Spatio-Temporal Smoothing for Very Large Data Sets: Satellite-Based Vegetation Measurements

• Activity Number: 89
• Type: Invited
• Date/Time: Sunday, August 9, 2015 : 8:30 PM to 9:15 PM
• Sponsor: Korean International Statistical Society
• Author(s): Johan Lindström*
• Companies: Lund University
• Keywords: Spatio-temporal ; Gaussian Markov random field ; Splines ; Satellite data ; iterative methods ; large data

Abstract

Satellite measurements provide large scale monitoring of several important ecological factors, including vegetation cover. However, satellite data is often noisy with considerable amounts of missing or biased data due to atmospheric disturbances. To characterise vegetation over large areas, ecologists are interested in smoothed complete time-series of vegetation measurements. Given these time-series important information regarding plant phenology and plant productivity can be extracted.

We present the post-processing of satellite data as a spatio-temporal smoothing problem with non-Gaussian heavy tailed errors. Using the link between Wahba-splines and Gaussian Markov random fields (GMRFs) numerically efficient algorithms can be construced. The resulting algorithms replace costly numerical operations such as matrix determinants and inverses with stochastic approximations of trace and diagonal elements. These approximations are computed using iterative methods; the gridded structure of the satellite data allows for efficient preconditioners to be used. As an example the suggested methods are applied to satellite based vegetation measurements over the African Sahel.