Lunar Crater Detection and Depth Analysis

Overview

This project leveraged deep learning techniques to detect craters on the lunar surface using satellite imagery from India's Chandrayaan-2 mission. It was built as part of a research initiative for lunar mapping and resulted in a publication in the Journal of the Indian Society of Remote Sensing.

Data Preparation

The imagery was captured by the Optical High Resolution Camera (OHRC) on Chandrayaan-2, covering areas of 12km x 3km with a ground resolution of 0.19m (19cm). Each image contained approximately 1 million pixels.

Given the enormous image size, we used the PDS4 (Planetary Data System) format to crop each image into smaller 640x640 pixel segments covering 120m x 120m each. Over 900+ images were manually labelled using RoboFlow.

Model Architecture

Two object detection models were trained:

• R-CNN for baseline crater detection
• YOLOv5 for faster, real-time capable detection

Both models were trained on Google Colab Pro to handle the computational requirements of the large dataset.

Geospatial Analysis

Beyond detection, we implemented a geospatial coordinate tracking and depth analysis algorithm to estimate the location and depth of detected craters. This information is critical for identifying potential landing spots on the lunar surface.

Results

• 85-92% detection accuracy on test datasets
• Successfully identified craters ranging from 10-100m in diameter
• Depth estimation based on crater diameter correlation

Challenges

The biggest challenge was manual labelling — some smaller craters (1-10m diameter) were missed during annotation. Since the research focused on suggesting landing spots, omitting sub-10m craters was acceptable. Techniques like oversampling and data augmentation were used to address underrepresentation of smaller features.

Outcome

Published in the Journal of the Indian Society of Remote Sensing — recognized as an innovative application of AI in space exploration.