ANMOL1140W/mae-remote-sensing

EuroSAT Land Cover Classifier

Interactive demo for Masked Autoencoder (MAE) fine-tuning on the EuroSAT remote sensing dataset.

Upload a satellite image or choose an example to classify land cover among 10 EuroSAT categories.

Features

• —Upload custom RGB satellite patches
• —Built-in example images (Industrial, Forest, River, Highway, Residential)
• —Top-5 predictions with confidence scores
• —CPU-only inference for Hugging Face Spaces deployment

Model

• —Encoder: ViT-style MAE encoder (64×64 input, 4×4 patches, 128-dim embeddings)
• —Head: Linear classification head (ViTClassifier)
• —Checkpoint: checkpoints/model_best.pth
• —Classes: Loaded from labels.json

Local run

pip install -r requirements.txt
streamlit run app.py

Project layout

app.py              # Streamlit UI
utils/              # Inference, preprocessing, checkpoint loading
models/             # MAE encoder + ViT classifier
labels.json         # EuroSAT class names
checkpoints/        # Fine-tuned weights
examples/           # Sample images
configs/config.yaml # Model/dataset settings