Urban Land Use Classification from Public Domain Imagery

In this project, we investigate the urban land use dataset from UC Merced that consists of 21 land use categories from various urban areas across the United States. We explore various features of the images, starting from simple ones such as mean color values of the color channels and histogram of oriented gradients, to complex features such as bag of visual words, contours of an image, and embeddings from pre-trained deep neural network models (e.g., VGG16 and ResNet101). We perform principal component analysis to reduce the dimensionality of our dataset for better generalizability. We identify features suitable for classifying the images into their respective classes through experiments with various classification models. Finally, we compare and contrast these classification models for efficiency (with respect to the time required for training and prediction) and model accuracy.