TerraClass is a deep learning project focused on the automated classification of satellite imagery for land-use analysis. The primary goal is to accurately categorize 28x28 pixel satellite image tiles into one of six distinct land-use classes: barren land, trees, grassland, roads, water, and buildings. The project explores the effectiveness of both a custom-built Convolutional Neural Network (CNN) and a transfer learning approach using powerful pre-trained models.
This project is a testament to the potential of deep learning in remote sensing, providing a robust framework for environmental monitoring, urban planning, and disaster management.
- Repository: TerraClass on GitHub
- Demo: No live demo yet—run the notebooks locally to see the magic! (Future work: WebApp UI Dev & Deployment)
- Custom CNN: A custom-designed, lightweight Convolutional Neural Network from scratch that serves as a highly effective baseline model.
- Transfer Learning: Integration of pre-trained VGG16 and ResNet50 models to leverage powerful feature extraction capabilities.
- High-Resolution Data: Utilizes high-resolution satellite imagery from the National Agriculture Imagery Program (NAIP) via the DeepSat-6 dataset.
- High Accuracy: The custom CNN model achieved a best validation accuracy of 97.5%, outperforming the transfer learning models. The transfer learning approach attained a maximum of 96.4% validation accuracy.
- Comprehensive Evaluation: Provides detailed analysis of model performance through training/validation plots, classification reports, and confusion matrices.
- Optimized Workflows: The project demonstrates a complete workflow from data preprocessing and model training to final evaluation and analysis of misclassifications.
- Framework: TensorFlow, Keras
- Language: Python
- Data Handling: NumPy, pandas,
scipy.io - Visualization: Matplotlib, Seaborn
- Machine Learning: Scikit-learn
- Environment: Jupyter Notebook
- Version Control: Git, Git LFS for large files
|
This project utilizes the DeepSat (SAT-6) Airborne Dataset. ➡️ Instructions to download are provided in
|
|
TerraClass-Terrain-Classification-using-Deep-Learning/
├── DL_Cust_Model/ # Custom CNN model (baseline)
│ └── Custom CNN.ipynb
│
├── DL_Res50_Model/ # Transfer learning with ResNet50
│ └── ResNet50.ipynb
│
├── DL_VGG16_Model/ # Transfer learning with VGG16
│ └── VGG16.ipynb
│
├── dataset/
│ ├── Dataset_Instructions.pdf # Kaggle dataset link & setup
│ ├── res_baseline.weights.h5 # ResNet50 trained weights
│ ├── res_history.csv # ResNet50 training history
│ ├── sat_img.png # Sample from DeepSat-6 dataset
│ ├── vgg_baseline.weights.h5 # VGG16 trained weights
│ └── vgg_history.csv # VGG16 training history
├── LICENSE.md
└── README.md To set up the project on your local machine, follow these steps:
- Clone the repository:
git clone https://yourusername/TerraClass-Terrain-Classification-using-Deep-Learning.git cd TerraClass-Terrain-Classification-using-Deep-Learning - Download the Dataset: Follow the instructions provided in
dataset/Dataset_Instructions.pdfto download theDeepSat SAT-6files from Kaggle and place it in thedataset/folder. - Install Dependencies: Ensure you have Python 3.8+ and all the necessary Python libraries installed.
pip install tensorflow keras numpy pandas matplotlib seaborn scikit-learn scipy
Once the installation is complete, you can explore the project's models and results by opening the Jupyter Notebooks.
- To run the Custom CNN model, open
DL_Cust_Model/Custom CNN.ipynb. - To run the VGG16 transfer learning model, open
DL_VGG16_Model/VGG16.ipynb. - To run the ResNet50 transfer learning model, open
DL_Res50_Model/ResNet50.ipynb.
The notebooks are configured to load pre-trained weights by default, allowing for quick evaluation of the models. If you wish to retrain a model, simply set the load_weights variable to False in the respective notebook.
The models were rigorously evaluated on a held-out (unseen) test dataset of 81,000 images. Notably, the custom-built CNN delivered strong performance, highlighting that a well-designed lightweight model can compete closely with deeper, pre-trained architectures for this specific task.
| Model | Test Accuracy |
|---|---|
| Custom CNN | 97.5% |
| VGG16 (Transfer) | 96.4% |
| ResNet50 (Transfer) | 79% |
Key Observations:
- The custom CNN outperformed the transfer learning models, achieving the highest accuracy.
- Evaluation using classification reports and confusion matrices revealed that the most frequent classification errors involved confusing
barren landwithgrasslandandgrasslandwithtrees. - These misclassifications suggest that the spectral or spatial characteristics of these land types may overlap in the feature space used by the model.
Based on the project's findings and analysis, future work could focus on several key areas to enhance the model's capabilities and practical utility.
- Refined Classification Workflow: A more refined workflow could be implemented to improve spatial accuracy. This would involve classifying partially overlapping image tiles and assigning classification results specifically to the center-most pixels of each tile.
- Enhanced Pixel-Level Resolution: This refined approach would increase the spatial accuracy of the final classification map by enhancing pixel-level resolution, leading to finer granularity in land use delineation across large areas.
- Spatial Metadata Restoration: Future work could extend the application to include a workflow for georeferencing, which involves restoring the spatial metadata of the reassembled land use classification.
- Complex Spatial Evaluations: This enhancement would enable more complex spatial evaluations, such as calculating statistics over different regions (e.g., class distribution by county).
- Interactive User Interfaces: Developing user interfaces to interact with the classified output would further enhance the utility of the system.
- Adapting to Diverse Datasets: Future research may focus on adapting the framework to more diverse satellite datasets for generalization.
- Improving Class Differentiability: Efforts should also be made to improve the model's ability to differentiate between similar classes, such as grassland and barren land, which were identified as challenging during misclassification analysis.
- We sincerely acknowledge the Original Creators of the DeepSat-6 dataset for making this data publicly available.
- Special thanks to the Open-Source community for tools like TensorFlow, Keras, and more.
This project is licensed under the MIT License. See LICENSE.md for more info.
If you have questions, find bugs, or want to contribute, feel free to:
- Open an issue in this repo
- Contact me at saisameer.kumar95@gmail.com
- Connect on LinkedIn
“The best way to get a project done faster is to start sooner.” – Jim Highsmith
Don’t hesitate to reach out — collaboration makes it better!