An empirical study of the links between NDVI and atmospheric variables in Africa with applications to forecasting vegetation change and precipitation Chris Funk UCSB, Climate Hazard Group Molly Brown, NASA Global Inventory Modeling and Mapping Systems

1
An empirical study of the links between NDVI and
atmospheric variables in Africa with applications
to forecasting vegetation change and
precipitationChris FunkUCSB, Climate Hazard
GroupMolly Brown,NASA Global Inventory
Modeling and Mapping Systems
E D C
1/13/2003
IRI Presentation
F E W S
2
Objective Improved FEWS NET Executive Summaries
3
Background Useful NDVI projections

• NDVI is a satellite measurement of vegetation
  used to monitor drought
• NDVI linked to locusts (Tucker, 1985 Hielkema et
  al., 1986), Malaria (Hay et al., 1998), and Rift
  Valley Fever (Linthicum et al., 1999)
• RVF in 1998/1999 cost the Greater Horn 100
  million
• NDVI is linked lagged precipitation
• (e.g. Nicholson, 1990 Potter and Brooks, 1998
  Richard and Poccard, 1998, and others)
• It seems logical to try to use lagged rainfall to
  project future NDVI
• These projections are distinct from and
  compatible with NDVI forecasts based on
  downscaled climate information
• e.g. the work of Matayo Indeje at the IRI
• Future work should look at combining approaches
• Best Skill Persistence lagged Rain Climate
  Forecast

4
Monthly Data

• NDVIe data from NASA GIMMS (1 global and 0.1
  degree African)
• GPCP rainfall rescaled to 1 degree
• Tim Loves CPC FEWS NET African Rainfall
  Climatology data (0.1 degree, Africa)
• Cold cloud duration precipitation estimates
  blended with automatic gauge data (Love et al.,
  2004)
• NCAR Reanalysis relative humidity fields
• Class C variable (Kalnay et al., 1996)

5
Talk Overview

• Empirical Models of NDVI Change
• Describe model
• Test model
• Works in most semi-arid regions
• But, only some regions have decent
  cross-validated skill when the seasonal cycle is
  removed
• However, most of Africa is explained well by
  either the seasonal cycle or the projection model
• From a decision support perspective we can tell
  people what the NDVI conditions will be a few
  months in advance
• Empirical analysis of lagged NDVI/precipitation
  relationships
• NDVI can maybe help predict precipitation in a
  few regions
• Brazil and Eastern Australia (perhaps).

6
Month ahead Max-to-Min NDVI Change Formulation
More veg ? higher evapotranspiration
Higher RH ? less evapotranspiration
We assume geographicallyvarying Nmin and Nmax
are fixed. These have been shown to be strongly
related to average precipitation, temperature and
latitude (Potter and Brooks, 1998)
Less veg ? higher rainfall efficiency
More rain ? more veg
7
Month ahead Max-to-Min NDVI Change Formulation
Growth term
Loss term
We model NDVI change
Loss stops whenwe reach historic minNDVI
Growth stops whenwe reach historic maxNDVI
Growth assumed to log-linear with precipitation
Loss assumed linearlyrelated to 100-RH
8
Zimbabwe test site revisited
9
Observed and Estimated NDVI Change for Zimbabwe
Test Site
Extremes under-estimated could consider
extending max/minbeyond historic values
10
Observed and Estimated NDVI for Zimbabwe Test Site
Extremes under-estimated could consider
extending max/min beyond historic values
Some inter-annual variability is captured
11
1-month max-to-min NDVI Change models
12
5-month max-to-min NDVI change models
Future Climatological
Averages
Past Observed Data
G1 G2 G3 G4 G5L1 L2 L3 L4 L5
FoL1
G1 G2 G3 G4 G5L1 L2 L3 L4 L5
FoL2
G1 G2 G3 G4 G5L1 L2 L3 L4 L5
FoL3
G1 G2 G3 G4 G5L1 L2 L3 L4 L5
FoL4
Climatological averages could be replaced with
forecast precipitation and relative humidity
This effort meant to complement forecasting
efforts by the IRI, CPC and others
13
Cross-validation results for Africa
R2 images contain the seasonal cycle Skill
1.0 Var(Nobs- Nest)/Var(Nobs), Michaelsen,
1987 Some drought-prone semi-arid locations
show good skill
14
Focus on 4 month forecast
Livestock dependentRVF-prone
In southern Africa low rainfall regions
predicted okay, but seasonal cycle appears
dominant In eastern Greater Horn region, good
skills found applications to pasture,
malaria and RVF feasible
15
Sample Application NE Kenya NDVI projections

• In 1997/98 an extensive outbreak of Rift Valley
  Fever occurred in northeastern Kenya
• Apx. 27,500 cases occurred in Garissa district,
  making this the largest recorded outbreak in East
  Africa, (Woods, Karpati and others, 2002)
• Early warning can allow prevention, monitoring
  and mitigation activities

16
Sample Application NE Kenya NDVI projections
Test Site
17
Sample Application NE Kenya NDVI projections
Test Site
Caveats Large area increases accuracy of
estimatesARC rainfall known to be accurate in
KenyaStill this analysis bodes well for RVF
detection
18
Conlcusions

•  We can model NDVI change in semi-arid regions
  with a simple max-to-min growth/loss formulation
• Skill levels are high in semi-arid, low in
  tropical forests and places with a strong
  seasonal cycle
• Future work will look at incorporating forecast
  information
• We hope to create integrated monitoring/projection
  information products