EuroSat-RGB-Classifiers

This repository contains three different models fine-tuned on the EuroSAT dataset, along with their performance comparisons.

Models Used

1. ResNet-18 (pre-trained on ImageNet 1k)
2. ResNet-50 (pre-trained on ImageNet 1k)
3. ViT-Base-Patch16-224 (pre-trained on ImageNet 21k and fine-tuned on ImageNet 1k)

Dataset

The dataset comprises JPEG composite chips extracted from Sentinel-2 satellite imagery, representing the Red, Green, and Blue bands. It encompasses 27,000 labeled and geo-referenced images across 10 Land Use and Land Cover (LULC) classes

Splits : Train 80% Validation 10% Test 10% (original dataset's label distribution is consistent in each split)

Link: Eurosat-RGB Dataset

Models' Performance

Links to final models:

• resnet18-eurosat
• resnet50-eurosat
• vit-base-patch16-224-eurosat

Model's loss and accuracy in each phase:

Model	Phase	Avg Loss	Accuracy
resnet18-eurosat	Train	0.097586	97.01%
	Validation	0.071375	97.70%
	Test	0.068443	97.74%
resnet50-eurosat	Train	0.076420	97.56%
	Validation	0.054377	98.30%
	Test	0.058930	98.07%
vit-base-patch16-224-eurosat	Train	0.012038	99.61%
	Validation	0.023757	99.04%
	Test	0.040557	98.67%

Model's accuracy, precision, recall and F1 score:

Model	Accuracy	Precision	Recall	F1
resnet18-eurosat	97.74%	0.97747	0.97741	0.97740
resnet50-eurosat	98.07%	0.98078	0.98074	0.98074
vit-base-patch16-224-eurosat	98.67%	0.98673	0.98667	0.98668

Confusion Matrices:

• resnet18-eurosat

• resnet50-eurosat

• vit-base-patch16-224-eurosat

Usage

For details on the model training procedure and its parameters, please refer to the Jupyter notebooks provided. For a list of required dependencies, refer to requirements.txt.

Conclusion

• ViT-base-patch16-224 appears to be the best performing model in terms of accuracy, precision, recall, and F1-score, making it a robust choice for the EuroSAT dataset.
• ResNet50 also performs exceptionally well and could be considered if a ResNet architecture is preferred or if there are computational constraints.
• ResNet18, while slightly behind the other two models, still offers strong performance and could be a good choice for scenarios where a smaller, less complex model is desired.