The Normalized Difference Vegetation Index (NDVI) data for this study was derived from the Global Inventory Modeling and Mapping Studies (GIMMS) dataset. It is highly reliable NASA dataset, with the spatial resolution of 8 km (Tucker et al., 2005). NDVI data, covering all continents except Antarctica is available for a 25 years period spanning from 1982 to 2006, were obtained from the Advanced Very High Resolution Radiometer (AVHRR) instrument onboard the NOAA satellite series 7, 9, 11, 14, 16 and 17. This dataset that has been corrected for sensor inter-calibration differences and other effects related to vegetation change, such as: solar zenith angle and viewing angle effects, volcanic aerosols, missing data in the Northern Hemisphere during winter using interpolation, short-term atmospheric aerosol effects, atmospheric water vapor effects, and cloud cover (Tucker et al., 2005).
Satellite sensor data with scan angles < ±40 from nadir was improved through remapping to the output bin closest to the centre location of each 8 km equal area grid cell. Maximum value NDVI was composed for a bimonthly time step, from first day of the month to the 15th day, and from 16th day till the end of the month. (Tucker et al., 2005) Maximum value compositing was used to simultaneously minimize atmospheric and directional reflectance effects (Holben and Fraser 1984, cited in Tucker et al, 2005). Every image was then manually checked for navigation accuracy and images with >±1 pixel navigation errors were investigated. The days with the navigation error were separately manually reprocessed, the NDVI data was calibrated and formed into maximum value composites. Cloud screening for the all continents, except Africa, was provided by a channel 5 thermal mask of 0°C (Tucker et al., 2005).
The quality of the AVHRR NDVI dataset was assessed in different ways through the time. One of the assessment methods to be mentioned was comparing the time-series NDVI values with desert objects, which were not used in NDVI calibration and normalization procedures, at different latitudes. The results showed that among different satellite periods, for all the investigated deserts, NDVI slopes are very close to zero (Tucker et al., 2005).
Some limitations of this database should be mentioned. At first, according to Tucker et al. (2005), due to the lack of consistent AVHRR water vapor records between 1981 and 2004, they were not able to correct atmospheric and directional reflectance explicitly. Limitations may also be found due to the coarse special resolution of 8 km (Defries et al., 2000, cited in Elmzoughi et al., 2008).
The precipitation data was derived from the Global Precipitation Climatology Project (GPCP) webpage. This project is an element of the Global Energy and Water Cycle Experiment (GEWEX) of the World Climate Research program (WCRP), established to provide time-series monthly mean precipitation data, beginning from the year 1979. This target was completed using rain gauge data from more than 6,000 stations merged with infrared and microwave satellite estimates of precipitation (http://cics.umd.edu/~yin/GPCP/main.html).
Precipitation data from the Global Precipitation Climatology Project are presented as interactive maps and graphs (see Appendix 1). Due to the lack of precipitation dataset in a raster form, precipitation patterns will be explored visually. Therefore, certain level of approximation should be taken into account as limitation for this study.
A climatic indicator of desertification is the drought index. The Self-Calibrating Palmer Drought Severity Index was extracted for the Azerbaijan study area, in order to statistically correlate the vegetation and climate changes (Wells et al., 2004, Appendix 2). The Drought Index was extracted from Grid cell with coordinates 49.25E and 40.25N, the cell size was 0.5° x 0.5°. The Palmer Drought Severity Index is based on a water balance model, and was in use since 1965 as a standard for measuring metrological drought, especially in USA. The classification of PDSI values has 11 categories from the extreme drought to extreme wet spell, with values 0.49 to –0.49 classified as normal conditions. The index calculates the difference between the volume of precipitation appropriate for the normal conditions and the amount of actual precipitation, taking into account regional and seasonal climatic differences. However, PDSI values were not suitable for all the diverse climatologic regions. Therefore it was improved by numerical methods replacing the averaged constants of the Palmer’s formula using local climate products. The self-calibrating index (SC-PDSI) has been created and tested, and confirmed to provide more accurate comparison of the index at different locations (Wells et al., 2004).
Visualization and Analysis
Visualization and analysis of the available data using image processing and analysis tools integrated with a Geographic Information System (GIS).
The first task for the visualization and analysis of the Normalized Difference Vegetation Index (NDVI) data obtained from the Global Inventory Modeling and Mapping Studies (GIMMS) dataset was the images processing. This task was conducted in ENVI software, best suited to process satellite images, where 600 NDVI sub-images were extracted for each of five Caspian region countries for the years 1982-2006 (24 images per year). Each image represented NDVI temporally l averaged over a 15-day period. ENVI software was utilized for creating 1 file per year that should comprise all 24 images. This task was achieved by applying the Layer Stacking tool that imported all 24 images per year into 1 output file, containing 24 bands. As a result of this process, 25 new files were created for each of the five countries.
