AI Capstone Project: Deep Learning
A deep learning study on land-use classification from satellite imagery, comparing convolutional backbones with Vision Transformers.
What I did
- Engineered an end-to-end deep learning pipeline that combines CNN feature extraction with Vision Transformers (ViT) for land-use classification.
- Designed and trained custom CNN and ViT architectures from scratch, using data augmentation, regularization and dynamic early stopping to reduce overfitting by ~15%.
- Tuned learning rates, batch sizes and optimizer schedules, speeding up training convergence by ~35%.
- Reached 97.2% classification accuracy on the EuroSAT dataset, benchmarking convolutional backbones against transformer latencies with confusion matrices.
Tech stack
PythonPyTorchTensorFlow/KerasCNNViTData Augmentation