Reconstruction of NDVI Data with Normal Variance-Mean Mixture Noise using Stochastic Gradient Methods

Abstract: Satellite data is useful for making inference on large areas, but there are plenty of problems associated with the technology. The observation errors need to be carefully modelled in order to make inference from the data. In addition, the large data-sets involved makes straight-forward computations infeasible. This thesis explores a non-Gaussian noise structure, and implements a modified EM algorithm where the M-step is replaced for taking a step in a stochastic gradient descent method. The estimated parameters are used for reconstruction both of simulated data and of a data-set retrieved from the PAL database. The algorithm performs well on a simulated data equipped with the non-Gaussian noise structure; as for the PAL data, the algorithm performed quite well, but convergence proved to be somewhat unstable.