Previous: Forecasting regional food/fiber production levels Next: Download page Table of contents Index Glossary Images Frames U09-NRM-127: The role of Distributed Data Access Technologies in NRM - for ITC-IDV version 2.7 > Thematic Expert Models > Food security > Forecasting regional food/fiber production levels

7.0.2.0 Does vegetation growth in Africa follow El Niño?
A large-scale atmospheric process in modelling Africa's vegetation status

Introduction

Since no tight boundaries between different regions exist in the atmosphere, changes in one of them can have noticeable effects in another, even if far apart. An excellent example hereof is the connection between ocean surface temperatures in the eastern Pacific and the climate around the world, also referred to as the El Niño teleconnection. As illustrated by the graphic below, this particular teleconnection has important consequences for weather, agro-ecosystems, and consequently economies around the globe. As the time-lag between sea-surface anomalies and local weather around the world is considerable (4~6 months), this large-scale atmospheric process is arguably most researched as a better understanding has the potential of yielding a timely predictor of potential events before they occur, and could thus help save lives and resources. For example, already in 1987 Skidmore (1987) demonstrated that bushfire activity could be predicted well in advance and reasonably accurate for large portions of the Australian continent.
Climatic impacts of warm El Niño events (Oct-Mar). Source: <a href=http://www.fao.org/sd/eidirect/EIan0008.htm>FAO's El Niño Primer.</a> last accessed: 29-10-2009.
Image 1: Climatic impacts of warm El Niño events (Oct-Mar). Source: FAO's El Niño Primer. last accessed: 29-10-2009.
During the northern hemisphere winter, El Niño’s expected impacts include drought (D) in southern Africa, continuing drought in northern Australia and Indonesia, high rainfall (R) in three continents and unseasonably warm weather (W) in parts of North America and eastern China.

We will look at the physics behind this phenomenon later, but teleconnections as the one described above represent a statistical relationship (i.e. correlation) from which causality cannot necessarily be inferred. Ocean surface temperatures, and most notably slight anomalies herein (+/- 1 to 3 degrees Celsius), can cause significant changes in the Southern Oscillation, the atmospheric component. The combination of these two, also referred to as El Niño/Southern Oscillation (ENSO), and it's effect on the African continent is the focus of our study. We will try to answer the question whether rainfall and vegetation growth in this continent is linked to El Niño. We structure our scientific activities following the GIScience process as proposed by Gahegan (2005) depicted in the chart below.

Scientific activities to answer: Does biomass in Africa follow El Niño?
Image 2: Scientific activities to answer: Does biomass in Africa follow El Niño?

Which variable categorizes El Niño best?

There is no single answer to this question, as there are many ways through which we can monitor the state of El Niño. Some ENSO studies place special emphasize on the ocean component of this phenomenon, however, data collection in the ocean has not been as detailed or long-running as in the atmosphere. Bouys are simply technological more challenging, and consequently also more costly to operate than regular weather stations. Instead, Troup et al. (1995) based their indicator on data collected through regular (terrestrial) surface weather stations. The southern oscillation index (SOI) resulted, available as early as the year 1882 and onwards, which is derived from normalized mean sea level pressure difference between Tahiti, French Polynesia, and Darwin, Australia. The index takes the following form (Troup et al., 1965):
Troup's SOI Calculation (1887-1989 base period). Source: <a href=http://www.longpaddock.qld.gov.au/SeasonalClimateOutlook/SouthernOscillationIndex/SOIDataFiles/index.html>The Long Paddock-Website.</a> Last accessed: 01-05-2009.
Image 3: Troup's SOI Calculation (1887-1989 base period). Source: The Long Paddock-Website. Last accessed: 01-05-2009.
A SOI of -10 means that Troup's Southern Oscillation Index is 1 standard deviation on the negative side of the long-term mean for that month. Stone et al. (1996) showed that El Niño characterization could be further improved by defining so-called "phases of SOI". A SOI Phase is determined by the change in average monthly SOI over the two previous months. This allows us to group all sequential two-month pairs of the SOI (from the year 1882 and onwards) into the following five (5) clusters: Phase 1, consistently negative; Phase 2, consistently positive; Phase 3, rapidly falling; Phase 4, rapidly rising; and Phase 5, consistently near zero. Boundaries between phases were mathematically defined using cluster analysis to further minimize errors. A comparison with other monthly SOI values can be found in Allan et al. (1996).

