Noise Modelling and Outlier Detection in Satellite Scatterometer Data

Abstract

This thesis describes a project in two parts. The project is an application of data modelling techniques to satellite scatterometer data. Scatterometer data is computed from radiation returned to a satellite as a result of its emitted radar beam interacting with the surface of oceans. The frequency of the beam is fixed to react to ripples on the water surface which are a result of the surface wind conditions, therefore the amount of power reflected back to the satellite is dependent on the surface wind vectors. The satellite has three radar beams so the scatterometer data is 3-dimensional and is believed to lie close to a cone-like manifold. The two parts of the project are: 1, The application of a Generative Topographic Mapping (GTM) to the data in order to create a manifold in data space which can be used identify outliers for the purposes of exclusion from further processing. 2. The modelling of the noise variance in order to see if there is any variation depending on the position within the data. The noise is modelled using an iterative technique involving 2 interacting neural networks, one modelling the mean of the data and one modelling the noise.