Work in progressSpace technologies · Super-resolution
A self-attention GAN that reconstructs high-resolution lunar terrain from coarse orbital elevation data, developed within İTÜ UZDES's space-technologies program.
- Setting
- İstanbul Technical University — UZDES (Space Systems Design and Test Laboratory), space technologies track.
- Method
- A Self-Attention GAN (SA-GAN) generator/discriminator pair upsamples 60 m/px lunar digital elevation models (DEMs) toward 2 m/px resolution.
- Data
- Digital Terrain Model and PDS-derived lunar elevation data, organized into training pairs with dataset and metadata manifests.
- Related program work
- The same UZDES space-technologies effort also includes autonomous lunar-rover path planning (global A* search with SAC-based local control) over NASA LRO topography.
Evidence boundary: this is an active work-in-progress research project. The repositories document architecture, data preparation, and training infrastructure; a completed quantitative evaluation (e.g. PSNR/SSIM against held-out lunar terrain) is not yet published.
Expanded project brief
Reconstructing lunar terrain detail with a self-attention GAN.
Lunar Image Enhancement is developed within İstanbul Technical University's UZDES program as part of a broader space-technologies effort. The core problem is resolution: publicly available lunar digital elevation models (DEMs) are widely available at roughly 60 meters per pixel, which is too coarse to resolve small craters, boulders, and slope hazards relevant to landing-site assessment and rover traverse planning.
The SA-GAN repository (ethosoftai/SA-GAN) implements a Self-Attention Generative Adversarial Network that learns to upsample 60 m/px DEM tiles toward 2 m/px resolution. Self-attention layers in the generator and discriminator are intended to preserve long-range structural consistency across a tile—ridgelines and crater rims—rather than only sharpening local texture, which is a known failure mode of convolution-only super-resolution networks on terrain data. Training pairs are built from Digital Terrain Model and PDS-derived elevation sources, tracked through dataset and metadata manifests, with the training loop optimized for multi-GPU nodes using mixed precision and channels-last tensors.
The same UZDES space-technologies track also maintains the TUA repository (Ahmet2001/TUA), which houses an autonomous lunar-rover navigation stack—global A* path planning combined with Soft Actor-Critic (SAC) reinforcement learning for local trajectory tracking over NASA LRO topography—alongside 3D models for an Ultracortex EEG rig used in a related hardware effort. The enhanced DEM output from SA-GAN is intended to eventually feed higher-fidelity terrain into that navigation and simulation pipeline.
Status: this is an active work in progress. Architecture, data preparation, and training infrastructure are implemented and documented in the repositories; a held-out quantitative evaluation (for example PSNR/SSIM against ground-truth high-resolution terrain) has not yet been published, and no landing-site or mission-readiness claim is made.
