Studies on the relationship between the Normalized Difference. Vegetation Index ... changes in the Kruger National Park, South Africa. The interannual and ..... Reisinger, A. (Eds.), IPCC, Geneva, Switzerland, p.p. 104. A.K. Knapp, P.A. Fay, ...
Spatial and temporal heterogeneity of phenology patterns in Kruger National Park, South Africa: different drivers for different areas. F. Parrini*, B.F.N. Erasmus School of Animal, Plant and Environmental Sciences, University of the Witwatersrand, Johannesburg, South Africa – (francesca.parrini, barend.erasmus)@wits.ac.za Abstract – In this paper we explore the spatio-temporal variation of vegetation greenness and the underlying causal processes in a southern African savanna region. From the AVHRR NDVI time series we derived the following phenological metrics for the period 1985-1999: start, length and end of wet season, NDVImax and seasonal ΣNDVI. We used a spatio-temporal clustering technique to find the most probable phenological clusters in space and time. All phenological metrics had high temporal and spatial variability, and we tested how well temperature, rainfall, soil and tree cover explained the observed patterns. Phenology metrics models which included tree cover and/or soil sand % together with rainfall received the best support. This highlights the importance of developing a finer scale spatially explicit rainfall–vegetation model in order to infer how biodiversity will be affected by climate change. Keywords: biodiversity, phenology, NDVI, savanna, spatiotemporal clustering, rainfall 1. INTRODUCTION Savannas are highly heterogeneous event-driven ecosystems. An analysis of the spatial and temporal variation of vegetation greenness and their causal processes is of primary importance to assess functional biodiversity responses to reported and expected changes in climate (IPCC, 2007). The amount of long-term annual rainfall, the seasonal distribution of rainfall and the intensity and frequency of rainfall events are predicted to change in the course of this century (Richard et al., 2001). These variables have an effect on primary production (Knapp et al., 2002) mainly through their control over water balance and growing-season length, and these effects are particularly evident in arid and semi-arid ecosystems (Vanacker et al., 2005).
(Pettorelli et al., 2005). The spatial configuration of such vegetation resources determines the cost of resource acquisition for any particular animal, and for humans. Therefore an understanding of the nature of this spatio-temporal configuration will inform population ecologists on the likely costs of resource acquisition, and subsequently, potential impact on population viability. Climate change studies typically focus on the consequences of variations in temperature or rainfall over time either at specific locations or averaged over regions (e.g. Goward and Prince, 1995; Los et al., 2001), with little consideration for changes in the spatial configuration of these variables. Explaining observed patterns of vegetation greenness as a function of climate’ opens the way to model the impact of predicted changes in the spatio-temporal configuration of vegetation greenness to climate change. In this study, a suite of 5 metrics, derived from 14 years of timeseries NDVI data, were calculated to describe phenological changes in the Kruger National Park, South Africa. The interannual and spatial variability of these metrics was assessed to provide information on spatial and temporal heterogeneity in this savanna region. In the highly seasonal environments of southern Africa, water balance and spatial and temporal rainfall distribution are the most important factors related to changes in plant phenology (Vanacker et al., 2005), therefore the influence of annual climatic fluctuations on the phenology dynamics was also assessed. The goal was to investigate the influence of climatic variables, vegetation cover and soil characteristics on the NDVI metrics in order to assess their potential for monitoring biodiversity changes. We discuss how a better knowledge of the variability in spatially explicit measures of water availability could improve our understanding of the functioning of the ecosystems under changing climate conditions. 2. METHODS
A delay of just few weeks in the production of new leaves can make a difference to the survival and reproductive success of the herbivores that depend on them (Owen-Smith and Cooper, 1989) and on the long term could have substantial consequences for terrestrial productivity and carbon dynamics (IPCC, 2007). The timing of late-summer rainfall, and the subsequent period for which green forage is available during the dry austral winter, has been shown to affect herbivore population dynamics in a semiarid savanna (Ogutu and Owen-Smith, 2003). Studies on the relationship between the Normalized Difference Vegetation Index (NDVI), rainfall and soil moisture availability have lead to a better understanding of the environmental constraints on vegetation growth. The NDVI is based on the differential reflection of green vegetation in the visible and infrared portion of the spectrum (Tucker, 1979), and provides information about the spatial and temporal distribution of vegetation greenness, and by proxy, - biomass and - quality * corresponding author
The Kruger National Park (KNP) is located in the eastern corner of the South African lowveld, adjoining the Mozambique border (22° 20’S - 25° 32’S and 30° 47’E and 32° 00’ E) and covers approximately 20 000 km². Its elevation ranges between 260 m and 839 m. Long-term annual rainfall is approximately 513 mm in the north and 660 mm in the south (20 years record) and interannual variability is 42 % in the north and 40 % in the south (coefficient of variation). Granitic soils dominate the western region while the eastern sector is predominantly underlain by basalt and associated clayey soils (Du Toit et al., 2003). Vegetation is mainly woodland savanna (Venter and Gertenbach, 1986) with Acacia spp., Combretum spp., marula Sclerocarpa birrea caffra, and mopane Colophospermum mopane as the dominant trees. As a measure of greenness, we used the NDVI 10 days composite data set acquired by the National Oceanic and Atmospheric Administration (NOAA) Very High Resolution Radiometer (AVHRR) from 1984 to 1999 (1994 excluded due to satellite
