2.2.   Sorting data

2.3.   Import images and ancillary files

2.4.   Converting digital number (DN) into geophysical values

2.5.   Image quality control
A visual assessment was conducted to evaluate the quality of the data. Some images presented some defects such as random noise, or stripes with extremely high or low data values. Several approaches were developed to remove damaged data as damaged images would alter the calculations of mean images. The most efficient approach was to detect odd pixel by comparing its DN values with its surrounding median value using a 3x3 kernel. If the difference was above a given threshold, that pixel was identified as “noise”, and removed from the image. When the damages were too important, the images were excluded, and replaced by blank images. According to that technique, 8 “damaged” images out of the 720 were replaced with blank images.

2.6.   Creation of historical decade image
720-decade AVHRR images were imported in ERDAS Imagine as a single file: (168 x 246 pixel, 720 Bands). From those 720 "bands", an historical image was produced. It represents the average vegetation status in Afghanistan over 20 years, where the DN value of a given pixel represents the mean DN value for that pixel over the 720 bands.

2.7.   Stratification of historical decade image

2.7.1 Stratification using ISODATA algorithm
The VPI approach relies on the correlation between the vegetation status and the statistical distribution of the NDVI. Such correlation could be performed on a pixel by basis. But according to the procedure developed by Sannier et al. (2000), it is better to stratified the study area into zones of homogenous NDVI value as it drastically reduces the amount of processing.

2.7.2 Land mask for area above 3000 m
To avoid misinterpretation of the data, a land mask was created to remove NDVI value localised beyond a threshold of 3000m. According to FAO, AAH reports and personal communication, agriculture is scarce at such altitude. Beyond 3000 m, it is assumed that agriculture is not successfully performed, only extensive grazing is likely. Actually, the threshold was set at 3100m to take in consideration the 100m error of the digital elevation model (DEM).

2.7.3 Maximum likelihood classification
Then, based on the signatures obtained with the ISODATA (Iterative Self-Organising Data Analysis Technique) algorithm, a supervised classification of the historical image was performed using Maximum Likelihood algorithm and a threshold of 95% probability of belonging to a class so that only the purest pixels will be classified into a class.

2.8. Extraction of mean NDVI value for each vegetation class and decade
Using the stratified historical image and the NDVI time series data, the mean NDVI value was calculated for each 9-stratum of each 720-decade. Zeroes were excluded from the calculation of the mean. Then, the mean NDVI time-series for each 9 stratum was imported into a spreadsheet (9 classes [column] by 720-decade mean values [rows]) to calculate the probability distribution of the NDVI from the historical data.

2.9. Calculation of Probability Distribution of NDVI
Based on the offset and gain coefficients given with the MVC NDVI data, time-series NDVI values were calculated.
The probability distribution of NDVI was calculated for each decade according to the methodology used for assessing the probability of extreme hydrological events (Linsley et al. 1975). In hydrology, the aim is often to extrapolate the distribution in order to predict the size of events that have very low return periods.
In our case, for each decade and stratum, the NDVI values were ranked from highest to lowest. Using the formula defined by Weibull (1939 : A statistical Theory of the Strength of Materials, Ing. Vetenskapsakad. Handl. (Stockh.) , 151, 15): where m is the rank in order of magnitude, and n is the number of years (i.e. 20 years), the probability (p) of having an NDVI less or equal to a given value was defined. This can also be expressed as a return period, (t=1/p), which is the average number of years between occurrences of the event (Linsley et al. 1975). Then the quintile ranges of NDVI for each vegetation class and each decade were calculated to define the five VPI classes as described in figure 2.

2.10. VPI maps production
The vegetation status maps produced for the peak growing season of each class, should demonstrate the spatial and temporal distribution of vegetation in Afghanistan (Sannier et al., 1998).
For a given decade, the vegetation status of given pixel was assigned depending on the class it belongs to (i.e. 1 to 9) and the VPI quintile range defined for that specific class. Vegetation status maps were produced for 9 consecutive decades (from March to May 1993, 1999, 2000, and 2001) according the class probability distribution and the growing season of each class.

3.  Result and Critical discussion
3.1. Analysis of the historical image stratification
In order to find the best stratification, different classifications of the historical decade image using ISODATA algorithm were performed. Each trial was assessed by visual assessment of cluster’s statistical distribution, and by visual comparison with the Global Land cover image produced by the University of Maryland (link) and Afghanistan biodiversity profile (ICOMOG, 1995).
A range of 4 to 16 strata classification was tested, with no significant improvement in term of classification accuracy and, class separability. The statistical distribution of the clusters was checked to ensure that it was unimodal. Consequently, a 9-class stratification was arbitrary chosen, admitting the fact that the 9th class statistical distribution was multimodal due to spectral confusion. Nonetheless, the 9-class stratification remained as such, as that class is mostly outside Afghanistan (i.e. Pakistan). Because of the coarse resolution of the image, clear differentiation within the strata was difficult, and spectral confusion within class 9 was inevitable.
Figure 3 presents the historical image stratified into 9 classes. The elevation mask (black) covers altitude above 3000 m.

