Satellite Imagery Classification: CNN & ViT Hybrid

Project Overview

This project provides a deep learning pipeline for the binary classification of satellite imagery (agricultural vs. non-agricultural land). Designed for scalability and high performance, the repository demonstrates the implementation, comparison, and integration of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) using both TensorFlow/Keras and PyTorch.

Data Pipeline

• Dataset: 6,000 perfectly balanced satellite images (3,000 class_1_agri, 3,000 class_0_non_agri).
• Data Loading: Emphasizes memory-efficient sequential/lazy loading (tf.data.Dataset / PyTorch DataLoader) over naive bulk loading to mitigate RAM bottlenecks and optimize GPU I/O synchronization.
• Preprocessing: Standardized image resizing (64x64), normalization, and on-the-fly data augmentation to improve model generalization.

Model Architectures

1. Baseline CNNs: Custom networks built in both Keras and PyTorch. Focuses on hierarchical spatial learning through stacked convolutional and pooling layers.
2. Vision Transformers (ViTs): Treats images as tokenized patch sequences, utilizing positional encodings and multi-head self-attention to model global spatial relationships.
3. CNN-ViT Hybrid: An optimized architecture utilizing transfer learning. A pre-trained CNN backbone (e.g., ResNet50) extracts local features and reduces dimensionality, feeding spatial feature maps into Transformer encoder blocks for global context modeling.

Training & Evaluation

• Hyperparameters: Adam optimizer, Binary Cross-Entropy loss, dynamic learning rate schedules (step decay), and dropout for regularization.
• Metrics Evaluated: Accuracy, Precision (minimizing false positives for resource allocation), Recall, F1-Score, ROC-AUC, and Confusion Matrices.
• Performance: Optimized hybrid models achieve >95% accuracy.

Tech Stack

• Frameworks: PyTorch, TensorFlow / Keras
• Core Concepts: Computer Vision, Transfer Learning, Self-Attention, Scalable Data Pipelines