Spectral Unmixing and Mapping of Coral Reef Benthic Cover with Deep Learning

Abstract

Coral reefs are important to the biological ecosystem and the communities and wildlife that rely on the habitat they create. However, coral reefs are in critical and rapid decline: reefs have degraded over recent decades and what remains is at increasing risk of loss. Despite this, only a small fraction of the world’s reefs have been studied quantitatively, when mapping these reefs would allow us to monitor and understand their health. We seek methods to better model reefs at global scales. In this paper, we train with high spatial resolution satellite imagery and use low-spatial, high-spectral resolution priors from labeled imagery to demonstrate better accuracy in generalizing to unseen examples than prior work. We also develop on a novel method using a Deep Conditional Dirichlet Model to perform unmixing of spectral data across different resolutions. Accounting for mixing yields improvements on the accuracy of the predicted class probabilities. We map coral benthic cover with this method and intend to generalize and apply it at global scales.