For the generalization and reduction the complexity of the NDVI data, it was decided to calculate numerical summaries, the annual Mean and Standard Deviation (STD). The purpose of these statistics is to describe individual variables in numerical terms, to summarize the row data (Walford, 1995). The Mean is a measure that in the form of single number presents a summary description of a series of numbers; it is calculated as a sum of the numbers, divided by the number of numbers. However its weakness is in giving the same weight to all the values, which makes it predisposed to the influence of extreme values (Longley et al, 2005).
Standard deviation is the square root of the variance (the mean squared difference from the mean). Variance also measures dispersion, however it not very convenient measure for descriptive purposes because variance measure is given in squared units, not in the same as units as originally variables were measured. STD overcomes this disadvantage by estimating squared root of variance. Dispersion measures are adopted to express the range of NDVI variation in highly seasonal environments (Liu et al., 2003). Standard deviation images were used to map vegetation decline under desertification (Liu et al., 2003). Hence, standard deviation statistics measure NDVI variation over time at every pixel, and therefore dynamics of the vegetation over time, as the main indicator of the scales of land cover deterioration processes (Tucker et al, 1991; Liu et al., 2003). Imagery displaying the mean and STD of NDVI were produced using ENVI image processing software. 50 images for each of the five countries of the Caspian Region were created (NDVI mean and STD per year for the 25 year study period). These data were exported to ArcMap for further processing (Figure 3).
As it was stated at the introduction, the processes of desertification were supposed to be detected and analyzed at the Caspian Region, which was defined as 160 km stripe from the coast line. In order to produce this strip, the following steps were undertaken in ArcGIS software. At first, the Caspian Sea coast line was digitized from the one of original NDVI images in ArcMap. This resulted in a shapefile defining the water mask applied to all the data in the NDVI processing. The next step was to digitize political boundaries of all five countries to be investigated. As a result, 6 shape files were created in ArcCatalogue (Figure 4a). After this, a buffer of 160 km was created around the Caspian Sea, and then subdivided on 5 separate polygons, 1 file per country sharing region ≥ 160 km from the Caspian Sea for the following analysis (Figures 4b & 4c).
Figures 4a-4c. Coastline and political boundaries digitizing; Buffering of the coast line; Creation of 5 new polygons, sharing countries’ region ≥ 160 km from the Caspian Sea.
After five zones for investigation were defined as polygon layers in ArcMap, the new NDVI images were created by dividing Mean and STD layers by study polygons. This task was performed by Extract by Mask tool in ArcMap (Figure 5).
Defining a visualization and analysis system for the NDVI images in this study to analyze desertification was a primary task of this investigation. Guided by the previous research methods described in the Theory section (above), I will consider desertification process as the process of vegetation degradation, especially in arid and semi-arid areas, lasting for long periods that could not be fully explained by fluctuations in precipitation. From the NDVI point of view, values less than 0,1 will be considered as desert, between 0,1 and 0,2 as semi-desert, areas with 0,2 – 0,3 values range will be considered as steppe, 0,3 – 0,4 as shrub and grassland, and finally pixels with more than 0,5 will be assumed having forest land cover. These definitions were based on the literature (cf. Anyamba & Tucker, 2005). An NDVI value less than 0.2 is therefore regarded as an indicator of the prevalence of drought conditions. Hellden and Tottrup (2008), classified regions that had a long-term mean monthly NDVI value between 0.1 and 0.5 as subject to desertification. Weier & Herring (1999) assumed NDVI values below 0.1 correspond to barren areas of rock or sand and values between 0.2 to 0.3 indicating shrub and grassland.
However, climate conditions between Eastern and Western coastal areas of the Caspian Sea differ as it was described (see Introduction, above). The East coast has lower precipitation, and arid to semi-arid climate, while the West coast climate types stretch from arid to sub-humid and humid in the south-western zone. Taking into account that vegetation cover characteristic features differ between littoral countries, it was decided to apply 2 different vegetation classifications for the NDVI mean value images (see Table 1). For classifying the vegetation types of the Caspian Region, the Vegetation map from the atlas of former USSR (Union of Soviet Socialist Republics) was investigated (see Appendix 3).
The mean yearly NDVI data was presented in raster form for visual analysis. After raster images of the mean NDVI were created for every year, and for each of five countries, data was classified in accordance with the Table 1. Vegetation cover was presented as areas with hue symbolization, which means that colors were applied to discriminate between classes. The final maps were created to visualize the time-series changes in vegetation cover in the Caspian Region countries (Appendix 4).
