This page describes the characteristics of the products output by SAMIS and provides guidance on their analysis and interpretation.
· Characteristics of SAMIS Products
· 10 day Rainfall Comparison to Average
· Cumulative Rainfall Comparison to Average
The products that can be generated by SAMIS when all the inputs are available, include:
· Rainfall related parameters – rainfall occurrence parameters such as number of rain days (over 10 days and a month) and length of active dry spell; 10 day rainfall amounts and derived parameters such as monthly and seasonal cumulative amounts and respective climatological departures.
· Vegetation related parameters – Vegetation index and departure from its long term average. At the time of writing, the system uses NDVI data downloaded from the ADDS (Africa Data Distribution Service) website.
The specific parameters produced by SAMIS are :
·Dekadal Rainfall Amounts
·Dekadal Rainfall Anomaly (ratio against climatology)
·Cumulative Rainfall from March to current dekad
·Cumulative Rainfall Anomaly (ratio against climatology)
·Dekadal Number of Rain Days
·Length of Currently Active Dry Spell (maximum over 30days)
·Vegetation Index
·Vegetation Index Difference from long term mean
·Monthly Rainfall
·Monthly Rainfall Anomaly
·Monthly Number of Rain Days
The display, analysis and further processing of these products at the Sudan Meteorological Authority (SMA) and the Sudan Early Warning Unit (SEWS) takes place within a low cost, widely used GIS/IP package, IDRISI32. Further processing can produce :
·Reports on the season’s evolution using on-screen display functions
·Colour coded maps for web / bulletin based dissemination
·Tables with target area statistics (averages,max,…) – the target areas can be crop production areas, provinces, etc,.
·Plots of parameters at given target point locations – e.g. NDVI with rainfall, current vs climatological cumulative rainfall, etc,.
All SAMIS outputs are raster data – raster data is the GIS term for regularly gridded data; SAMIS outputs are produced in a lat-long grid with the following limits :
24°N, 20°E to 2°N, 40°E
and with a resolution of 20 pixels per degree (about 5.5 km – the nominal spatial resolution of METEOSAT imagery).
Raster data are usually stored as binary files and are displayed as images in conventional GIS/IP software where each pixel value represents the value of some physical quantity (rainfall amount, vegetation index). A typical image display is shown in Fig 1a.
For public dissemination, these data images are further processed by :
·masking out non-Sudanese territory
·tranformation of original values into suitable intervals of values (classes)
·overlay of an appropriate title and legend
They are displayed with suitable title and legend and overlaid with geographical information such as country and state borders, main towns, etc,. These charts are saved as graphics files (gif, jpg) for incorporation into bulletins and webpages. An example is shown in Figure 1.
Figure 1 - Example of a SAMIS 10 day rainfall image (left) and the corresponding chart (right)
Figure 2 - Example of chart of 10 day rainfall estimate (Dek 1 July 2003)
The 10 day rainfall product is derived from a combination of satellite and gauge data. A satellite-only estimate is derived first and then adjusted with contemporaneous rain gauge data. Should gauge values be unavailable, the satellite only product is used.
The input is composed of images of a satellite parameter called CCD (Cold Cloud Duration) which represents the duration (over a period of a dekad) of (assumed) rain bearing storm clouds. This is derived from half hourly thermal infra-red imagery from the METEOSAT satellite. Several CCD levels are used, each representing duration of storm clouds at different intensities. These images are converted into a 10 day rainfall estimate by means of a multivariate linear relationship, whose spatially variable coefficients are derived from a calibration against historical rain gauge data.
In each dekad, this satellite estimate is likely to overestimate and underestimate actual rainfall in different places. To correct for these errors, we use the rain gauge data available for the same dekad at SMA, which consists of 28 rainfall values for their synoptic stations.
Using these contemporaneous gauge values and the satellite estimate, differences between gauge and satellite are formed at the gauge locations and interpolated into a difference (or bias) image which contains negative values (where the satellite underestimated the gauge rainfall) and positive values (where it overestimated). This difference image is then subtracted from the first satellite estimate to yield the final combined estimate.