Model: break up data by period and El Niño event

Most agro-ecosystems models using the El Niño/Southern Oscillation (ENSO) as predictor are analogue models; a model which explains a phenomenon by reference to some other occurrence. In analogue models, the predictand and predictor are determined first, to which a close analogue is sort from historical cases (Unganai, 1998). Hereafter, scientist generally use a statistical test to prove relations between the two are indeed significant. Here, we will use a non-parametric test (Kruskal-Wallis) as our data are not normally distributed. The relating hypothesis reads as follows:

The null hypothesis (Ho) for research is that there is no systematic difference in the predictand (i.e. rainfall, NDVI) between the seasons grouped according to the different El Niño events (the predictor);
versus the alternative hypothesis (Ha), that there is systematic difference in the predictand (i.e. rainfall, NDVI) between the seasons grouped according to the different El Niño events (the predictor).

The ENSO cycle, and its effect on vegetation biomass in Africa, is best described through a combination of three components: data collected in the ocean, data collected in the atmosphere, and a terrestrial dataset collected over the African continent. All three data components are published using Distributed Data Access technology through the ITC TDS Catalog (http://thredds.itc.nl:8180/thredds/ensoCatalog.html). Finally, for each component three datasets were produced by sorting close analogue from historical cases: El Niño, La Niña, and one for Normal conditions. The resulting datasets are still large. In the ocean, they cover 30 N. Lat to 30 S. Lat, 60 W. Lon to 240 W. Lon, and 0-232 meters below sea level, with data every 2 degrees of latitude and longitude. There are 15 vertical points in the data. The atmosphere data covers the same range of latitude and longitude, but with data every 2.5 degrees of latitude and longitude. On land, the NDVI dataset geographically extends from 43.5 N. Lat to 42.1 S. Lat, 24.7 E. Lon to 64.7 W. Lon, with data every 8-km mapped using a conic, equal area projection (Albers Conical Equal Area).

Note: Distributed Data Access is a technology consisting of interconnected servers and clients, who communicate (requests for) data through the Internet. Servers provide capability to remotely sub-set data collections based on the user's spatial, temporal, and parameter requirements. Internet-bandwidth requirements are thus drastically reduced, often resulting in download sizes by a factor 10 smaller than the original. Software packages as FERRET, GrADS, matlab, IDL/ENVI, IDV, and through add-ons even packages such as ESRI ArcGIS, are examples of clients that 'understand how to access data remotely'. This feature is made possible through open-source software standards developed by Unidata (http://www.unidata.ucar.edu/) and the Open Geospatial Consortium (OGC, http://www.opengeospatial.org).

Atmospheric Data

The atmospheric data originates from the National Center for Environmental Prediction (NCEP), which has been collecting, archiving and quality controlling data for the atmosphere over the entire globe for the last 50 years. Once the data is collected, the Climate Diagnostics Center of the National Oceanic and Atmospheric Administration (NOAA) makes it available to the public via the web . Our atmospheric data is divided into three 'events': normal conditions, El Niño conditions, and La Niña conditions. To make these events from a 50-year-long span of data, we binned each year into one of the three categories, based on the Southern Oscillation Index:

To ensure a clear signal, we only considered data from December to January, when ENSO phases peak. Then, we averaged all of the years in each bin to produce three atmospheric datasets of 'typical' El Niño, La Niña, and Normal Years.

Oceanic Data

The datasets relevant to this study also contains data from the ocean. Due to scarcity of oceanographic observation (see Figure 1.3), scientists use numerical models of the ocean instead to get a more complete dataset. The models are checked against the data that oceanographers do have and then used primarily to fill in the gaps in time and space. Our ocean data were kindly provided by Dr. Peter Gent (NCAR Climate and Global Dynamics Division) and were generated from the ocean component of the Climate System Model (CSM) (http://www.cgd.ucar.edu/csm/index.html) run by the National Center for Atmospheric Research (NCAR). As with the atmospheric data, the ocean data is binned and averaged to produce a 'typical' El Niño, La Niña ocean dataset.

