Deep Learning For Feature Tracking In Optically Complex Waters

Abstract

Environmental monitoring and early warning of water quality from space is now feasible at unprecedented spatial and temporal resolution following the latest generation of satellite sensors. The transformation of this data through classification into labelled, tracked event information is of critical importance to offer a searchable dataset. Advances in image recognition techniques through Deep Learning research have been successfully applied to satellite remote sensing data. Deep Learning approaches that leverage optical satellite data are now being developed for remotely sensed multi- and hyperspectral reflectance. The combination of spectral with spatial feature extracting Deep Learning networks promises a significant improvement in the accuracy of classifiers using remotely sensed data. This project aims to re-tool and optimise spectral-spatial Convolutional Neural Networks originally developed for land classification as a novel approach to identifying and labelling dynamic features in waterbodies, such as algal blooms and sediment plumes in high-resolution satellite sensors.