Pixel values represent 10 day rainfall in mm (1mm = 1 litre/m2 = 10m3/ha). Like most SAMIS products these values should be interpreted in their spatial context. Remember that the spatial resolution of these images is 5 - 6 km, giving pixels of the order of 30Km2 and will represent the average rainfall over an area of this size.
Do not expect single pixel values to compare well with individual gauge values and keep in mind that two gauges even in close proximity can measure widely different values. The final combined estimate always lies between the two sources (satellite and gauge); the nearer one is to gauges, the closer the value will be to that of the gauge and vice-versa.
Within Sudan, these estimates are expected to behave well over most of the territory. Until a quantitative evaluation is completed and reported on, caution should be exercised in the following areas :
· Near NE coastal areas (Port Sudan region)
· North of 17N and away from the Nile (where there are reporting gauges), the values of the estimates will never be high and should be taken as merely indicative of presence of rainfall. Small amounts in these regions (<5mm) may be missed altogether.
· On the extreme SE of the country near the border with Kenya, in case the estimates indicate high values (>20mm or so), particularly when the satellite underestimated rainfall in the Juba region. Low or zero values are expected to be reliable.
In coastal areas, rainfall may arise from mechanisms other than large storms, and in this case the methodology used performs poorly. Also the presence of mountains close to the shore creates additional problems.
In the other areas, there is a near total absence of rain gauges and hence calibrations for these regions are less reliable. Even where gauges exist, note that calibrations are based on non-zero rainfall values and very arid regions by definition do not provide many of these.
Evaluation of absolute amounts of rainfall obviously depends on the purpose of the evaluation : are you trying to evaluate whether this dekadal rainfall was enough for crops or whether you are likely to face localised flooding / water logging? There are fairly obvious things you should look out for :
· low (<20mm) amounts midway through the crop gowing period in a typical production area – but even this needs to take into account the situation in the previous dekads (as you may have enough water stored in the soil).
· very high (>100mm) amounts (corroborated by even higher gauge values) over mountainous regions – still, there are many other factors at play namely the distribution of the rainfall during the 10 day period (you could look into the number of rain days product) and the nature of the terrain (soil, vegetation, topography).
· As a general rough rule, you may look for dekadal amounts between 30mm to 50mm as likely to satisfy general crop water requirements over the central third of Sudan within the July to September period.
Having a standard against which to compare the 10 day amount simplifies matters to a degree (see next section).
Figure 3 - 10 day rainfall (Dek 1 July 2003) expressed as a fraction of the long term average 10 day rainfall for the same dekad (Dek 1 July). Note the very high values in the North of Sudan (see text for details)
Comparing the rainfall to a given standard helps our interpretation. Common practice is to compare to an average long term value which is meant to represent the usual conditions, so that the comparison shows how unusual the current rainfall amount is. Of course this can apply to any parameter, not just rainfall.
The comparison can be expressed in the following ways :
· As the simple difference between current value and long term average value
· As the ratio between current value and long term average value
· As the standardized difference, i.e. the difference divided by the long term standard deviation (z score)
Note that the first option is difficult to interpret since a difference of 30mm (say) is quite significant in areas of low rainfall but not so in areas of high rainfall. The third option is atractive but requires an estimate of the long term standard deviation as a map for the whole country, which was not available. Also, it is difficult to interpret by non-specialists. The second option was chosen after consultation with SMA.
A requirement for this work was the preparation of the long term average 10 day rainfall amount (we’ll refer to this data as LTA rainfall) for all dekads since March Dek 1 to October Dek 3 as images covering the whole of the Sudanese territory. These were prepared by a combination of gauge rainfall long term (1971-2000) averages and shorter term (1990-2000) satellite data (CCD parameter, see previous section for details).
