Exploring Avenues for Future Research: Coastline Feature Extraction
Abstract
Coastline mapping is safety critical problem for worldwide shipping. The complexity of global coastlines and lack of labelled images has been a challenging task for modern techniques to map accurately. This paper looks to find a human centred approach by creating a tool that allows experts to label large satellite image datasets. The initial problem space of feature extraction from satellite images via artificial intelligence networks is scoped.
Then dimensionality reduction to map those latent features into a two dimensional tool. A pipeline for such a process has been implemented with a example implementation using Auto-Encoders. In this we find that Auto-Encoders are suitable for feature extraction and can cluster the data well. However when considering coastal features the complexity of problem is clearly shown with features being so varied. Future work looks to implement new models discussed in the paper and to expand the pipeline proposed.
Download Tulsi's Thesis