Evaluating the performance of multiple remote sensing indices to predict the spatial variability of ecosystem structure and functioning in Patagonian steppes

Abstract

Assessing the spatial variability of ecosystem structure and functioning is an important step towards developing monitoring systems to detect changes in ecosystem attributes that could be linked to desertification processes in drylands. Methods based on ground-collected soil and plant indicators are being increasingly used for this aim, but they have limitations regarding the extent of the area that can be measured using them. Approaches based on remote sensing data can successfully assess large areas, but it is largely unknown how the different indices that can be derived from such data relate to ground-based indicators of ecosystem health. We tested whether we can predict ecosystem structure and functioning, as measured with a field methodology based on indicators of ecosystem functioning (the landscape function analysis, LFA), over a large area using spectral vegetation indices (VIs), and evaluated which VIs are the best predictors of these ecosystem attributes. For doing this, we assessed the relationship between vegetation attributes (cover and species richness), LFA indices (stability, infiltration and nutrient cycling) and nine VIs obtained from satellite images of the MODIS sensor in 194 sites located across the Patagonian steppe. We found that NDVI was the VI best predictor of ecosystem attributes. This VI showed a significant positive linear relationship with both vegetation basal cover (R² = 0.39) and plant species richness (R² = 0.31). NDVI was also significantly and linearly related to the infiltration and nutrient cycling indices (R² = 0.36 and 0.49, respectively), but the relationship with the stability index was weak (R² = 0.13). Our results indicate that VIs obtained from MODIS, and NDVI in particular, are a suitable tool for estimate the spatial variability of functional and structural ecosystem attributes in the Patagonian steppe at the regional scale.

Introduction

Drylands cover about 41% of Earth's land surface, and are home to more than 38% of the total global population (Millennium Ecosystem Assessment, 2005). Because of climatic restrictions, only 25% of the world's drylands are devoted to agriculture, but they are of paramount importance for grazing, as 65% of the drylands are used for grazing of managed livestock on native vegetation (Millennium Ecosystem Assessment, 2005). These areas also support 78% of the global grazing area (Asner et al., 2004), and over 50% of the world's livestock (Puigdefábregas, 1998).

The establishment and adjustment of land management practices in drylands requires routine monitoring of land functionality (Pyke et al., 2002). This is particularly important for areas that are subject to uses that can promote desertification, such as grazing (Asner et al., 2004). Measuring ecosystem functionality in situ requires assessing variables such as the retention of water and nutrients on landscapes (Valentin et al., 1999), the plant productivity (McNaughton et al., 1989) and soil properties related to nutrient cycling (Maestre et al., 2012). These measurements are very time-consuming and costly, and require technical equipment and expertise that may not be always available, particularly in developing countries. Therefore, methods based on easy-to-measure indicators are being increasingly used when monitoring drylands (de Soyza et al., 1997, Herrick et al., 2002, Pyke et al., 2002). A number of methodologies have been developed in the last decades for this aim, which are based on measures of structural attributes of vegetation and soil surface characteristics related to ecosystem functioning (National Research Council, 1994, Herrick et al., 2005, Tongway and Hindley, 2004). One of these methods that have attracted most attention to date is the landscape function analysis (LFA) methodology, developed in Australia by David Tongway and co-workers (Tongway, 1995, Tongway and Hindley, 2004). The LFA uses easily observable vegetation structure attributes and soil surface indicators to assess ecosystem functionality. These indicators are combined in three indices (stability, infiltration and nutrient cycling), which assess the degree to which resources tend to be retained, used and cycled within the system. Several studies have shown significant relationships between the LFA indices and quantitative measurements of these functions in multiple ecosystems and countries, including Australia (Holm et al., 2002), Iran (Ata Rezaei et al., 2006), South Africa (Parker et al., 2009), Spain (Maestre and Puche, 2009, Mayor and Bautista, 2012), and Tunisia (Derbel et al., 2009). The LFA methodology has been selected to develop the MARAS system (Spanish acronym for “Environmental Monitoring for Arid and Semi-Arid Regions”), a large-scale network of long-term monitoring sites across Patagonia (Argentina) aiming to detect early changes in ecosystem structure and function that could indicate the onset of desertification processes (Oliva et al., 2011). The first MARAS permanent sites were set up in 2008, and until now about 200 MARAS have been established. The effort and time required to collect field data for the MARAS system is costly, and this limits the number of sites that can be routinely measured.

Scaling up or extrapolating measurements from small plots to larger, more representative landscapes is an important objective of the MARAS system, as well as of similar initiatives such as Western Australian Rangelands Monitoring System (Pringle et al., 2006) or Land Degradation Assessment in Drylands (Nachtergaele and Licona-Manzur, 2009). Remote sensing tools are extremely important to achieve this objective (Ludwig et al., 2007, Reynolds et al., 2007). Field-based surveys facilitate the interpretation and extrapolation of satellite images by providing data to calibrate empirical models relating ecosystem functionality with remote sensing data (Wessman, 1994).

Vegetation indices (VIs), based on satellite observations, are mathematical transformations of reflectance measurements in different spectral bands, especially the visible (usually red) and near-infrared bands, that are widely used to obtain information about land surface characteristics (Jackson and Huete, 1991). Over the years, a great number of VIs of varying complexity have been proposed, each with advantages and limitations (Bannari et al., 1995). The most commonly used VI is the Normalized Difference Vegetation Index (NDVI, Rouse et al., 1973). Different proportions between vegetation cover and background soil may affect the relationship between NDVI and vegetation attributes in sparsely vegetated areas such as drylands (Huete and Jackson, 1988). NDVI is also sensitive to attenuation and scattering by atmospheric gases and aerosol particles (Carlson and Ripley, 1997). Thus, several alternative VIs have been developed to account for factors such as the background soil (e.g. the Soil-Adjusted Vegetation Index – SAVI –, Huete, 1988), or the atmosphere (e.g. the Atmospherically Resistant Vegetation Index – ARVI –, Kaufman and Tanre, 1992). A wide range of satellite sensors has been used to construct VIs. The Moderate Resolution Imaging Spectroradiometer (MODIS) sensor represents a suitable compromise between spatial and temporal resolution, as it provides free-cost products with atmospherically corrected and georeferenced surface reflectances at spatial resolutions down to 250 m, and with temporal frequencies ranging from 1 to 16 days (Justice et al., 1998). Thus, the use of MODIS images to calibrate field-obtained indicators of ecosystem structure and functioning is highly attractive, particularly when economical and/or technical constraints preclude the use of images with higher spatial resolution.

Recent studies have shown that NDVI can satisfactorily predict LFA indices in restored mines in Australia (Ong et al., 2009) and semi-arid grasslands in Spain (García-Gómez and Maestre, 2011). However, and to the best of our knowledge, no previous study has attempted to evaluate the ability of VIs other than NDVI to predict LFA indices, or other surrogates of ecosystem functionality. We aimed to do so by evaluating the relationships between LFA indices, key features of perennial vegetation (basal cover and richness) and several VIs obtained from the sensor MODIS over a large area (800,000 km²) in the Patagonian steppe. The objectives of this study were to: (i) test whether we can predict the spatial variability in ecosystem structure (species richness and plant cover) and functioning (LFA indices), over a large area using VIs obtained from MODIS data and (ii) evaluate which VIs are the best predictors of these ecosystem attributes.