Figure 4 - The 10 day rainfall (left) and the long term average rainfall (right). Note that the LTA image has a minimum of 1mm everywhere. The ratio between them originates the image shown in Figure 3.
To obtain the ratio it is a simple matter to divide the two images (current and LTA 10 day rainfall) and express the result in percentage (see Figure 4).
Pixel values represent percent ratio of 10 day rainfall over LTA rainfall. Values below (above) 100 indicate below (above) average rainfall. The charts produced follow the usual convention of warm colours for drier than average conditions and cool colours for wetter than average conditions. The percent interval 90%-110% is considered normal and displayed in neutral grey. Keep in mind that conclusions apply strictly to the current dekad.
This product will suffer from limitations in the same areas as the 10 day rainfall amounts. The LTA data is only as good as the data used to build it and far away from gauged areas it is relying mainly on a fairly short sequence of satellite data.
These charts help you to frame the 10 day rainfall amounts in a climatological perspective by relating them to average conditions and hence allowing you to evaluate the 10 day amounts better in terms of quality.
You may expect crops and vegetation to be tuned to the average conditions and hence substantially lower than average conditions may indicate potential problems. Keep in mind that this chart only applies to the current dekad and that any impact of strong deviations has to be judged against the situation in previous time steps. Reporting to Fig 4, the small percentages obtained in the Juba region does indicate significantly below average rainfall for this dekad, but any judgement on its impact would have to consider the situation in the previous dekads : if this is an isolated below normal dekad in the middle of normal or above normal dekads, there is no need to worry. In any case, a situation like this one should be flagged for follow up in the next 10 day period.
Evaluation of this parameter needs particular care where LTA rainfall is very low, i.e. Northern third of Sudan (roughly north of about 16N) and extreme SW, near Kenya. Here it is frequent to find very strong deviations (see Fig 4 for an example). These should not be over-interpreted :
·The high values are due to the fact that in these places it is very easy to substantially exceed the LTA – you only need 10mm rainfall to get 500% of LTA rainfall where the LTA is 2mm. In these areas, well above average rainfall does not imply large rainfall amounts.
·The low values are due to the fact that mostly the estimated values are quite small and hence it is also frequent to find small percentages – estimating 3 mm where the LTA is 9mm gets you 33% of LTA rainfall. In these areas, well below average rainfall does not imply drastically drier conditions.
Figure 5 - Cumulative rainfall at the end of Dek 1, July 2003. Note the usual northward decreasing trend in values. Some patterns in the Northern extreme of Sudan are due to misinterpreted cloud features
The cumulative rainfall in a particular dekad is derived by simple accumulation of all 10 day rainfall amounts since the beginning of the season. This is taken as Dek 1, March.
Pixel values represent cumulative rainfall in mm (1mm = 1 litre/m2 = 10m3/ha). Usual convention of cool colours for the higher values was followed. This product will suffer from limitations in the same areas as the 10 day rainfall amounts, but because it is a cumulative product relative inaccuracies will tend to decrease the further you progress through the season.
Even then, do not expect agreement between pixel values and gauge values, in particular in the early stages of the rainfall season. Bear in mind that significant variation in the pattern of cumulative seasonal rainfall can occur even over very short distances which satellite based methods typically do not represent well.
Like any absolute parameter, cumulative rainfall may be difficult to interpret in the absence of a comparative standard or professional experience about the region in question. The comparative standard is the long term average (LTA) cumulative rainfall – see next section.
In Sudan, you should expect to see a regular trend of decreasing rainfall along SW-NE direction, veering Northwards in the northern half of the country. Deviations from this regular pattern (e.g. pockets of low values) may highlight interesting areas for more detailed monitoring.
Figure 6 - Cumulative rainfall (Dek 1 March 2003 – Dek1 July 2003) expressed as a fraction of the long term average cumulative rainfall for the same period. Note the very high values in the North of Sudan (see text for details)
Please refer to Section 4.2 for details of what to and how to compare rainfall amounts.
