Satellite Land Classification with CNNs and Vision Transformers

This portfolio project classifies satellite image tiles as agricultural or non-agricultural land using deep learning models built in both Keras/TensorFlow and PyTorch.

The work progresses from data-loading experiments to CNN baselines, then to CNN-Vision Transformer hybrid models that combine local feature extraction with global attention.

Project Highlights

• Binary land-use classification from satellite image tiles
• 6,000 image dataset with balanced classes
• Keras and PyTorch data pipelines
• CNN baselines in both frameworks
• CNN-ViT hybrid models in both frameworks
• Cross-framework evaluation using accuracy, precision, recall, F1, ROC-AUC, loss, and confusion matrices

Dataset

The dataset contains 6,000 JPG satellite tiles:

Class	Meaning	Images
class_0_non_agri	Non-agricultural land	3,000
class_1_agri	Agricultural land	3,000

The project data pipeline downloads the dataset from IBM Skills Network cloud storage:

https://cf-courses-data.s3.us.cloud-object-storage.appdomain.cloud/4Z1fwRR295-1O3PMQBH6Dg/images-dataSAT.tar

Large local data files are kept out of Git. See data/README.md for setup notes.

Results

Model	Accuracy	Precision	Recall	F1 Score	ROC-AUC
Keras CNN	0.9925	1.0000	0.9850	0.9924	1.0000
PyTorch CNN	0.9988	0.9983	0.9993	0.9988	1.0000
Keras CNN-ViT Hybrid	0.9958	0.9990	0.9927	0.9958	0.9998
PyTorch CNN-ViT Hybrid	0.9990	0.9990	0.9990	0.9990	1.0000

The PyTorch models achieved the strongest scores in these runs, with the final PyTorch CNN-ViT hybrid reaching 99.90% accuracy.

Repository Structure

.
├── scripts/
│   ├── 01_data_loading_memory_vs_generator.py
│   ├── 02_keras_data_pipeline.py
│   ├── 03_pytorch_data_pipeline.py
│   ├── 04_keras_cnn_classifier.py
│   ├── 05_pytorch_cnn_classifier.py
│   ├── 06_keras_vs_pytorch_cnn_comparison.py
│   ├── 07_keras_cnn_vit_hybrid.py
│   ├── 08_pytorch_cnn_vit_hybrid.py
│   └── 09_final_cnn_vit_evaluation.py
├── src/
│   ├── config.py
│   ├── data_utils.py
│   ├── metrics.py
│   └── visualization.py
├── LICENSE
├── data.md
├── models.md
├── results_summary.md
└── requirements.txt

Python Scripts

The scripts/ folder contains the project workflow as Python source code, organized from data loading through final model evaluation.

The Python scripts are:

1. 01_data_loading_memory_vs_generator.py
2. 02_keras_data_pipeline.py
3. 03_pytorch_data_pipeline.py
4. 04_keras_cnn_classifier.py
5. 05_pytorch_cnn_classifier.py
6. 06_keras_vs_pytorch_cnn_comparison.py
7. 07_keras_cnn_vit_hybrid.py
8. 08_pytorch_cnn_vit_hybrid.py
9. 09_final_cnn_vit_evaluation.py

These scripts are intended for code review, search, and future refactoring.

Setup

Create an environment and install the dependencies:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run or inspect the Python scripts in scripts/ for the source workflow.

Project Background

This project began as an IBM/Coursera deep learning capstone sequence. I reorganized it into a portfolio-ready repository with a clear Python workflow, documented results, reusable helper modules, and GitHub-friendly handling for large data and model artifacts.

Next Improvements

• Add a small inference script for classifying a new satellite tile.
• Export selected plots from model runs into reports/figures/.
• Add a lightweight Streamlit demo for interactive predictions.
• Track experiments with a reproducible configuration file.

License

This project is licensed under the Apache License 2.0. See LICENSE for the full license text.