Autonomous Surface Analysis on Moon & Mars using Convolutional Neural Networks
This repository implements a Deep Learning framework designed for Automated Crater Detection on planetary surfaces. Utilizing high-resolution imagery from Lunar and Martian missions, the project treats craters as "anomalies" or specific features in a vast landscape, enabling autonomous navigation and geological mapping for future space exploration.
The core objective is to leverage Convolutional Neural Networks (CNNs) to identify and segment craters of varying sizes, shapes, and illumination conditions.
The model is trained on diverse surface imagery from Moon and Mars datasets. Unlike standard object detection, planetary imagery presents unique challenges such as:
- Low Contrast: Distinguishing shallow craters from surrounding terrain.
- Variable Lighting: Shadows that mimic craters or obscure edges.
- Scale Invariance: Detecting both massive impact basins and small "pockmark" craters.
To improve model robustness, the following image processing techniques were applied:
- Histogram Equalization: Enhancing contrast in low-light Martian terrain.
- Data Augmentation: Random rotations, flips, and zoom to simulate different orbital perspectives.
- Normalization: Scaling pixel values to optimize gradient descent during training.
The model utilizes a layered Convolutional Architecture to extract hierarchical features:
- Convolutional Layers: Capture edge detection and circular patterns.
- Pooling Layers: Reduce spatial dimensions while retaining critical "crater" features.
- Dropout Regularization: Implemented to prevent overfitting on specific surface textures.
The model was evaluated using standard Computer Vision metrics to ensure high precision in autonomous environments:
- Mean Average Precision (mAP): High accuracy in identifying crater boundaries.
- IOU (Intersection Over Union): Measuring the overlap between predicted craters and ground-truth labels.
- F1-Score: Balancing detection sensitivity to avoid "False Positive" boulders or shadows.
- Frameworks: TensorFlow / Keras / PyTorch
- Computer Vision: OpenCV
- Data Handling: NumPy, Scikit-Image
- Visualization: Matplotlib (for prediction overlays)
- Transfer Learning: Fine-tuning pre-trained models like ResNet or U-Net for higher segmentation accuracy.
- Real-time Detection: Optimizing the model for deployment on edge-computing devices for rovers.
- Multimodal Data: Integrating LIDAR/Depth data to distinguish crater depth from surface shadows.
Maintained by DHARKIVE-STUDIO