Linear trends in seasonal vegetation time series and the modifiable temporal unit problem

R De Jong,S De Bruin

doi:10.5194/bg-9-71-2012

Abstract

Abstract. Time series of vegetation indices (VI) derived from satellite imagery provide a consistent monitoring system for terrestrial plant productivity. They enable detection and quantification of gradual changes within the time frame covered, which are of crucial importance in global change studies, for example. However, VI time series typically contain a strong seasonal signal which complicates change detection. Commonly, trends are quantified using linear regression methods, while the effect of serial autocorrelation is remediated by temporal aggregation over bins having a fixed width. Aggregating the data in this way produces temporal units which are modifiable. Analogous to the well-known Modifiable Area Unit Problem (MAUP), the way in which these temporal units are defined may influence the fitted model parameters and therefore the amount of change detected. This paper illustrates the effect of this Modifiable Temporal Unit Problem (MTUP) on a synthetic data set and a real VI data set. Large variation in detected changes was found for aggregation over bins that mismatched full lengths of vegetative cycles, which demonstrates that aperiodicity in the data may influence model results. Using 26 yr of VI data and aggregation over full-length periods, deviations in VI gains of less than 1% were found for annual periods (with respect to seasonally adjusted data), while deviations increased up to 24% for aggregation windows of 5 yr. This demonstrates that temporal aggregation needs to be carried out with care in order to avoid spurious model results.

Highlights

Vegetation systems provide a quick and measurable response to many environmental changes at a wide range of spatial and temporal scales
The aim of this paper is to demonstrate possible Modifiable Temporal Unit Problem (MTUP) effects in analysis of time series of satellite imagery using both real and simulated vegetation index (VI) data and to provide, in this sense, a framework for linear time series regression
The risk of artefacts is minimal at an aggregation level corresponding to a full period, for instance a calendar year

Summary

Introduction

Vegetation systems provide a quick and measurable response to many environmental changes at a wide range of spatial and temporal scales. The availability of historical time series from satellite observations with daily global coverage makes operational monitoring of vegetation condition a matter of detecting and interpreting changes within these datasets. Change detection is often complicated by a number of statistical preconditions that are intrinsic to time series of spectral vegetation indices with dense sampling intervals. In few cases linear models were fitted directly to seasonal data Pelkey et al, 2000), but seasonality is typically remediated using temporal aggregation, where the aggregation window (or bin size) corresponds to the length of a calendar year. The resulting bins can be regarded as temporal units, which, like spatial units, are modifiable (Taylor, 2010). In case of spatial units, it has been demonstrated that the size may influence

Objectives

Results

Conclusion