Jonathan Seaquist

The aim of this study is to systematically compare parametric and non-parametric techniques for analyzing trends in annual NDVI derived from NOAA AVHRR sensor in order to examine how trend type and departure from normality assumptions affect the accuracy of detecting long-term change. To generate annual data, the mean NDVI of a four-month long ‘green’ season was computed for fifteen sites (located in Africa, Spain, Italy, Sweden, and Iraq) from the GIMMS product for the periods 1982-2006. Trends in these time series were then estimated by Ordinary Least-Squares (OLS) regression (parametric) and the combined Mann-Kendall test with Theil-Sen slope estimator (non-parametric), and compared using slope value and statistical significance measures. We also estimated optimal polynomial model for the annual NDVI, by using Akaike Information Criterion (AIC), to determine the trend type at each site.
Results indicate that slopes and their statistical significances obtained from the two approaches at sites with low degree polynomials (mostly linear) and steep monotonic (gradually increasing or decreasing) trends compare favourably with one another. At sites with weak linear slopes, the two approaches had similar results as well. Exceptions include sites with abrupt step-like changes resulting in departures from linearity and consequently high degree polynomials where the least-squares method outperformed the Mann-Kendall Theil-Sen method. In sum, we conclude that OLS is superior for detecting NDVI trends using annual data though further investigation using other techniques is recommended.