Machine-Learning: Mars Frost Detection

🔷 Identification of Frost in Martian HiRISE Images

🔶 Overview

This project focuses on binary classification of Martian terrain images to detect frost using deep learning models. A custom CNN + MLP model and transfer learning (EfficientNetB0, ResNet50, VGG16) were used for feature extraction and classification. Data augmentation, dropout, batch normalization, and L2 regularization were applied to improve model generalization.

🔷 Dataset Used

• HiRISE Mars Terrain Images (Dataset Link)
• The dataset contains 119,920 image tiles extracted from 214 HiRISE subframes.
• Each image tile is labeled as either "frost" or "background".
• Provided splits for training, validation, and testing.

🔷 Libraries Used

• TensorFlow/Keras - Building deep learning models.
• OpenCV - Image augmentation (cropping, zooming, rotating, flipping).
• Matplotlib & Seaborn - Data visualization.
• scikit-learn - Model evaluation metrics.
• Pandas & NumPy - Data processing and numerical computations.

🔷 Steps Taken to Accomplish the Project

🔶 1. Data Preprocessing & Augmentation

• Resized images to 299x299 pixels.
• Applied image augmentation: random cropping, zooming, flipping, contrast adjustment, and translation.
• Normalized pixel values between 0 and 1.

🔶 2. Training a CNN + MLP Model

• Built a 3-layer Convolutional Neural Network (CNN) followed by a Multi-Layer Perceptron (MLP).
• Used ReLU activation for all layers.
• Applied softmax function for binary classification.
• Regularization techniques used:
  • Batch Normalization
  • Dropout (30%)
  • L2 Regularization
• ADAM optimizer and cross-entropy loss were used.
• Model trained for at least 20 epochs, with early stopping based on validation loss.
• Precision, Recall, and F1-score were reported.

🔶 3. Transfer Learning with Pre-trained CNNs

• Utilized EfficientNetB0, ResNet50, and VGG16 for feature extraction.
• Froze all layers except the final fully connected layer.
• Extracted features from the penultimate layer and trained a classifier.
• Used ReLU activation, batch normalization, dropout (30%), and softmax activation.
• Trained for at least 10 epochs (preferably 20 epochs) with early stopping.
• Compared results with CNN + MLP model.

🔶 4. Model Evaluation & Analysis

• Reported Precision, Recall, and F1-score for all models.
• Compared CNN + MLP vs. Transfer Learning performance.
• Plotted training and validation loss curves to analyze convergence.

🔶 5. Findings & Comparison

• CNN + MLP required more training data but performed well with augmentation.

• Transfer Learning (EfficientNetB0, ResNet50, VGG16) achieved higher accuracy due to pre-trained feature extraction.

• EfficientNetB0 provided best performance in terms of validation loss and classification accuracy.

  ---

📌 Note

This repository contains a Jupyter Notebook detailing each step, along with results and visualizations.