Trend detection and analysis in Eastern Europe and European Russia

Wednesday, 17 December 2014
Kirsten de Beurs, University of Oklahoma, Geography and Environmental Sustainability, Norman, OK, United States, Geoffrey M Henebry, South Dakota State University, Brookings, SD, United States and Braden Owsley, Texas State University San Marcos, San Marcos, TX, United States
A confluence of computing power, cost of storage, ease of access to data, and ease of product delivery make it possible to harness the power of multiple remote sensing data streams to monitor land surface dynamics. Change detection has always been a fundamental remote sensing task, and there are myriad ways to perceive differences. From a statistical viewpoint, image time series of the vegetated land surface are complicated data to analyze. The time series are often seasonal and have high temporal autocorrelation. These characteristics result in the failure of the data to meet the assumption of most standard parametric statistical tests. Failure of statistical assumptions is not trivial and the use of inappropriate statistical methods may lead to the detection of spurious trends, while any actual trends and/or step changes might be overlooked. Methods for the analysis of messy satellite data, which are often influenced by discontinuity, missing observations, non-linearity, and seasonality, are still being developed within the remote sensing community. We have found several examples of research that compares trends from different datasets. However, there is a dearth of information on the comparison of trend detection methods themselves for standardized datasets.

Here we describe three different trend detection methods, and compare their results for a set of synthetic time series exhibiting monotonic trends as well as step changes. We will vary the length of the time series, the number of observations per year and the number of missing values. We will also vary the seasonality and the strength of the autocorrelation. We will then discuss a case study for Eastern Europe and European Russia where we investigate time series of MODIS Nadir BRDF-adjusted (NBAR) data at 8-day and 500m resolution between 2001 and 2013. We investigate basic vegetation indices such as NDVI and EVI but also extend the analysis towards a disturbance index which identifies how pixels differ from comparable regions.