happy36/eurosat-efficientnet

Model Card: EuroSAT Land Cover Classifier

Model Details

• —Architecture: EfficientNetB0 (transfer learning)
• —Input: 64x64 RGB satellite tiles
• —Output: 10-class softmax (AnnualCrop, Forest, HerbaceousVegetation, Highway, Industrial, Pasture, PermanentCrop, Residential, River, SeaLake)
• —Framework: TensorFlow/Keras

Training Data

• —Dataset: EuroSAT (27,000 images, 10 classes)
• —Split: 80% train / 20% validation

Performance

• —Validation accuracy: 0.9124 (frozen base, epoch 10)
• —Note: fine-tuning caused temporary degradation; see training notes.

Intended Use

• —Land-use/land-cover classification from Sentinel-2-style satellite tiles.
• —Not validated for other sensors, resolutions, or geographic regions outside the EuroSAT source imagery.

Limitations

• —Confusion observed between visually similar classes (e.g., PermanentCrop vs AnnualCrop, Highway vs Residential).
• —Not tested for adversarial robustness in production settings (see FGSM analysis, cell 30).

Ethical Considerations

• —Satellite classification models can be used for surveillance or land-monitoring purposes; consider use-case context and consent/regulatory frameworks in deployment.