Abstract:
Google Earth Engine has a multi-petabyte catalog of open source satellite imagery and geospatial datasets available to scientists and researchers to analyze. One interesting problem is extracting demographic information from these satellite images, such as the population of a given area. Our goal is to build neural network models to estimate the population from satellite images. This is a difficult task, as the model must learn land use types within an image in order to identify high- versus low-population areas. Previous sources have worked on a similar problem using deep sequential models and UNet architectures. We expand this by building many models including FFNNs, CNNs, UNets, and also investigating the value of transfer learning for this task.
See the "AM216_Final_Writeup.ipynb" file in this repository for the full report.