 Normal Ocean Temperatures, Vertical Cross-Section Pacific Ocean, 3D iso-surface at 29 degrees celsius
Image 4: Normal Ocean Temperatures, Vertical Cross-Section Pacific Ocean, 3D iso-surface at 29 degrees celsius
 Ocean Temperatures, Vertical Cross-Section Pacific Ocean, from top-to-bottom: El Niño, Normal, and La Niña conditions
Image 5: Ocean Temperatures, Vertical Cross-Section Pacific Ocean, from top-to-bottom: El Niño, Normal, and La Niña conditions

Terrestrial Data

Assuming NDVI is an accurate indicator of vegetation greenness and biomass, the NDVI dataset (http://iridl.ldeo.columbia.edu/SOURCES/.USGS/.ADDS/.NDVI/) we use for modelling the terrestrial component is based on satellite sensor observation from the Advanced Very High Resolution Radiometers (AVHRR) instrument aboard the National Oceanic and Atmospheric Administration (NOAA). This dataset was re-processed as described in Tucker et al. (2005) and originates from FEWS/Africa Data Dissemination Service (ADDS) located in Harare, Zimbabwe. The same dataset is mirrored at the IRI/LDEO Climate Data Library (http://iridl.ldeo.columbia.edu) of the International Research Institute for Climate and Society. The same Data Library holds one more datasets (predictand) relevant to our study, notably estimated rainfall as described in Love et al. (2004), both geographically covering the whole of Africa:

To remotely access these data sources, the data access URLs takes the following form (NDVI Dekadal Maximum 1981-2004):

dods://iridl.ldeo.columbia.edu/SOURCES/.USGS/.ADDS/.NDVI/.NDVIg/.dekadal/.maximum/dods
NDVI Dekadal Maximum 2004-now:
dods://iridl.ldeo.columbia.edu/SOURCES/.USGS/.ADDS/.NDVI/.NDVIrg/.dekadal/.maximum/dods
For the RFE imagery the data access url takes the following form:
dods://iridl.ldeo.columbia.edu/SOURCES/.NOAA/.NCEP/.CPC/.FEWS/.Africa/.DAILY/.ARC/.daily/.est_prcp/dods
Note: the dods:// prefix (instead of http://) which signals the OPeNDAP data server a DODS-data object is being requested instead of the metadata (text). If no error occurs in accessing and subsetting the data, the server responds by returning a data object to the client. Try the same url, but now with the http:// prefix and observe the difference.

Now, having NDVI imagery for a 28-year-long timespan (since December 1981, when the first NOAA/AVHRR sensor was launched) we again binned each year into one of the three El Niño categories as with the atmospheric/oceanic components:

Contrasting to the first two components, for the terrestrial dataset we extend the end of the period from January to April, so that December - April results. This to ensure a clear signal results given the time-lag between the moment ENSO peaks and for corresponding rainfall (and vegetation growth) anomalies over Africa to manifest. Note: 1982 here represents December 1981 through April 1982, etc.

Is there systematic difference in NDVI between El Niño, Normal, and La Niña years?

Results indicate that variability of vegetation greenness for Africa is evidently the result of climatic trends: in addition to the anticipated seasonal trends ('breathing of Africa'), we identify signals of interannual variability. The most readily identified is one that periodically affects Southern Africa. It is shown by a simple difference in NDVI between typical El Niño and La Niña events that in this region red color-tones dominate suggesting that the temporal loadings for this region exhibit a very strong relationship with the ENSO. However, we also detect that large portions of the continent do not exhibit a consistent ENSO relationship, notably Madagascar and the Sahel.
 NDVI Difference (El Niño vs. La Niña years), Map-View Africa. Red indicates less biomass during typical El Niño events (and more during typical La Niña events), green indicates no biomass difference between El Niño vs. La Niña events, and blue indicates more biomass during typical El Niño events (and less during typical La Niña events). White indicates data gaps; absence of data (missing) can be either due to waterbody, cloudcover or other (unknown) reasons.
Image 6: NDVI Difference (El Niño vs. La Niña years), Map-View Africa. Red indicates less biomass during typical El Niño events (and more during typical La Niña events), green indicates no biomass difference between El Niño vs. La Niña events, and blue indicates more biomass during typical El Niño events (and less during typical La Niña events). White indicates data gaps; absence of data (missing) can be either due to waterbody, cloudcover or other (unknown) reasons.
 NDVI Difference (El Niño vs. La Niña years), Cross-Section East Africa (transect depicted in previous figure, North-South is here plotted from left-right).
Image 7: NDVI Difference (El Niño vs. La Niña years), Cross-Section East Africa (transect depicted in previous figure, North-South is here plotted from left-right).

Do rainfall anomalies coincide with vegetation status?