A requirement for this work was the preparation of the long term average (LTA) cumulative rainfall amount at all dekads since March Dek 1 to October Dek 3 as images covering the whole of the Sudanese territory. This LTA cumulative rainfall data set was derived by simple cumulation of the individual 10 day LTA data images (see Section 4.2).
Figure 7 - The current cumulative (left) and the cumulative LTA (right) rainfall (Dek 1 June). Note that the LTA image never registers zero value. The ratio between them originates the image shown in Figure 7.The 10 day rainfall (left) and the long term average rainfall (right). Note that the LTA image has a minimum of 1mm everywhere. The ratio between them originates the image shown in Figure 6.
To obtain the ratio it is a simple matter to divide the two images (current cumulative and cumulative rainfall LTA, see Fig 8) and express the result in percentage (see Fig 7).
Pixel values represent percent ratio of cumulative rainfall over LTA cumulative rainfall. Values below (above) 100 indicate below (above) average rainfall. The charts produced follow the usual convention of warm colours for drier than average conditions and cool colours for wetter than average conditions. The percent interval 90%-110% is considered normal and displayed in neutral grey.
Note that unlike the 10 day rainfall ratio, these cumulative values present an integrated view over the season to the current dekad and smooth the inaccuracies in the cumulated data; hence it is a more useful parameter in that it allows you to make more definite statements about the quality of the season in terms of rainfall.
Considering this is a seasonally integrated parameter and that crops and vegetation are likely to be tuned to the average conditions, substantially lower than average values may indicate potential problems. However, it is important to consider the timing, when lower than average values are commented on :
· Early in the season these indicate a later than usual arrival of the rains, and the situation may be reversed in a couple of dekads.
· Past the middle of the season and any recovery is not only less likely, it may arrive too late to minimise negative impact on crop production.
Though cumulative data integrate the season rainfall to the current dekad, it is still important to consider past recent dekads. As important as knowing if you are under lower or above average conditions is to have an idea of the trend of the situation.
So, always analyse the the current dekad in conjunction with the past 2 or 3 dekads, so you can state whether the situation is improving, worsening or stable. Referring to Fig 6, the area in East central Sudan in orange colours indicates lower than average rainfall but within a continuous trend of improvement when compared to the situation of the past 3 dekads (not shown).
For the exact same reasons outlined in section 4.2, evaluation of this parameter needs particular care where cumulative LTA rainfall is very low, i.e. Northern third of Sudan (roughly north of about 16N) and extreme SW, near Kenya. Here it is frequent to find very strong deviations (see Fig 6 for an example), which should not be over-interpreted. Please refer to section 4.2 for details.
Figure 8- Chart of number of rain days in Dekad 1 July 2003
The number of rain days within a dekad (or month) is derived from the simple cumulation of a product which has so far not been mentioned – the daily rainfall occurrence. This product estimates on a daily basis, whether rainfall occurred (yes/no) without regard for the actual amount.
The daily rainfall occurrence, unlike the 10 day rainfall amount, is derived exclusively from the satellite data without recourse to any rain gauge information. Like the 10 day rainfall amount it is also derived from METEOSAT CCD data (see section 4.1 for details) but daily CCD rather than dekadal.
Satellite derived rainfall estimates based on CCD data are too unreliable when made for periods of less than 10 days. However, on a daily basis, CCD can act as an indicator of rainfall occurrence, even if no reliable estimates of actual amounts can be made – the longer the duration of storm clouds (as measured by the CCD) more likely it is for rainfall to have occurred.
Although we can’t have daily values, from the daily occurrence we can still extract useful information on the distribution of rainfall.
Important - In the context of SAMIS, when we talk of rainfall occurrence, we mean rainfall occurrence above 1mm. Tiny rainfall amounts (<1mm) are prone to error by a variety of reasons and better results are obtained by defining a suitable non-zero low value.