failure), aggregated to 5 x 5 km blocks. Clouds and poor atmospheric conditions result in low NDVI values, and therefore sudden drops in NDVI can be regarded as noise and removed. We applied a line smoothing algorithm to the NDVI time series, following Chen et al. (2000). For each of the 36 10-days time periods (t) in one year, we calculated the midpoint NDVI value between the previous period (t-1) and the following period (t+1). Then we compared this midpoint value with the actual NDVI value at time t. If the calculated midpoint was greater than the original value, the original was replaced by the midpoint value as the NDVI for that 10-day period; if the calculated midpoint value was smaller than the original NDVI value, the original was retained. This technique was effective at preserving the essence of the NDVI time series, while eliminating much of the contaminated data. From the interannual variations in the NDVI time series it is possible to determine phenological metrics describing the growth cycle of vegetation. We derived 3 temporal metrics and 2 greenness metrics (adapted from Reed et al., 1994): the start of the growing season, the growing season length, the end of the growing season; and the gross primary production over the growing season and the peak of photosynthetic activity. We used the phenological detection algorithm described by White et al. (2002) to define the phenological metrics. First, for every pixel, we selected the annual cloud free minimum and maximum and calculated the midpoint between them. This was repeated for every year and averaged across years (NDVIhalfmax). We used NDVIhalfmax as a threshold to identify the start and the end of the growing season in each year. This ensures consistency across years in the definition of the date of onset, therefore allowing detection of changes between years. The start of the growing season (GSS) and the end of the growing season (GSE) is a date expressed in units of numbers of temporal composites from the start of the year i.e. one unit equals 10 days for this AVHRR dataset. The length of the growing season (GSL) was calculated as the number of compositing periods from the start to the end of the growing season. From the time series, we derived other two phenological indicators: the maximum NDVI value (NDVImax) was the maximum NDVI value in any growing season; the growing season production (iNDVI) was the sum of all the NDVI values in one growing season. We used scan statistics software (SaTScan - Kulldorff et al., 1998) to detect clusters of each of the 5 metrics in space and time. The scan statistics identifies clusters of measurements where the occurrence of the event is more likely within the cluster than outside the cluster. The scanning procedure is a 3D moving window approach, using a moving virtual cylinder to detect clusters within and across years. The procedure allows both the base and the height of the cylinder to vary continuously as scanning progresses through space and time, obtaining an infinite number of overlapping cylinders of different size and shape covering the entire study area, where each cylinder reflects a possible cluster. The cylinder or circle (if only two dimensions) corresponding to the maximum likelihood ratio, obtained by a Monte Carlo simulation that repeats the analysis for a large number of random replications of the original data set (Kulldorff, 1997), represents the most likely or primary cluster. Secondary clusters are identified by the ranks of the likelihood ratios for all cylinders. In our analysis the size of the base of the cylinder was permitted to vary from 0 to 50% of the study area, the height
from one year to the maximum number of years in the time series and 9999 replications were used in the Monte Carlo simulation. Data were analysed separately from 1984 to 1993 and from 1995 until 1999 because of the missing year in 1994. We used the default setting of not reporting secondary clusters that overlap the most likely cluster as they provide little additional information. For the period 1985-1999, we obtained monthly precipitation totals and monthly averages of mean, minimum and maximum temperature from the South African Weather Service for 30 meteorological stations distributed throughout the park. For each station we calculated 1) total wet season rainfall in year t as the sum of precipitation from the start of the growing season in year t-1 (as identified from the NDVI time series) to the end of the growing season in year t; 2) wet season rainfall up to the month of photosynthetic peak as the sum of the precipitation from the start to the growing season in year t-1 to the month of highest NDVI value in year t; 3) wet season rainfall up to the month prior to the month of photosynthetic peak as the sum of the precipitation from the start to the growing season in year t-1 to the month before the month of highest NDVI value in year t; 4) the spread of the rainfall across the wet season months as the evenness (Shannon’s index H - Bronikowski and Webb, 1996) of rainfall during the growing season months; and 5) dry season rainfall in year t as the sum of precipitation from the start of the dry season in year t to the end of the dry season in year t. In the KNP rainfall is spatially highly variable due to its convective nature (see http://metsys.weathersa.co.za/SZ.html for snapshot of rainfall events in real-time). High spatial variability in rainfall reduces effective use of interpolation measures (Veenendal et al., 1996). Therefore, we examined the response of the NDVI derived metrics to rainfall and temperature variables only for NDVI pixels that contain a weather station. For these pixels, we extracted NDVI metrics and monthly climate data for the 14 years of the study period. Soil background colour and the ratio of woody to herbaceous vegetation are both spatially variable and further restrict the reliable use of NDVI metrics to make direct spatial comparisons. Hence, for each phenological metric we performed a time series regression analysis with the pixels containing a weather station and soil, tree characteristics and weather variables as the explanatory variables. We used soil sand % (Batjes, 2004) as a proxy measure for soil water retention: sandy soils will drain quicker, with less moisture available to plants. For a measure of tree cover, we used the % tree cover data from the MODIS-VCF global database. It overestimates tree cover but it’s the best approximation we have at this point in time for the KNP. We used the adjusted Aikake’s Information Criterion (AICc) to choose the best time series regression models. The larger the AIC distance between any two models, the less relatively plausible is the model with the higher value (smaller AIC values are more parsimonious). Models were sorted by AICc value and assigned weights (wi) representing the probability that the ith model was the best model in the candidate group. Models having w within 90% of the best model defined our confidence set (Burnham & Anderson, 2002). 3. RESULTS The results from the scan statistics confirm the high spatial and temporal heterogeneity of the KNP. Table 1 shows that only 5 of the clusters encompass more than one year suggesting that any
Table 1. Summary of spatio-temporal clusters of NDVI metrics in KNP from 1985 to 1993, and from 1995 to 1999 as indicated in the column headings. All clusters were significant at p