Rainfall is an important, but certainly not the only source of moisture to support plant growth. After water is lost in transpiration it could well be replenished by uptake from the soil through other hydrological mechanisms, such as lateral water flow through the soil. Normally this factor is ignored as its contribution is relatively small, but uptake can also come from water (vapour) flow through the upper boundary of the rooting zone and flow through the lower boundary of the rooting zone. Tree species have been reported (Obakeng, 2007) which are capable of developing roots to depths of more than 70 m in order to reach deep sources of moisture in water-scarce ecosystems such as the Kalahari desert. To rule out the possibility that the differences observed in NDVI as described above can not be attributed to ENSO, we will look closer at whether seasonal rainfall totals also show systematic difference for this teleconnection. For two sites, one located in the northern part (left) and one in the southern part (right) of the above NDVI transect, we plot the spread of seasonal rainfall totals within and between each SOI Phase (Phase 1, consistently negative; Phase 2, consistently positive; Phase 3, rapidly falling; Phase 4, rapidly rising; and Phase 5, consistently near zero).
Box & Whisker Diagram for Karoi, Zimbabwe.
Image 8: Box & Whisker Diagram for Karoi, Zimbabwe.
As confirmed by the chart, during a typical El Niño-year (SOI Phase 1, consistently negative) rainfall over Northern Zimbabwe (Karoi) is below-normal vs; a La Nina-year (SOI Phase 2 , consistently positive) which shows rainfall is above-normal. In contrast, north from the Equator, now for Naivasha (Kenya), things are reverse as can be seen from a similair chart.
Box & Whisker Diagram for Naivasha, Kenya
Image 9: Box & Whisker Diagram for Naivasha, Kenya

Are the patterns that emerge from the visualizations persuasive?

From the above visualizations we now have confidence that, at least for some regions in Africa, ENSO is related to vegetation biomass. But visualizations are still subject to interpretation, and not free from bias, be it on the part of the producer or the consumer of the chart. For example, if we would select another site in the NDVI transect for further analysis, a less positive picture emerges.
Box & Whisker Diagram for Narok, Kenya.
Image 10: Box & Whisker Diagram for Narok, Kenya.
Based on this visualization, one observer may argue that the small, but apparent difference in medians (thick, horizontal line in the center of each box-plot) are proof of a relationship, while another observer may equally convincingly argue that the spread between SOI Phases overlaps too much to warrant such a statement. As stated earlier, generally scientists also use statistical tests to prove relationships are not subject to 'chance' since we often use small (representative) samples and not the whole population to prove a hypothesis. To provide hard(er) evidence, we here deploy a non-parametric H-test (Kruskal-Wallace) to answere our earlier formulated hypothesis.
Hypothesis test results for Karoi, Zimbabwe.
Image 11: Hypothesis test results for Karoi, Zimbabwe.
The high probability of 99.2% leads us to believe that our initial conclusion in favor of the predictive strength of El Niño, based on visualization alone, is indeed correct. However, further analysis is required for consensus-building. Note that the disadvantage of a non-parametric test such as the one deployed here is that only measures of central tendencies of populations, i.e. medians, are compared. The practical value is therefore limited, since forecast figures should be taken from the whole population to be more objective, taking into account the intra-distribution also. Thus, a (even high) statistical correlation does not necessarily imply that the prediction error of a forecast model based on this premise is also low.

Also for the two sites in Kenya (Naivasha and Narok) interpretations from the data visualizations seem to be confirmed by the statistical test, and results for Narok remain negative.

Hypothesis test results for Naivasha, Kenya.
Image 12: Hypothesis test results for Naivasha, Kenya.
Hypothesis test results for Narok, Kenya.
Image 13: Hypothesis test results for Narok, Kenya.

Talking about predictive (GIS) models, eventually the 'real' proof of any concept lies in its adaptation in society. This is the case for ENSO in South-East Africa, such so that even major seed trade companies see stock values fluctuate due to (un)founded fears for another El Niño. For example, de Jager et al. (1998) based their weather forecast component of a calibrated CERES maize model on this same ENSO teleconnection, with the intention to forecast the extent and severity of drought in maize in the Free State Province of South Africa one month before the growing season starts. The high correlation value (r²=0.86) for simulated vs. actual maize yields is an encouraging sign. Because every month for every year can be placed into a particular analogue of months, weather parameters other than rainfall can also be placed together to take out daily values (i.e. temperature, solar radiation, etc.). These data can then be used as input for a crop growth simulation model, a procedure made possible by the Agro-meteorological Rainfall Analysis and Forecast model (ARAF) as developed by Venus (1999, 2000). Based on this model he proved that roughly 75% of the rainfall variability and a corresponding 80% of maize yield variability could be explained for Mashona-Land West Province, Zimbabwe. Further justification of this technique can be found in Meinke and Hammer (1998) who demonstrated that highly significant differences in peanut yields in Australia exist among seasons grouped according to the SOI phases of Stone (1996). Despite being data demanding, El Niño/Southern Oscillation (ENSO) based agro-ecosystems predictions remain useful, i.e. for preliminary estimations of regional agricultural productivition-levels, also since they are available well in advance of actual conditions to allow for adaptive management.