The daily rainfall occurrence product is internal to SAMIS. This means it is produced and used in the derivation of other products, but it is not analysed by itself or disseminated to outside users. From the daily rainfall occurrence two products are derived :
·Number of rain days on a dekadal and monthly basis
·Length of currently active dry spell (dekadal product, but analyses past 30 days)
Pixel values represent the number of significant rain days. Charts use the same colour bands of the dekadal rainfall estimates.
The number of rain days images is derived from daily rainfall occurrence estimates tuned to daily rainfall above 1mm. Being based on a probabilistic approach, the method will be more uncertain the closer one is to the chosen rainfall threshold. This means that the method is expected to perform best for large daily falls. Put another way, it is unlikely to miss significant daily rainfall, but it is unsurprising that it misses daily falls of 5mm or so, though not in any consistent or systematic way.
Because of this, the images of number of rain days and the images of dekadal rainfall estimates will not have the same geographical outline. This means that you will find areas where dekadal rainfall has been estimated as above 0, but where no rain days are estimated (the reverse may also happen but much less frequently). Note that by the simple fact that rainfall occurrence is estimated as rainfall above 1mm, you could have dekadal rainfall of up to 10mm for 0 rain days and still consider the methods as consistent.
However, for dissemination purposes, the two estimates are made consistent by assigning 1 rain day to all pixels where (dek rainfall > 0 and n rain days = 0). This is to avoid misinterpretation of the quality of the products and the presentation of lengthy technical explanations to clarify their characteristics.
Number of rain days over a dekad or month are useful indicators of the distribution of rainfall within the period. You should expect a general agreement between the spatial pattern of dekadal/monthly number of rain days and rainfall amounts : in general, the areas of highest number of rain days match the areas of highest rainfall.
On a monthly basis, identification of areas of low number of rain days and high rainfall may help point out areas of poor temporal distribution of rainfall. At this stage, it is not recommended to draw conclusions about the occurrence of very large daily falls from cursory rainfall intensity calculations, i.e. rainfall amount divided by number of rain days. This may yet prove a reasonable indicator of such events, but at the time of writing, it had not been verified even in a qualitative way.
Figure 9 - Chart of length of currently active dry spell by the end Dekad 1 July 2003
Length of dry spells are one of the most useful parameters for crop monitoring since it allows an evaluation of the quality of the distribution of rainfall within a given period. Two consecutive dekadal rainfall estimates of 20mm say, could hide a dry spell of up to 18 days, if the rain had fallen on the first day of the first dekad and last day of the second dekad.
Length of dry spell is computed by counting the consecutive non occurrences of rainfall as estimated by the daily rain occurrence product. For details on the daily rain occurrence product see previous section 4.5.
Pixel values represent the length in days of the active dry spell. Spells shorter than 5 days were lumped into an initial class, under the assumption that they are unlikely to cause any significant impact on crops. Warm colours were chosen for the longer dry spells and cooler colours for the shorter ones.
The SAMIS dry spell parameter is produced each dekad and evaluates the length of the active dry spell over the 30 days previous to the last day of the dekad being considered. Active since it considers dry spells that are still in existence on this last day. Given the 10 day interval, dry spells of up to 9 days may escape detection or be evaluated as being up to 9 days shorter than reality.
Note that length of a dry spell is a somewhat fragile parameter in that you need only to estimate a single false rain day for a long dry spell to be considered significantly shorter than it is.
Because of this, as a rule of thumb, the dry spell length should be evaluated when large values (say above 8 or 9 days) are present over extensive areas of half a degree or more; small numbers of pixels in isolation may be unrepresentative variability. Spatial consistency is also of importance, i.e. larger values of dry spell length should occur in contiguous groups of pixels, usually (but not always) surrounded by or linked to pixels of smaller values. A typical example is presented in the picture below :
Figure 10 - Zoom in the same product as in previous figure. Note the extensive and spatially consistent area of high values of dry spell length. The highest values tend to cluster in the middle of the area or join around the areas pictured in white, which represent areas where no significant rain days have yet occurred.
