Planes in Satellite Imagery Classification

Concise example notebook that trains a small CNN to classify 20×20 satellite image crops as Plane or NotPlane.

Contents

• Satellite_Plane_Classification.ipynb — the primary notebook that prepares data, builds and trains the model, evaluates results, and saves plane_classifier_cnn.h5.
• Dataset path expected by the notebook: data/planesnet/planesnet (place the dataset here before running).

Quick start

1. Ensure you have Python 3.8+ installed.
2. Install the core dependencies (recommended):

pip install numpy pandas tensorflow scikit-learn matplotlib seaborn pillow

3. Place the downloaded dataset under data/planesnet/planesnet relative to the repository root.
4. Open Satellite_Plane_Classification.ipynb in Jupyter or VS Code and run the cells in order.

What the notebook does (high level)

• Inspects and balances classes in the dataset (the notebook removes 10k random "NotPlane" images as part of balancing).
• Creates TrainValidTest splits and ImageDataGenerators.
• Builds a small CNN (Conv2D layers with 32/64/128 filters, Dense 256) and trains for 20 epochs using class weights.
• Evaluates on a held-out test set, plots metrics and confusion matrix, and saves the trained model as plane_classifier_cnn.h5.

Important notes & warnings

• The notebook performs file operations (randomly deleting and moving images). Keep a backup of your original dataset if you need to preserve it.
• Images are resized to 20×20 prior to training — this is intentional in the example but may not suit all use cases.

Next steps (suggestions)

• Add a requirements.txt for reproducible installs.
• Extract the notebook into a script for non-interactive training runs.

If you'd like, I can also add a requirements.txt or convert the notebook into a runnable training script.