The UC Merced Land Use dataset contains 2100 images, spanning 21 different categories of land use. I was tasked with developing, and subsequently comparing between, two methods for land cover detection.

I opted for practicality and developed two Convolutional Neural Networks (CNNs) using TensorFlow. The first CNN was tailored specifically for the task, trained on the UC Merced Land Use dataset. For the second method, I used transfer learning, leveraging a pre-trained model from the VGG16 dataset — a larger and more diverse collection.

The term paper below details and evaluates the development process of each CNN, providing justifications for the choices made. It then dives into the experimental results of comparing the two methods.