Aridity influence on vegetation patterns in the middle EbroValley (Spain): evaluation by means of AVHRR images and climate interpolation techniques
Sergio M. Vicente-Serrano1,2*, José M. Cuadrat-Prats3 and Alfredo Romo4
1 Instituto Pirenaico de Ecología, CSIC (Spanish Research Council), Campus de Aula Dei, P.O. Box 202, Zaragoza 50080, Spain
2Unit for Landscape Modelling, University of Cambridge, SirWilliamHardyBuilding, Tennis Court Road, Cambridge, CB2 1QB, UK
3Departamento de Geografía. Universidad de Zaragoza. C/ Pedro Cerbuna 12. 50009. Zaragoza. Spain.
4Laboratorio de Teledetección. Departamento de Física Aplicada I. Facultad de Ciencias. Universidad de Valladolid. 47071. Valladolid. Spain
* e-mail:
ABSTRACT.This paper analyses the role of climatic aridity on the spatial differences of vegetation activity and its inter-annual variability in a semiarid region of the North East of the Iberian Peninsula. The vegetation activity was quantified by means of a monthly NDVI database from NOAA-AVHRR satellite images (1987-2000) at 1 km2 of spatial resolution. Coefficients of variation (CoV) from temporal NDVI series were also calculated. The greater temporal variability of NDVI was recorded in areas with low vegetation cover (steppe and dry farming lands), whereas the lowest temporal variability of NDVI was recorded in the irrigated lands and forests located in the most humid areas. There is a strong and negative relationship between the CoV and the NDVI distribution, but we recorded important differences among land uses. Using maps of annual precipitation and temperature obtained by means of a regression-based method, we studied the influence of aridity on spatial patterns of NDVI and CoV values. The results show the high influence of aridity on the spatial distribution of vegetation activity in the MiddleEbroValley. Aridity causes a general decrease of NDVI and an increase of CoV. Nevertheless, non-linear relationships between aridity and NDVI and CoV were recorded for the whole of the study area and for the different land uses. The relationships between aridity and vegetation patterns in the semiarid region studied are discussed in depth.
KEY WORDS. Remote Sensing, NDVI, Temporal variability of vegetation, regression-based interpolation, land cover types, Mediterranean region.
- Introduction
Arid and semiarid regions of the world are highly sensitive to human-induced climate and/or land transformation (Nicholson and Farrar 1994; Nicholson et al. 1998; Evans and Geerken 2004). In these regions, lack of water is the main constraint for vegetation development (Hadley and Szarek 1981; Le Houreou 1984). In the semiarid areas of the Mediterranean region there is high ecological uncertainty related to the consequences of possible climatic change during the twenty-first century. Some climatic models show an important decrease in water availability in this area (Houghton et al. 2001; Jones et al. 1996), which could cause dramatic consequences for natural ecosystems and dry farming. The semiarid regions of southern Europe are ecotones where the impacts of climate change can be identified early (Lavorel et al. 1998). Knowing the influence of climate on vegetation in these regions is very important to determine the possible consequences of climatic change on vegetation characteristics.
Remote sensing is an useful tool to analyse the vegetation dynamic on local (Ringrose and Matheson 1991), regional (Nicholson et al. 1990), or global scales (Lucht et al. 2002; Kawabata et al. 2001; Kogan 2001) and to determine the impact of climate on vegetation (Wang et al. 2001 and 2003). Several studies have shown that interannual differences in vegetation parameters are mainly driven by water availability (Farrar et el. 1994; Santos and Negrin 1997; Weiss et al. 2004; Sannier and Taylor 1998; Maseli et al. 1992; Groten 1993; Al-Bakri and Taylor 2003). Ichii et al. (2002) analysed the relationships between precipitation and vegetation activity, quantified by means of remote sensing, on a global scale and found the highest correlation in the arid and semiarid regions of the world in South Africa, Australia and Central Asia. Moreover, other studies have shown by means of remote sensing that interannual variability of vegetation activity in semiarid regions is higher than in humid and sub-humid regions (e.g., Wang et al. 2003; Tucker et al. 1991). Also deserts and herbaceous covers were found to have the highest variability (Fang et al. 2001, for China).
