Vegetation cover degradation assessment in Madagascar savanna based on trend analysis of MODIS NDVI time series

Abstract

Like other African countries, Madagascar is concerned by vegetation cover degradation especially in savanna ecosystems. In this article, we describe an approach to quantify and localise savanna vegetation cover degradation. To this end, we analyse using STL decomposition method the trends measured between 2000 and 2007 of two phenological indicators which are derived from NDVI MODIS time series and characterizing vegetation activity during the growing season. Three types of trend were observed – null, positive or negative – over the study period with which we can associate a state of vegetation cover degradation. Future work will provide validation of this result. Next a comparison between the spatial variations of vegetation cover degradation and fire pressure for the same period should improve knowledge on the effect of fire on savanna vegetation activity. This information will be useful for local managers in order to implement savanna management strategies.

Introduction

Vegetation cover degradation in African savanna ecosystem is leading to the desertification of some areas, with the disappearance of the vegetation strata (regressive vegetation dynamics), or to an encroachment and a landscape closure, with the development of shrub-tree strata (progressive vegetation dynamics). In both cases, the equilibrium that maintains the savanna ecosystem is modified (Anyamba and Tucker, 2005, Olsson et al., 2005).

In African countries, savanna is an essential ecosystem for the local population because of its agricultural, environmental and economic importance. Due to recent increases in the degradation of the vegetation cover in this ecosystem, there is a pressing need for a quantitative and reproducible assessment of the phenomenon to support policy development for food security and resource conservation (Harrison and Pearce, 2000).

Changes in ecosystems can be classified in three groups, seasonal, gradual and abrupt change (Verbesselt et al., 2009). The phenomenon of vegetation cover degradation results from a modification in the level of vegetation activity over time that can be measured through the study of trends in the vegetation index or phenological indicators (Reed et al., 2003). Vegetation cover degradation belongs to the class of gradual change, which is why, in this article, we assume that multi-year trends are smooth and change slowly; they can be modelled by a 1st degree polynomial.

Two requirements were defined to select the method for analysing image time series:

• To be able to identify the specific phenology cycle of the local savanna, from which phenological indicators are derived.

• To allow the characterisation of gradual change in savanna ecosystems by deriving the direction of change within the trend of the time series and without the need to define the change trajectory or to identify breakpoints.

Among recent methods proposed for analysing image time series, some were only developed to identify seasonal change such as functions fitted to time series data (Jönsson and Eklundh, 2002), or abrupt change by discriminating noise from the signal by its temporal characteristics (Millward et al., 2006), or seasonal and gradual changes using the breaks for additive seasonal and trend (BFAST) method (Verbesselt et al., 2009). In this paper, we chose seasonal decomposition of time series by Loess (STL) because it provides an accurate and robust estimation of trend and seasonal components thanks to its capacity to deal with outliers or missing values within the time series. The Loess method was also used to decompose the 16,886 time series (composed of 178 values) to cover the savanna study area because, according to Johnson et al. (2008) and Lu et al. (2001), STL method proved to be flexible, computationally efficient and simple.

STL is a non-parametric method which, using an additive model (Eq. (1)), flexibly decomposes time series data into three separate components – trend (Z_t), seasonal (S_t) and remainder (ɛ_t) – for each pixel. This type of model implies that the magnitude of fluctuations in the original series resulting from the seasonal pattern and the residual component is not affected by the level of the trend.

$$X_{\text{t}}=Z_{\text{t}}+S_{\text{t}}+{\epsilon }_{\text{t}}$$

STL is an iterative procedure that repeatedly uses different types of LOcally wEighted regreSion Smoother (LOESS) (Cleveland et al., 1990). To evaluate the Loess fit g(x) at a given x, all data points in the neighbourhood of x are assigned tricubic weights, so that the closer a point is to x, the larger its weight. Weighted least squares are then used to fit a polynomial though the points and g(x) equals the value of the polynomial at x. The parameters to be defined are the size of the neighbourhood and the degree of the polynomial (constant, linear or quadratic).

The seasonal component provides the phenology cycle of the local vegetation for the study period while the trend component, modelled by a piecewise linear function, enables determination of the direction of change during the study period by analysing the slope sign of the trend.

Satellite remote sensing has long been a source of data to detect and monitor vegetation cover dynamics over time (Coppin et al., 2004). Coarse and medium spatial resolution satellite images are well-suited for this task. They provide consistent, valuable and repeatable measurements at a spatial scale which is appropriate for detecting the effects of many processes that cause degradation of the vegetation cover.

Changes in vegetation cover can be measured by remote sensing of the normalised difference vegetation index (NDVI) as a strong correlation has been established between NDVI and vegetation cover. Time series of MODIS NDVI data have been successfully applied to quantify vegetation activity and to measure vegetation dynamics (Ahl et al., 2006, Zhang et al., 2003).

The purpose of this paper is to describe and compare two approaches to quantify vegetation cover degradation in savanna ecosystems using remote sensing time series NDVI data. The first approach consists in measuring the trend of a phenological indicator time series characterising vegetation cover degradation over time (Bai et al., 2005, Coppin et al., 2004). Characteristics of the phenological indicator are derived from the analysis of the NDVI profile given by the STL seasonal component. Then the phenological indicator is calculated for each growing season from the NDVI time series. The second approach is based on the measurement of the trends of NDVI time series using the STL trend component. This change detection technique enables gradual change to be isolated from seasonal and abrupt changes (Eklundh and Olsson, 2003). In both approaches, NDVI time series and phenological indicators represent an integrated measure of vegetation activity. A deviance of its values from a local reference value is assumed to be a measure of vegetation cover degradation.