References

Allan R.J., Lindesay J.A., Parker D.E., 1996. El Niño-southern oscillation and climatic variability. CSIRO Publishing, Melbourne pp 405.

Anyamba, A., Eastman, J. R., 1996. Interannual variability of NDVI over Africa and its relation to El Niño/Southern Oscillation, International Journal of Remote Sensing, Volume 17, Issue 13 September 1996 , pages 2533 - 2548.

Gahegan, M. 2005. ‘Beyond tools: visual support fot the entire process of GIScience’, in Exploring Geovisualization, ed. by Dykes, J., MacEachren, A. M. and Kraak, M. J., pp. 8399, Elsevier, Amsterdam.

Love, T.B., Kunar, V., Xie, P., and Thiaw, W., 2004. A 20-Year Daily Africa Precipitation Climatology Using Satellite And Gauge Data, 2004 AMS Conference on Applied Climatology.

Meinke, H., and Hammer, G.L., 1998. Forecasting Regional Crop Production and Climatic Risk During SOI Phases; a Case Study for the Australian Peanut Industry Australian Journal of Agricultural Research, Australia.

Obakeng, T.O., 2007. Soil moisture dynamics and evapotranspiration at the fringe of the Botswana Kalahari with emphasis on deep rooting vegetation. PhD Thesis, Library of ITC, Enschede, The Netherlands.

QDPI, 1995. AUSTRALIAN RAINMAN Version 2.1 Users Guide. QueensTerrestrial Department of Natural Resources and the Department of Primary Industries, Australia.

Skidmore, A.K., 1987. Predicting bushfire activity in Australia from El Niño/Southern Oscillation events. Australian Forestry, 50(4): 231-235

Stone, R.C., Hammer, G., and Marcussen, T., 1996. Prediction of Global Rainfall Probabilities Using Phases of the Southern Oscillation Index Nature 384, pp. 252-255.

Tucker, C. J., J. E. Pinzon, M. E. Brown, D. A. Slayback, E. W. Pak, R. Mahoney, E. F. Vermote, and N. E. Saleous, 2005. An Extended AVHRR 8-km NDVI Data Set Compatible with MODIS and SPOT Vegetation NDVI Data. International Journal of Remote Sensing.

Troup, A.J. Quart. J., 1965. Roy. Meteor. Soc. 91, pp. 490-506.

Unganai S.L., 1998. Seasonal Climate Forecast for Farm Management, In Proceedings of the Training Zimbabwe Meteorological Services, Ministry of Transport, and Energy, Harare, Zimbabwe.

Venus, 1999. Proceedings of the Seasonal Climate Forecast for Farm Management, Zimbabwe Meteorological Services, Ministry of Transport, and Energy, Harare, Zimbabwe

Venus, 2000. Towards A Framework For Maize Yield Forecasting; The Case of Mashona-Land West Province, Zimbabwe. MSc Thesis, Library of ITC, Enschede, The Netherlands.

ZIMMET, 1998. Proceedings of the Post Season National Climate Stakeholders Seminar, Apr. 1998, Zimbabwe Meteorological Services, Ministry of Transport, and Energy, Harare, Zimbabwe.

Zucchini, W., and Adamson, P.J., 1984. The Occurrence and Severity of Droughts in South Africa. WRC Report No. 91/1/84. Publisher unknown.

 


Previous: Forecasting regional food/fiber production levels Next: Download page Table of contents Index Glossary Images Frames Français Español Portugese Italian Japanese Korean Russian Chinese-simp Chinese-trad Dutch German Greek Français Español Portugese Italian Japanese Korean Russian Chinese-simp Chinese-trad Dutch German Greek

U09-NRM-127: The role of Distributed Data Access Technologies in NRM - for ITC-IDV version 2.7 > Thematic Expert Models > Food security > Forecasting regional food/fiber production levels