Nevertheless, in the semiarid regions of southern Europe there are few studies on the relationships between climatic conditions and vegetation cover quantified by means of remote sensing (e.g., González-Alonso et al. 1995 and 2001; Ascaso 1997).
This paper analyses the spatial patterns of vegetation activity and its temporal variability in the MiddleEbroValley, a semiarid Mediterranean region located in the North East of the Iberian Peninsula, where the transformation of vegetation cover due to human activity has been intense (Frutos, 1976). In the MiddleEbroValley, there are agricultural areas with an important economic, social, and environmental interest (Martí 1992; Austin et al 1998). Also, there are some natural areas modified by humans (steppes) with great landscape diversity and biological richness (Suárez-Cardona et al. 1992; Pedrocchi 1998) and forests located in the boundaries of their climatic limitations (Braun-Blanquet and Bolós, 1957).
Understanding the influence of climate on the spatial distribution of vegetation cover and vegetation activity is very important, mainly for the areas where climate change is predicted during the Twenty-first Century. These studies may help determine possible structural and functional modifications of vegetation, corresponding to the present climate models. The study presented here has been done in a region highly humanised and prone to soil degradation (Lasanta, 2003), that may be affected by some of the problems related to global change.
2. Study area
The MiddleEbroValley is one of the most arid regions of the Iberian Peninsula (Figure 1). The vegetation activity in this region is highly determined by the water availability (Braun-Blanquet and Bolos 1957; Austin et al. 1998). The vegetation cover is low due to aridity, poor soils and frequent droughts (Guerrero et al. 1999; Vicente-Serrano, 2004). Lithology is characterised by millstones and gypsums (Peña et al. 2002) that contribute to aridity because soils poorly retain the water (Navas and Machin 1998). Relief determines the climate since the Valley is surrounded by mountainous chains that cause continental climatic characteristics and important dry conditions (Cuadrat 1999). In the centre of the Valley the annual precipitation is less than 350 mm. Moreover, throughout the study area there is a negative water balance (precipitation – evapotranspiration), which is very important in the central areas (> 900 mm). Temperature has an important seasonality. Frost periods are frequent in winter, which limit, noticeably, the vegetation development. In the centre of the valley the winter average temperature is 10º, but minimum temperatures below –10 ºC are frequent. In summer, the average temperature is 24 ºC, although maximum temperatures above 40 ºC are also frequent. Seasonal differences of precipitation are less important, although the dry season is usually recorded during summer months.
The landscape has for centuries been highly determined by human land use, which has substantially altered the natural vegetation (Frutos 1976; Pinilla 1995). During the second half of the twentieth century, the transformations affected large areas due to the creation of irrigated lands (Frutos 1982) and land abandonment caused by the extensification policies of the European Union (Errea and Lasanta 1993). At present, the dominant land uses are steppes (mosaic of shrubs and pastures) (25.7%), dry farming areas with herbaceous cultivations (21.4%), and the mosaic of both these previous land uses (13.6%). Coniferous forests (mainly Pinus halepensis) cover 7.9% of the total surface of the study area. Moreover, an important percentage of the study area has been transformed to introduce irrigation (20.8%) (Figure 1).
3. Methods
3.1. Remote sensing capability to ecosystems monitoring
We used remote sensing to analyse the influence of climate on vegetation activity and its interannual variability. The use of remote sensing to analyse vegetation activity is based on the spectral properties of vegetation, which absorbs an important percentage of electromagnetic energy in the visible region and reflects a high percentage in the near-infrared region (Knipling 1970). Due to these properties, different vegetation indices based on combinations of red and infrared bands can be calculated. The most widely used is the Normalized Difference Vegetation Index (NDVI) (Tucker 1979). High NDVI values indicate high vegetation activity due to the strong relationship of this index to radiation absorbed in photosynthetic processes (Gallo et al. 1985). But NDVI is also highly related to net primary production, vegetation cover and leaf area index (Sellers 1985).