For better analysis of the supposed desertification patterns in the Region, data was collated into Excel tables, with the classes organized as rows, and NDVI values as columns. Pixel counts were converted into square kilometers, on the assumption of 64 km2 per pixel when pixel resolution is 8 km2. Histograms were produced to compare the area distribution of each vegetation class every year (Appendix 5).
For more clear visualization of changes in the Caspian Region countries’ vegetation cover, and therefore possible desertification processes detection, Standard deviation (STD) of NDVI values maps and pie charts were also created. Pixels with high STD values correspond to intensive vegetation cover, while pixels with low STD values should correspond to the areas with scarce vegetation. Difference between pixels natural grouping ranges of STD values is presented by saturated color symbolization, the darker color the higher STD value (Appendix 6). The circular pie charts show percentage of areas with different STD values range. Time-series presentation of STD in the form of pie-charts is used for detecting changes in vegetation cover over the period of time (Appendix 7 & 8).
(for maps, tables and graphs, see Appendixes 1 – 8)
In the study area of Azerbaijan, the precipitation graphs show a reduction in precipitation of more than 50% over the period 1982-2006 (Appendix 1a). However there were several years, between 1986 and 1988, with better than average precipitation. The peak of mean monthly precipitation in these years was approximately 4.7 mm/day; in 1993-1994, when the maximum value was about 4.4 mm/day. Even though precipitation exhibited an approximately 6-year periodic patterns in Azerbaijan coastal area from 1982 to 1995, precipitation since appears to be in a steady decline.
The Drought Index for Azerbaijan supports the precipitation data with regard to the long period of drought conditions at the late years of the study period. According to Drought Index data, since the year 1997 the climate conditions in the study area are gradually worsening from incipient drought to extreme and severe drought in years 2000- 2002. The correlation statistical analyses were applied for revealing dependency of vegetation classes on the changing climate conditions represented in Drought Index. The land cover classes were correlated with the annual mean of Drought Index variable by Spearman’s rank correlation analysis. The correlation coefficients (r) of – 0.56 and -0.54 for the desert and wormwood-solonchak semi-desert area respectively shows negative relationship between total area of this land cover classes and the drought index. The p-value in these cases shows high significance of correlation (Table 2). However, in Table 2 it could be noticed that results of the correlation analysis for the rest classes did not reveal relationship with drought index, with exception of the forest class slight correlation. The forest land cover class appear better correlation of r = 0.48 with the drought index during dry season from July to October, and r = 0.47 in August (the driest month) (see Table 3). The wet and dry seasons’ definition was based on the mean monthly drought index for 1901-2002. The wet season in Azerbaijan covers months from January to April.
The correlation analysis of the Drought Index with the ephemeral half-shrub and semi-desert class, meadow steppe and shrub and grassland class did not reveal correlation. However, this could be explained by imprecise land cover classification definitions due to the lack of field measures. Merging middle-value (semi-arid classes) of mean NDVI classes, which are ephemeral half-shrub semi-desert, meadow steppe and shrub and grassland classes, correlation analysis with the Drought Index return stronger correlation (Table 3). Therefore, it could be supposed that changes of middle-value land cover classes occur internally between these classes. Hence, classes’ boundaries that are difficult to define properly could be limitation for this study, causing problems with correlation analysis.
In Azerbaijan the mean of NDVI images, display a degradation of the vegetation in 2006, relative to the year 1982, mostly in Absheron, Aran and Shirvan plain areas. Areas mapped in green show high mean NDVI value, while areas in red show low NDVI values, and classified as deserts (Appendix 4). By comparing images of mean NDVI for 25 years, we can visually estimate changes in vegetation. The images of 1994 – 1995 show growth patterns in vegetation (Figure 6), and this could be explained by drastic rise break in the precipitation graph in years 1993-1994.
Table of contents :
1.2 Geography and climate of the study area
1.3 Aim of the study
2.1 The concept of desertification
2.2 NDVI as an indicator of desertification
3.1 Source data
3.2 Visualization and Analysis
Appendix 1a-c: Monthly mean precipitation in mm/day, for the years 1982-2006
Appendix 2: Drought Index, Azerbaijan
Appendix 3: Clipping of the USSR atlas
Appendix 4: Vegetation classification maps
Appendix 5: Histograms, showing vegetation types distribution through 1982 – 2006
Appendix 6: NDVI Standard deviation maps
Appendix 7: NDVI Standard deviation charts
Appendix 8: NDVI Standard deviation charts and histogram, Azerbaijan