This project investigates satellite image classification using the EuroSAT dataset. A custom Convolutional Neural Network (CNN) was first developed as a baseline model and subsequently compared against two state-of-the-art pretrained architectures using transfer learning: ResNet-50 (Microsoft, 2015) and ConvNeXT-tiny (Facebook/Meta, 2022). The objective is to evaluate the effectiveness of transfer learning for remote sensing image classification, with focus on predictive performance, computational efficiency, and practical deployment considerations. Results demonstrate that transfer learning achieves up to 96.26% test accuracy, representing a gain of approximately 15 percentage points over the custom baseline. ConvNeXT-tiny is identified as the optimal deployment candidate, achieving 95.16% accuracy with only 7,960 trainable parameters trained over five epochs.
The repository contains 3 main file:
- Python code .py format
- Project Report in pdf format
- Presentation in pdf. format