Dry spell lengths are relatively easy to evaluate. It is advantageous to have an idea of the length of dry spell that particular crops can withstand so as to assess likely impact. Remember that a long dry spell may be preceded by plentiful rainfall which will mitigate its impact. For detailed interpretation, look at the 10 day rainfall amounts for the dekads preceding the beginning of the dry spell.
It happens sometimes that the information provided by the dry spell image and the rainfall image may seem in disagreement, for example a given area is under a dry spell but the dekadal rainfall is not zero. There are two main reasons for this :
·A SAMIS dry spell does not imply total absence of rainfall. It is better described as period of predominantly dry conditions. Within a 20 day dry spell you may find two or three isolated days with small amounts of rain (<5mm) which the method tends to discard. The important thing to keep in mind is that absence of significant daily falls (> 5mm, say) is well represented.
·In some regions, the dekadal rainfall estimation method may not be able to generate values below a given threshold, say 5 or 10 mm.
Total clarification can only be obtained with recourse to the daily rain gauge values. This is not available to users outside SMA; SMA personnel will look up daily gauge values if available and provide summary explanations to users if and when required.
Figure 11– Vegetation index images separated by one month, June Dek 1 (left), July Dek 1 (right). Yellow to green denotes low to high NDVI values. Note the northwards progression of the vegetation
To derive a vegetation index, you need data from a satellite that measures the solar radiation reflected by the Earth’s surface (known as reflectance). Different surfaces on Earth reflect solar radiation in different ways – snow and ice reflect almost all solar radiation (they are very bright), while deep forests reflect very little (they are very dark). Differences are not only in amount, but also in the way the reflectance varies with wavelength – vegetation reflects in the green wavelengths (that is why it is green), very little in the red wavelenght (where they absorb radiation for photosynthesis) and a lot in the Near InfraRed (NIR) wavelength. Other surfaces have different reflectance behaviour – water reflects almost nothing in the NIR and also little in the Red.
It is by exploiting these differences in the way light is reflected, that satellites can identify different surfaces – snow, bare soil, vegetation, water, …,. For the monitoring of vegetation, we use so called vegetation indices. All vegetation indices exploit the fact that vegetated surfaces reflect solar radiation much more strongly in the NIR (Near InfraRed, wavelength around 1.1μm) region than in the Red (Visible, wavelength around 0.7μm) an effect that is the more pronounced the more vegetated the surface is. This contrast also exists for bare soil, but to a much weaker degree and is largely absent or reversed for water bodies and clouds.
There a large number of different vegetation indices, but the “default” is known as NDVI – Normalised Difference Vegetation Index. For the case of the NDVI, we have :
It is easy to see that the larger the difference between the two reflectances, the higher the NDVI. Reflectances vary from 0 (no reflectance) to 1 (full reflectance). The NDVI can theoretically vary between –1 and +1, but in real life values are mostly confined within the interval -0.3 to 0.75. Water and clouds mainly have negative values while bare soil has low positive values (0.05 is the soil NDVI rule of thumb value). Values for not too sparse vegetation are usually above 0.1.
The satellite from which NDVI data is derived is the american polar orbiting satellite series NOAA-N (where N identifies a particular satellite in the series, e.g. NOAA-11). These satellites have a sensor on board known as AVHRR (Advanced Very High Resolution Radiometer). You will hear about “NDVI from NOAA-AVHRR”. Another satellite that is able to produce NDVI since 1998, is the french SPOT satellite (SPOT-VEG).
In principle, the problem looks of easy resolution – acquire reflectance data in the two key wavelengths (NIR and Red), combine them and derive a parameter that identifies vegetation. However, within this simple scheme there are complicating factors :
·Atmosphere – the atmosphere will interfere in the measurement of reflectance. The main intervenients are water vapour and aerosols (haze). These cause the NDVI to be lower than what it is.
·Viewing Geometry – depending on details of its orbit, the satellite that measures the reflectances will look at the same point from different directions, with the Sun in varying positions. This causes the NDVI to vary – usually, NDVI measured from off-nadir (away from the vertical) positions is lower than NDVI measured directly overhead.
