cm93/resnet-50-eurosat

ResNet-50 EuroSat RGB Classifier

Model Details

Model Description: This project implements a ResNet-50 model (pre-trained on ImageNet 1k) for classifying images from the EuroSat RGB dataset across 10 Land Use and Land Cover (LULC) classes.

Model type: ResNet-50

Finetuned from model: Pre-trained ResNet-50 on ImageNet 1k

Model Sources:

• Repository: resnet-50-eurosat
• Demo [optional]: Google Colab

Uses

Direct Use: Classification of images in the EuroSat RGB dataset into 10 LULC classes.

How to Get Started with the Model

import torch
from datasets import load_dataset
from transformers import AutoImageProcessor, AutoModelForImageClassification

# Eurosat-RGB data
eurosat = load_dataset("cm93/eurosat")
image = eurosat["test"][2]["image"]

processor = AutoImageProcessor.from_pretrained("cm93/resnet-50-eurosat")
model = AutoModelForImageClassification.from_pretrained("cm93/resnet-50-eurosat", trust_remote_code=True)

processed_image = processor(image, return_tensors="pt")
processed_image = processed_image.pixel_values

with torch.no_grad():
    prediction = model(processed_image)['logits'].argmax(1).item()
print(model.config.id2label[prediction])
>>> Residential

Training Details

• Base Model: ResNet-50
• Pre-training Dataset: ImageNet 1k
• Target Dataset: EuroSat RGB
• Transfer Learning Method:
  • Phase 1: Feature Extraction: Used the pre-trained ResNet-50 convolutional base to extract features from EuroSat RGB images, replacing and training only the final classification layer.
  • Phase 2: Fine-tuning: Fine-tuned the entire model with a lower learning rate to optimize performance on the EuroSat dataset.
• Image Dimensions: 64x64 pixels
• Optimizer: Adam
• Learning Rate: 1e-3
• Sheduler : ReduceLROnPlateau on validation loss
• Early Stoping patience : 5 epochs
• Batch Size: 64
• Epochs:
  • Feature Extraction: 10 epochs
  • Fine-tuning: 20 epochs
• Loss Function: Cross-entropy loss

Metrics

These resuls were achieved within 7 epochs in fine tuning.

Metric	Training Set	Validation Set	Test Set
Overall Accuracy	99.55%	98.74%	98.70%
Avg Loss	0.015223	0.041319	0.044116