasets.
• On the GID dataset, obtained 96.99% overall accuracy, 92.17% mean IoU, and 95.83% mean F1 score.
• Demonstrated rapid convergence, achieving good performance within 30 training cycles.

Conclusions:

• The proposed diffusion model approach significantly improves semantic segmentation accuracy for remote sensing images.
• The method offers reduced parameters, faster training speeds, and superior performance compared to existing models.
• Effective for handling complex spatial information and texture structures in high-resolution remote sensing imagery.