·Sensor Degradation – the sensors that measure the reflectances change their characteristics in time and are periodically renovated. This may cause problems when comparing NDVI values from different sensors.
Because of the fact that most interference in the NDVI will tend to depress it, the NDVI products usually available are so called Maximum Value Composites. This means that a period is chosen (nearly always 10 days) and within this period one takes the maximum NDVI; this 10 day maximum NDVI is taken as representative of vegetation conditions. The 10 day interval is a compromise – hopefully long enough to catch a clear observation day, short enough so that vegetation doesn’t vary significantly.
To derive NDVI images, one needs to follow a long processing task. Typically :
·Geometric correction and Geographic co-location – to locate the image on Earth
·Calibration – to derive reflectance from raw satellite values
·Cloud Masking – to remove cloud affected data
·Atmospheric Correction – to minimise atmospheric interference (optional)
·Forming the MVC – deriving the maximum over the reference period (dekad)
This is usually taken care by processing centres using specialized software. The final result can be made available in a range of spatial resolutions. The original data has pixel sizes of about 1Km, and this may be degraded down to 4Km or 8Km.
NDVI data can be obtained from a variety of sources. In the SAMIS context these are :
·Directly from the SMA NOAA satellite receiver (waiting for start of operations)
·Download from the SPOT-VEG website (with a delay of 3 to 6 months)
·Download from the ADDS (Africa Data Dissemination Service) website.
The first two data sets can be provided at 1Km resolution. Here we consider the third possibility which has data at 8Km resolution but is available in near real time.
Pixel values represent the NDVI value scaled by a factor of 0.003 (i.e. pixel value 150 means NDVI of 0.45). These are usually colour coded in such a way as to contrast vegetation with soil, by choosing earth tones for low NDVI values and green shades for medium and high NDVI values. Cloud covered pixels (to which no NDVI value can be assigned) are coded in white.
The ADDS data has a fairly coarse resolution (8Km) and hence it is unlikely for any pixel to be pure, i.e. to contain only vegetation of a single type. Frequently the pixel will contain a fraction of bare soil in particular within the transition zone between fairly vegetated and arid regions of Sudan. Where it looks at vegetation only, you should expect each pixel to contain a variety of natural vegetation and/or crop types.
For NDVI at finer resolutions such as the 1Km products, you are more likely to find pure pixels.
Assigning particular significance to any given NDVI value is very difficult, if possible at all. The NDVI does have a relationship with vegetation but its quantitative meaning is somewhat obscure. It will increase with increasing vegetation amount/cover and increasing vegetation photosynthetic activity. NDVI of stressed/wilting vegetation will be lower than that of the same vegetation cover in unstressed conditions. Under equal conditions of stress, NDVI will be higher for the denser vegetation cover.
Looking at NDVI images in isolation you can make a general statement about where the vegetation has started to grow, or how far north the vegetation has advanced in environments subject to large intraseasonal variations in vegetation cover such as Sudan. A general rule of thumb is that for NDVI above an interval of 0.05 to 0.1 you will be looking at increasing amounts of vegetation.
Looking at the NDVI images in a temporal perspective is much more useful, in particular if you can examine time series at particular locations. This allows you to examine the NDVI seasonal cycle.
Note – NDVI is subject to atmospheric interference even after compositing over ten days and NDVI curves will tend to be affected by short term fluctuations which do not relate to any vegetation feature.
Assuming atmospheric interference is minimized, NDVI in seasonally arid climates follows a very well defined pattern, characterized by :
· an initial phase lasting until the first significant rains during which fluctuations are small and mainly due to atmospheric interference
· a period of rapid increase usually following the first abundant rains, corresponding to the initial vegetation development.
· a period of stable NDVI around a peak value, corresponding to the maximum vegetation development
· a period of rapid decrease in values corresponding to the wilting of the vegetation with the end of the rains; the speed of decrease depends on the type of vegetation, grasslands displaying particularly rapid ones.