The NDVI presents some problems for vegetation analysis due to the fact that relationships between NDVI and biomass or vegetation cover are often not linear (Choudhury et al. 1994; Gillies et al. 1997). Nicholson et al. (1990) indicated that the NDVI is a poor indicator of vegetation biomass in areas with high vegetation density. The soil background may also affects NDVI in semiarid regions because as the percentage of vegetation cover decreases the signal become increasingly contaminated by the soil reflectance (Huete and Jackson 1987). Nevertheless, in arid and semiarid areas in which the vegetation cover is not dense, several studies have shown the strong relationship between NDVI and vegetation biomass, vegetation cover and vegetation activity (i.e., Tucker et al. 1981; Gutman 1991; Diallo et al. 1991; Fuller 1998).
The NDVI is obtained from different satellites, but due to the elevated temporal frequency and the spatial resolution, the NOAA-AVHRR images are widely used in NDVI calculation (Gutman 1991; Gutman et al. 1995; Ehrlich et al. 1994). The AVHRR sensor registers radiometric data in the red and infrared spectral regions. The spatial resolution (1.1 km at nadir) reduces radiometric problems caused by soil (Jasinski 1990) or topography (Burgess et al. 1995).
The NDVI data used in this paper were created in Spain by the LATUV (Remote Sensing Laboratory of Valladolid University) at local area coverage resolution (LAC, 1 km2 of grid cell size). The images are received daily, calibrated (Kaufman and Holben 1993; NOAA 2003) and atmospherically corrected based on a modification of the 5S code (Tanré et al. 1990) using standard atmospheric measures for ozone absorption and molecular dispersion. For water vapour a method based on brightness temperatures of channels 4 and 5 is used (Illera et al. 1997). Later, the NDVI is obtained and geometrically corrected (Illera et al. 1996a). Residual atmospheric errors are reduced with the creation of monthly indices using the Maximum Value Composite method (MVC) (Holben 1986). The method is based on the selection of the maximum monthly NDVI from the daily images because the maximum values are recorded under optimum atmospheric and observation conditions: near Nadir, low aerosols and free of cloud (Gutman, 1989). The NDVI data set used in this paper has been extensively applied in the Iberian Peninsula for numerous purposes: identification of droughts (González-Alonso et al. 2000 and 2001), yield productivity and natural vegetation monitoring (Vázquez et al. 2001), or forest fire danger (Illera et al. 1996b).
We obtained one NDVI image per month from 1988 to 2000. For analysis, we selected the months between April and September, a period when vegetation is active in the different land-uses. We used the monthly NDVI data for each month, but also calculated annual values to summarise the vegetation activity between April and September. For this purpose we used integrated NDVI values during the season (ΣNDVI) because these values are highly related to net primary production (Tucker et al. 1981; Tucker and Sellers 1986; Prince and Tucker 1986). Integrated values were calculated following a trapezoidal approximation from monthly data (Justice and Hiernaux, 1986; Samson, 1993). Each annual value is obtained by means of the sum of the monthly NDVI values between April and September.
The average of monthly (NDVI) and annual (April to September) values (ΣNDVI) were obtained for each pixel of 1 km2. Moreover, coefficients of variation (CoV) at monthly and annual time scales were calculated to determine the temporal variability of NDVI following Weiss and Milich (1997).Coefficient of variation is a simple statistic calculated from the average and the standard deviation of the NDVI series in each pixel. This statistic has been widely used to determine the spatial differences of temporal variability in the vegetation activity of different arid and semiarid regions of the world (i.e., Tucker et al., 1991; Milich and Weiss, 1997; Fang et al. 2001).
The CORINE land cover map of the study area (CLC, 1989), reclassified to 12 categories, was used to determine the role of different land uses in spatial patterns of NDVI and CoV. For this purpose, we calculated the dominant land cover for each pixel of 1 km2, using a mode criterion (each pixel of 1 Km2 was assigned to the land cover type that represents the higher percentage of surface within the pixel). The classes used in this study were: 1: Dry farming areas (herbaceous), 2: Dry farming areas (permanent), 3: Dry farming areas (mosaic of herbaceous and permanent), 4: Irrigated lands, 5: Leafy forests, 6: Coniferous forests, 7: Steppes (shrubs and pastures) and 8: Mosaic of dry farming areas and steppes.
3.2. Continuous aridity mapping
To determine the role of aridity on the spatial patterns of NDVI and CoV we used precise climatic maps to take into account the climatic spatial differences caused by topographic and geographic factors. For this purpose we used the mean monthly precipitation from 159 and the mean monthly temperature from 72 weather stations. Random sampling was carried out on the original data set: seventy-percent of the data was used for interpolations and the remaining was reserved for subsequent test.
