Projects:2018s1-109 High-Resolution Change Prediction using Sparse Spatio-temporal Data

From Projects
Jump to: navigation, search

Supervisor: Brian Ng | Students: Liam Mellor, Lucas Sargent | Consilium PoC: Anthony Milton


This project aims to develop a machine learning system capable of predicting changes in high resolution satellite imagery. A convolutional neural network can learn the spatial transforms undergone by a series of images regardless of size, this project will focus on applying transforms learnt in one resolution to images of a higher resolution. Starting with a sequence of lower resolution images taken at various points in time, and given an initial high resolution image, the system should be able to generate a higher resolution counterpart to match any time spanned by the lower resolution series. The lower resolution series needs to be considered sparse as a variety of factors including exact satellite position and cloud coverage will affect the information available for the network to learn. The application of this project is to increase the frequency of available high resolution satellite images, however the approach taken and techniques learned could be applied to a variety of lossy data problems, such as video compression or small sensor amalgamation. This project requires students to: gain knowledge in deep and convolution machine learning technologies, develop and adapt machine learning techniques for satellite imagery (GIS data) and build robust machine learning networks capable of operating with minimal homogeneous data.