Abstract
The digital elevation model (DEM) provides critical data for 3D terrain modeling. However, obtaining continuous, high-density elevation data is challenging due to complex terrains and high measurement costs. Traditional spatial interpolation methods often suffer from low complexity and high computational costs, leading to poor real-time performance. This study treats DEM interpolation as an image generation task, where missing values are filled to create a complete DEM image. To address privacy concerns surrounding sensitive geographic data, we propose a DEM interpolation model based on Federated Learning (FL) and a MultiScale U-Net. Experimental results demonstrate that this model achieves faster processing speeds and higher precision compared to traditional methods, offering a secure and efficient approach for terrain information reconstruction.
Introduction
Digital Elevation Models (DEMs) are fundamental to various applications, including infrastructure planning, environmental monitoring, and disaster management. Despite their importance, high-resolution DEMs are often incomplete due to the prohibitive costs of field surveys. Spatial interpolation is typically used to fill these gaps. However, traditional geostatistical methods like Kriging are computationally intensive and struggle with the non-linear complexities of global terrain.
Furthermore, high-quality DEM data is often restricted by national security and privacy regulations. Centralizing such data for deep learning training is frequently impossible. This research addresses these challenges by combining the privacy-preserving capabilities of Federated Learning with the high-performance feature extraction of a MultiScale U-Net.
Methodology
Federated Learning Framework
To maintain data privacy, the study employs a Federated Learning (FL) architecture. Instead of uploading raw terrain data to a central server, local clients (e.g., regional data centers) train models on their private datasets. Only the resulting model parameters are transmitted to a central server for aggregation.
The authors specifically implemented the SCAFFOLD algorithm. SCAFFOLD improves upon standard federated averaging by using control variates to mitigate "client drift," which occurs when local data distributions (e.g., different geographic regions) vary significantly.
MultiScale U-Net Architecture
The core interpolation engine is a modified U-Net designed to capture terrain features at multiple scales. Unlike standard convolutions with fixed receptive fields, the MultiScale U-Net incorporates parallel branches with varying kernel sizes (1x1, 3x3, and 5x5). This allows the model to simultaneously learn fine-grained local details and broader topographic trends.
The model is trained using a joint loss function that combines L1 Content Loss for pixel-level accuracy and Perceptual Loss (via a pre-trained VGG19) to ensure the structural and semantic realism of the reconstructed terrain.
Experimental Setup and Results
The model was evaluated using the ETOPO 2022 Global Relief Model. Data was partitioned into 64x64 patches, with 1,200 pixels randomly sampled to simulate sparse observations. The framework was tested across eight simulated local clients representing different global regions.
Performance Comparison
The proposed federated MultiScale U-Net significantly outperformed traditional Kriging interpolation:
- Accuracy: The model achieved a 29.4% reduction in Mean Absolute Error (MAE) and a 25.0% reduction in Root Mean Square Error (RMSE).
- Efficiency: Processing time was reduced to just 0.046 seconds per patch, representing only 8.9% of the time required by Kriging (3.302s).
- Visual Quality: A Structural Similarity Index (SSIM) of 96.3% was achieved, indicating high fidelity to the original terrain.
Conclusion
This research demonstrates that treating spatial interpolation as a computer vision task within a federated learning framework is both viable and superior to traditional methods. By leveraging MultiScale U-Nets, the model effectively captures the fractal nature of terrain while ensuring that sensitive geographic data remains secure on local servers. Future work will focus on optimizing the model for resource-constrained hardware and testing its robustness against even higher levels of data sparsity.
Summary updated