There are several methods to obtain continuous climatic surfaces over the territory (Isaaks and Strivastava, 1989; Goovaerts 1997; Borrough and McDonnell, 1998). In this paper we have used regression-based methods by means of independent geographic and topographic variables. Regression methods are based on the creation of dependence models between climatic data and geographic and topographic variables. The climatic data in a location where information is not available is predicted according to:
Where z is the predicted value in point (x), b0,...,bn are the regression coefficients and P1,...,Pn are the values of the different independent variables in point x. The relationship between climatic data and geographical and topographical variables has been widely analysed in the scientific literature (Basist et al. 1994, Daly et al. 2002). These relationships may be used to create empirical models that predict the climatic data from geographic and topographic variables. In this paper, to map precipitation and temperature we have used an approach similar to that of Agnew and Palutikof (2000), which mapped mean temperature and precipitation in the MediterraneanBasin. The independent variables used in this paper can be consulted in Vicente-Serrano et al. (2003). They were obtained from a digital elevation model of the study area at the same spatial resolution as the NDVI images (1 km2).
Results obtained from regression models are inexact. The predicted values z(xi) do not coincide with the climatic data obtained in the weather stations. There is a known error in the prediction (residual). To improve the climatic maps and to correct this known error we interpolated the residuals (Ninyerola et al. 2000; Agnew and Palutikof, 2000; Brown and Comrie, 2002) by means of splines with tension (Mitasova and Mitas, 1993). The sum of prediction and residual maps modifies the initial results of the model and real climatic values are obtained in the location of weather stations. To validate the models and to know the final quality of the maps we used different accuracy statistics (See Vicente-Serrano et al. 2003) by means of 30% of the weather stations that were retained to validate the models.
To obtain a general measure of aridity we calculated an Aridity Index (AI) based on the ratio between the mean precipitation and temperature maps on monthly and annual time scales by means of Geographic Information Systems (GIS). This index has been widely used for climatic classification (Köppen 1923; UNESCO 1979) and aridity quantification (De Martonne 1957; Korzun et al. 1976; Botzan et al. 1998). Low values of the index indicate higher aridity.
Relationships between the AI, the NDVI and the CoV were measured by means of a non-parametric correlation coefficient (Rho-Spearman) because it is more robust than parametric coefficients, it is not affected by outliers and its use is convenient when relationships between variables are not linear (Siegel and Castelan, 1988).
4. Results
4.1. Spatial distribution of aridity
The results of the regression models obtained for precipitation and temperature are shown in Table 1. In general, the models have a great quality, with high percentages of variance explained and low errors in the predictions in relation to the range of spatial variability of precipitation and temperature over the territory.
Table 2 shows the results of validation. In general the models have a high quality. The mean absolute error is less than 6 mm for monthly precipitation maps and less than 0.8 ºC for maps of monthly mean temperature. Annual maps (April to September) also show a high quality. The agreement index (Willmott’s D), which scales with the magnitude of the variables retains mean information and does not amplify outliers (Willmott, 1981), also shows values higher than 0.86 in all cases, but in general the values are higher than 0.90. This indicates a high agreement between measured and predicted data in the weather stations reserved for validation.
Figure 2 shows the spatial distribution of the averaged Aridity Index (AI) between April and September. Important spatial differences are observed, mainly related to the relief distribution, which determines the spatial distribution of precipitation and temperature. Aridity is more intense in the centre and east of the study area, where low precipitation and high temperatures are recorded. In the North, close to the Pyrenean range, the most humid conditions are shown as a consequence of high precipitation and low temperatures. In the southeast, close to the Mediterranean Sea and also corresponding to a Mountainous area, the conditions are also humid.
4.2. Spatial patterns of NDVI
Vegetation activity has an important seasonality. The highest vegetation activity is recorded in the months of April and May. In spring the highest NDVI values are recorded in the north of the study area, corresponding to forests, steppes and dry farming areas. In June, as a consequence of the start of the summer dryness, vegetation activity decreases and the mean NDVI decreases dramatically in dry farming areas and in steppes. However, spatial diversity is very important and it is caused by the presence of irrigated lands and some forested areas.