A convolutional neural network that classifies land type from satellite imagery.
Background
Final project for an undergraduate Data Science course. Given 30,000 labeled satellite images across land types such as forest, agriculture, construction, desert, river, and mountain, the goal was to build a model that identifies the land type of any new image.
My role
Built a 3-layer convolutional neural network on top of a pre-trained model to reduce training effort.
Trained on GPU with CUDA and tuned channel widths layer by layer to improve AUC.
Results
Reached over 70% classification accuracy on the held-out data.
Transfer learning made a small custom network practical on a limited compute budget.