Haifald/vit-eurosat

ViT Fine-Tuned on EuroSAT

This model is a fine-tuned version of google/vit-base-patch16-224 for satellite land-cover classification on the EuroSAT dataset.

Model Details

• Architecture: Vision Transformer (ViT)
• Base model: google/vit-base-patch16-224
• Task: Image Classification
• Dataset: EuroSAT
• Number of classes: 10
• Fine-tuning epochs: 3

Classes

• AnnualCrop
• Forest
• HerbaceousVegetation
• Highway
• Industrial
• Pasture
• PermanentCrop
• Residential
• River
• SeaLake

Evaluation Results

The model was evaluated on the held-out EuroSAT test split of 2,700 images.

Metric	Result
Test Accuracy	98.70%
Macro F1	98.67%

For comparison, a fine-tuned ResNet50 trained on the same data split achieved:

Metric	Result
Test Accuracy	97.41%
Macro F1	97.36%

ViT achieved higher classification performance in this experiment, while ResNet50 trained substantially faster.

Fine-Tuning

The pretrained ViT classification head was adapted to the 10 EuroSAT classes and the model was fine-tuned end to end using Hugging Face Transformers.

Training configuration:

• Optimizer: AdamW
• Learning rate: 5e-5
• Weight decay: 0.01
• Epochs: 3
• Best checkpoint selected using validation macro F1

Usage

from transformers import pipeline

classifier = pipeline(
    "image-classification",
    model="Haifald/vit-eurosat"
)

predictions = classifier("path_to_image.jpg")
print(predictions)

Limitations

• Results are based on one EuroSAT experiment and one dataset split.
• Performance on other satellite datasets was not evaluated.
• Training time depends on the hardware environment.

Project

Full experiment and comparison with ResNet50: GitHub link will be added after the repository is published.