This project implements a satellite image classification system using ResNet50 and transfer learning on the EuroSAT dataset. The model learns to classify satellite imagery into 10 different land-use and land-cover categories.
The training process was performed in two stages:
- Feature Extraction using a frozen ResNet50 backbone.
- Fine-Tuning by unfreezing the last 30 layers of ResNet50.
The final model achieved 98.02% validation accuracy, demonstrating the effectiveness of transfer learning for remote sensing image classification.
EuroSAT Dataset
↓
Data Augmentation
↓
ResNet50 (ImageNet Weights)
↓
Feature Extraction
↓
Fine-Tuning
↓
Classification Layer
↓
Prediction
Dataset: EuroSAT
Classes:
- Annual Crop
- Forest
- Herbaceous Vegetation
- Highway
- Industrial
- Pasture
- Permanent Crop
- Residential
- River
- Sea/Lake
Image Type: RGB Satellite Images
Confusion matrix of the fine-tuned ResNet50 model on the EuroSAT dataset.
Sample predictions generated by the trained model on satellite images.
Class distribution of the EuroSAT dataset used for training and evaluation.
- ResNet50 (Pre-trained on ImageNet)
- Transfer Learning
- Global Average Pooling Layer
- Dense Classification Layer
- Softmax Activation
- All ResNet50 layers frozen
- Trained only the classification head
- Validation Accuracy: 95.37%
- Unfroze the last 30 layers of ResNet50
- Reduced learning rate for stable convergence
- Validation Accuracy: 98.02%
- Data Augmentation
- Transfer Learning using ImageNet pretrained weights
- Fine-Tuning of ResNet50 layers
- Learning Rate Adjustment
- Model Checkpointing
| Training Phase | Validation Accuracy |
|---|---|
| Frozen ResNet50 | 95.37% |
| Fine-Tuned ResNet50 | 98.02% |
- Deep Learning
- Transfer Learning
- Convolutional Neural Networks (CNNs)
- Computer Vision
- Satellite Image Classification
- Data Augmentation
- TensorFlow/Keras
- Model Fine-Tuning
- Performance Evaluation
- Confusion Matrix Analysis
- Python
- TensorFlow
- Keras
- NumPy
- Matplotlib
- Scikit-Learn
- Google Colab
satellite-image-classification/
│
├── resnet50_training.ipynb
├── model_evaluation.ipynb
├── README.md
├── requirements.txt
└── screenshots/
├── confusion_matrix.png
├── model_predictions.png
└── dataset_distribution.png
- Compare performance with EfficientNet and Vision Transformers (ViT)
- Perform hyperparameter optimization using Optuna or Grid Search
- Deploy the model using Streamlit or Flask
- Extend to higher-resolution satellite imagery datasets
- Explore ensemble learning techniques
- Implement explainable AI techniques such as Grad-CAM
Developed a high-accuracy satellite image classification model using transfer learning with ResNet50, achieving 98.02% validation accuracy on the EuroSAT dataset through data augmentation and fine-tuning of pretrained ImageNet weights.
Yash Raj Ravi
B.Tech – Artificial Intelligence and Data Engineering Indian Institute of Information Technology (IIIT) Kota