The exact shape of the curve varies and is thought to be related to the prevalent type of vegetation (if there is one). When monitoring rain fed agriculture there is a usual assumption that rain fed crops will develop in tandem with natural vegetation. Bear in mind that crops are likely to be more sensitive than natural vegetation to episodes of water stress.
The value of the maximum NDVI and the magnitude of the upward and downward trends either side of the peak may be used in land cover classification schemes. Detecting upward trends in NDVI has been used to identify beginning of crop or vegetation growing cycle.
Based on the link between NDVI and photosynthetic activity, there is reason to believe that some form of seasonally cumulated or integrated NDVI would relate to seasonal biomass production or crop yield. However, you need to be very careful with the temporal limits of the integration. Published experimental results are mixed, but few have looked at relationships to crop yield using long term series of data.
Best results can be obtained by comparing current NDVI conditions to a standard, be it a long term average or a reference year (see next section). Note that in areas such as Sudan where the NDVI is very dependent on the timing of the rainfall, the seasonal NDVI curves may be shifted back and forth in time depending on the arrival of the rainfall without the peak value being affected. This has implications when interpreting deviations from the normal.
Figure 12– NDVI difference relative to long term average on Dek1 July 2003. In green positive differences, in yellow and red negative differences.
Review the notes presented in sections 4.2 and 4.4 regarding comparison to usual scenarios. In the case of NDVI, current practice is to express the comparison as a simple difference between the current NDVI and a reference value and SAMIS follows this convention.
Pixel values represent differences from current to LTA (Long Term Average) NDVI. The same scaling as for the NDVI applies, i.e. (pixel value * 0.003) yields actual NDVI difference. Long term here is understood to mean no more than 20 years or so. For other data sets, such as the NDVI data from SPOT-VEG, there is no long term record and differences have to be made against a reference year or an average of 4 or 5 years at most.
Colour coding conforms to usual practice – it assigns shades of green to positive differences (current NDVI above LTA NDVI) and shades of yellow/red to negative differences.
Pixels for which current values are set to the cloud value, are assigned the colour white.
Evaluation of NDVI differences is highly assymetric. This means that while positive differences are easy to interpret and have a single well defined significance (more or better developed vegetation than average) the negative differences are much harder to interpret.
This arises because NDVI is affected by atmospheric (among other) interferences and these interferences have the systematic effect of decreasing NDVI values. Hence it is difficult to separate what is a negative difference arising from atmospheric interference from one arising from serious delay or collapse in vegetation development. Here, we are not talking about complete cloud interference since cloud affected pixels are easily identified, flagged and removed from further processing. We are talking about interference from haze and water vapour which does not depress the NDVI values in an unmistakeable way.
If however, the NDVI values have undergone sophisticated atmospheric correction procedures, the above warning may not apply, though it is convenient to bear in mind that atmospheric corrections are not 100% effective.
The warning applies to another usual practice of differencing the current NDVI and the NDVI of the previous dekad, in order to evaluate rates of NDVI growth and identify areas of developing vegetation. In the absence of atmospheric correction, this is prone to serious error, irrespective of the sign of the difference, since any of its two components (the current or the previous NDVI) can be affected by atmospheric interference and be lower than reality.
Is there a solution to this problem? Not for dekad to dekad differencing. When differencing from long term average (LTA) NDVI, which is a smooth curve where atmospheric interference has little effect, the advised technique is to look at (current-LTA) differences for a number of consecutive dekads, say 3 or 4. Atmospheric interference is spatially quite variable and hence negative differences where it plays a role do not as a rule stay fixed in the same spot for more than one dekad.
The advice is to look at 4 consecutive dekadal differences and identify negative differences that affect the same broad locations in all 4; look out for possible slow changes in the values which may indicate worsening conditions or redress from water stress. Another clue to identify negative differences due to atmospheric interference is to look out for cloud affected pixels within the negative difference area.