3.2.   Annual Variation of NDVI profile
The average NDVI value for each decade and each stratum was plotted in figure 4. It shows the variation of the vegetation response to the sensor through the year. The NDVI ranges between –0.05 and 0.43, which confirms the general low level of vegetation cover in Afghanistan. Indeed, Afghanistan is essentially semi-arid to desert, with less than 10% of the total land area for crop, and most of the rest is extensive grazing, desert or high mountain and permanent ice (Thieme O. & Suttie J.M., 1996).

Figure 4 reflects seasonal “green-up” in springtime from March to April, and drop in NDVI response in autumn from September to December.
The graph confirm the presence of Desert for Class 1 & 2, as NDVI does not vary throughout the year because of extreme arid conditions (ICOMOG, 1995).
Class 5, 6, 7, 8 tend to flatten and fall by the end of May. This is could be taken as an indication of grass steppe vegetation as “it quickly dries up in May after having covered 90% of the surface” (ICOMOG, 1995).
Class 9 represents a mixture of dense vegetation (forest, woodland, cropland) (De Fries, 1998 and ICOMOG, 1995). Its NDVI variation clearly defines 2 specific picks of vegetation, one in March and the second in September. Class 9 is mainly present in the southeastern subtropical area, over Pakistan. According to ICOMOG (1995) we can assume that the September NDVI peak for class 9, results from the influence of Indian Monsoon.

3.4.  Caracterisation of spatial pattern of vegetation using VPI
The VPI map indicates the qualitative status of the biomass for that decade, as being “very low”, “low”, “average”, “high” and “very high” than “normal” based on 20 years of NDVI data. Figure 7 is an example to show how VPI maps restituate the spatial pattern of vegetation status for the second decade of May 1993 - considered as a “good year” against 1999, 2000 and 2001 “drought years” in Afghanistan.

A successive comparison of Vegetation status map maps made possible to monitor the advance and severity of the drought. A quick comparison between 1993 and 2001 maps indicates how the on-going drought has affected Afghanistan vegetation.Areas located beyond the altitude of 3000m were masked. The area supervised by AAH is delimited by the white triangle located in the central region of Hazarajat, across the Bamian and Ourozgan provinces (close to the “mountain mask”).

Within 3 years, we clearly see the shift of vegetation status from “very high” to “very low”. The Hindu Kush region (near the “mountain mask”) seems to be the last area affected by the drought as its VPI remained positive in 1999 and 2000 and is qualitatively better than in other regions. The reason for this could be the snowmelt, which is a vital element for in agriculture. Indeed, Snowmelt brings water to lower valleys and rivers.
Moreover, the main irrigated areas can be easily identified on the 2000 and 2001 VPI maps as river-waters fed vegetation and irrigate cropland in the surroundings of the rivers map (i.e. Helmand River in the South, Oxus River at the North border).

Figure 8 witnesses the progression of the drought since its beginning in 1999, following a Southwest to Northeast gradient. In 2000, the intermediate area between the central regions and southern provinces were affected by the drought as their VPI shifted from “Very high” in 1999 to “Average” and “Low” in 2000. In 2001, few variations were detected, all VPI are “Very low”, which clearly indicates that the entire country is affected by the drought. As for figure 8, we can note that watersheds can be identified on the 2000 and 2001 maps. Northern eastern region of Hindu Kush (close to the “Moutain Mask”) seems to be the last areas affected by the drought The reason for this is the important role played by the snowmelt as explained previously. It is interesting to notice that back in 1999, most of Afghanistan had a positive vegetation productivity indicator, excepted for the Central region -Hazarajat included- whose “Very low” productivity indicator witnesses the beginning of the drought.

This statement tends to match with the USAID/ODFA activities and IDP movements map released in March 2001 (link).
Several areas marked as “Severe drought” and “extremely severe drought” are clearly distinguishable on the seasonal VPI map. The methodology and data used by USAID/OFDA to produce such map is unknown. Therefore it is not possible to validate that conclusion with certainty (cf. Appendix 6 presents VPI seasonal map).

Conclusion and recommendations

• On the whole, the development of the Vegetation Productivity Indicator to monitor the extent and severity of the drought in Afghanistan was successful. Regarding the limited “official” documentation available, VPI succeeded to monitor the drought progression throughout the country and for the Hazarajat region.

• The quality of the stratification of the historical image could be improved. It would be interesting to perform a four-channel classification with AVHRR as suggested by Kerber and Schutt (1986).

• Regarding Action Against Hunger needs in term of spatial information, it would be interesting to undertake the following study to complement the qualitative information given by the VPI :
1- Develop an approach to estimate food deficit in Hazarajat region, based on the methodology developed by Sannier et al. (2000) in Zambia. To do so, a set of precise ground data (geo-referenced production sites, yield…) is mandatory.
2- 70% of Afghanistan agriculture rely on irrigation, which is highly depending on snowmelt collected by watersheds during spring. Consequently, it would be interesting to produce a hydrological model based on snow cover estimation, as described by Baghdadia N. et al. (1996) using ERS-1 SAR data in selected areas. Then it would be possible to perform runoff forecasts for selected drainage basins in Hazarajat Region.
Such indication would be useful to estimate, in winter, water availability for crop production in spring, and consequently mitigate the coming of a new disaster.

Acknowledgement