Spatial interpolation of global DEM using federated deep learning.

Area of Science:

• Geoinformatics
• Remote Sensing
• Computer Science

Background:

• Digital Elevation Models (DEMs) are crucial for 3D terrain modeling but acquiring high-density data is challenging and costly.
• Traditional spatial interpolation methods for DEM restoration suffer from low real-time performance and precision due to high computational costs.
• Deep learning excels at image generation tasks, but DEM data privacy concerns limit centralized training.

Purpose of the Study:

• To develop a novel DEM interpolation model addressing data scarcity and privacy issues.
• To enhance the efficiency and security of terrain information processing.
• To explore the application of Federated Learning (FL) in DEM data restoration.

Main Methods:

• Proposed a DEM interpolation model integrating Federated Learning (FL) with a multiScale U-Net architecture.
• Utilized FL to enable local model training across multiple nodes, preserving data privacy.
• Treated DEM interpolation as an image generation task, inputting incomplete DEMs to generate complete ones.

Main Results:

• The FL-based multiScale U-Net model demonstrated a faster processing speed compared to traditional methods.
• The model achieved a lower interpolation precision than traditional methods, indicating a trade-off between speed and accuracy.
• The study successfully provided a privacy-preserving solution for DEM data utilization.

Conclusions:

• Federated Learning combined with multiScale U-Net offers an efficient and secure method for DEM interpolation.
• This approach is particularly valuable for applications with strict DEM data privacy and security requirements.
• The research opens new avenues for leveraging sensitive terrain information.