Context & Problem
Monitoring land cover changes — such as urban sprawl, deforestation, and agricultural health — is historically manual and expensive. The introduction of the EuroSat dataset (based on Sentinel-2 satellite images) established a benchmark for automating this process via machine learning.
The core challenge is building a computer vision model that can accurately distinguish between visually similar terrain categories (e.g., permanent crops vs. pastures) across thousands of satellite tiles.
Architecture & Approach
I built an end-to-end computer vision pipeline using TensorFlow and Keras to ingest, process, and classify the image dataset into 10 distinct land-use categories.
The architecture relies on custom Convolutional Neural Networks (CNNs). To prevent the model from overfitting to specific geographic artifacts in the training data, I implemented robust regularization techniques, including aggressive dropout layers and dynamic learning rate adjustments.
Key Results
- Successfully engineered a scalable multi-class image classification model.
- Utilized data augmentation to artificially expand the training set, improving the model's ability to generalize to novel satellite imagery.
- Demonstrated proficiency in deep learning lifecycle management, from tensor preprocessing to model evaluation.