Lunar Landing Dataset
Contributors
Supervisor (2):
-
1.
University of Colorado Boulder
-
2.
Politecnico di Milano
Description
This dataset was synthetically generated as part of a Master’s thesis in Space Engineering. It is based on Digital Elevation Model (DEM) data from the Lunar Reconnaissance Orbiter (LRO) mission, specifically the SLDEM - Selene/LRO Digital Elevation Model at 60 m/pixel resolution. The dataset spans longitudes from 0° to 360°E and latitudes from -60° to 60°N, subdivided into 32 tiles of 30° latitude by 45° longitude each. Surface textures were derived from the LROC WAC Hapke Normalized Mosaic, ensuring photometric consistency.
Methods
The dataset was generated using Python-based scripts leveraging the open-source Blender software. The CORTO (Celestial Object Rendering TOol) framework, developed at the DART Lab, Politecnico di Milano, was employed for high-fidelity synthetic rendering. CORTO enables both offline dataset generation and online closed-loop simulations, making it suitable for vision-based spacecraft navigation applications.
Technical info
Geometric Variability:
To simulate realistic spacecraft observations, each render is generated with randomized geometric parameters, ensuring a diverse dataset for generalization across different lunar perspectives. Key parameters include:
- Altitude: 20–120 km
- Latitude & Longitude: -60° to 60°, 0° to 360°
- Azimuth Angle: 0° to 360°
- Elevation Angle: 30° to 90°
- Sun Inclination: 5° to 15°
- Sun Azimuth: -45° to 45°
Tile Distribution:
The lunar surface is divided into 5° × 5° tiles (longitude × latitude) to systematically cover diverse terrains. Each tile undergoes multiple renderings under different conditions, resulting in a dataset of 21,144 high-fidelity synthetic images, distributed as follows:
- Total Number of Tiles: 1,762
- Renders per Tile: 12
Rendering Configuration & Computational Tradeoffs:
Rendering is performed using Blender’s Cycles engine with physically accurate lighting and shading models. Given the need for efficient generation of over 20,000 images, the following optimizations were implemented:
- Reduced Sampling Rate: 4 samples per render, balancing speed and crater detection quality.
- Optimized Resolution: 512 × 512 pixels for computational efficiency.
- Filmic View Transform: Enhancing contrast while preserving grayscale dynamic range.
- CPU-Based Multi-Threading: Enabling parallelized batch rendering.
- Exposure & Energy Scaling: Adjusted based on tile dimensions for consistency.
Final rendering parameters:
- Field of View (FOV): 75°
- Resolution: 512 × 512 pixels
- Film Exposure: 10
- Color Band: Black & White
- Bit Encoding: 8-bit
- View Transform: Filmic
- Sun Angle: 0.53°
- Sun Energy: 10 (tile-dependent)
- Render Engine: Cycles
- Rendering Device: CPU
- Samples per Render: 4
This dataset provides a comprehensive, high-fidelity representation of the lunar surface, supporting the development of robust vision-based navigation algorithms.
Notes
Dataset Folder Structure
The dataset is organized into multiple subfolders, each containing specific data types to support supervised learning and synthetic lunar surface analysis.
1. img/
- This folder contains the synthetic lunar surface renders generated using Blender and the CORTO framework.
- Images are provided in 512 × 512 pixel resolution and rendered using physically accurate shading and lighting models.
2. crater_masks_png/
- Contains crater masks overlaid on the original renders in PNG format.
- The crater annotations are based on the Robbins Crater Database, ensuring accuracy in crater locations and dimensions.
3. YOLO1_labels/
- Provides YOLO-compatible segmentation labels for craters.
- Each crater is annotated using a series of normalized pixel coordinates forming an elliptical contour.
- Format:
⟨class-index⟩ x1 y1 x2 y2 ... xn yn- class-index: Represents the crater category (0 = crater).
- xi, yi: Normalized pixel coordinates of the crater rim.
4. YOLO2_labels/
- Extends YOLO1_labels by including both segmentation and size classification information.
- Enables supervised learning models to differentiate craters based on their scale.
- Format:
⟨class-index⟩ x1 y1 x2 y2 ... xn yn- class-index: Represents the crater’s size category (0 = small, 1 = medium, 2 = large).
- xi, yi: Normalized pixel coordinates of the crater rim.
5. altimeter/
- Contains TXT files simulating altimeter measurements in kilometers.
- These values represent the estimated altitude of the spacecraft above the lunar surface.
6. depth_txt/
- Provides TXT files containing a pixel-wise depth map in kilometers.
- Each value represents the elevation of the lunar surface at that specific pixel.
7. lat_lon_exr/
- Contains OpenEXR (32-bit floating point, RGBA) files storing the pixel-wise coordinates in the lunar reference frame.
- Channels:
- R channel: Stores latitude values.
- G channel: Stores longitude values.