Geospatial Learning Playground

A hands-on learning repo working through deep learning on geospatial and image data — from satellite classification to neural architecture search. Each sub-project is small, self-contained, and documents what was learned along the way.

---

Projects

01 — EuroSAT Classification

Fine-tunes a ResNet18 (pretrained with MoCo on Sentinel-2) on the EuroSAT dataset for 10-class land-use classification. Includes an interactive matplotlib app that stitches random test tiles into a fake scene and overlays predicted classes with color-coded outlines.

RGB Scene	Classification Overlay

---

02 — MiniNAS

A lightweight Neural Architecture Search framework on FashionMNIST. Defines a search space of 144 CNN configurations, runs a random search over 50 sampled configs, and ranks them by accuracy, parameter count, and energy consumption. Includes a full 6-plot analysis and visualization pipeline.

---

Tech Stack

• TorchGeo — datasets, pretrained models & transforms for geospatial data
• PyTorch — deep learning framework
• Kornia — differentiable data augmentation
• Matplotlib — visualization & interactive UI
• Seaborn — statistical heatmaps & plots

Setup

# Clone the repo
git clone https://github.com/jonaebel/GeospatialLearning
cd GeospatialLearning

# Create a virtual environment
python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Notes

• Large data (data/), model weights (*.pt), and secrets (.env) are intentionally excluded via .gitignore.
• Datasets are downloaded automatically on first run.

---

_{Learning project · work in progress}