Early assessment of forage availability for An ASSET Protection Insurance scheme
EARLY ASSESSMENT OF FORAGE
AVAILABILITY FOR AN ASSET
PROTECTION INSURANCE SCHEME
Anton Vrielinga, Michele Meronib, Andrew Mudec, Sommarat Chantaratd,
Caroline Ummenhofere, Kees de Biea
aUniversity of Twente, Enschede, The Netherlands
bEuropean Commission, Joint Research Centre, Ispra (VA), Italy
cInternational Livestock Research Institute, Nairobi, Kenya
dThe Australian National University, Canberra, Australia
eWoods Hole Oceanographic Institution, Woods Hole (MA), USA
12 June 2015 – ADRAS workshop – Nairobi, Kenya
11:40 – 12:10
CONTENT
NDVI time series
Basics
Evolving of processing sequence
Temporal integration period of NDVI
determine unit-level start- and end- of season from NDVI series
evaluate earlier predictability of index
Conclusions
Challenges
INDICES RELATED TO SEASONAL FORAGE SCARCITY
Rainfall
Station-data limited
Many satellite-derived RFEs, but accuracy for area?
Vegetation indices
NDVI (but also others like EVI, fAPAR)
a real measurement, available from many satellites
Alternatives products exist:
soil moisture
evapotranspiration (from LST)
temporary water bodies
Not only what to use, also how to use it!*
* See also: de Leeuw, J., Vrieling, A., Shee, A., Atzberger, C., Hadgu, K., Biradar, C., Keah, H., Turvey, C.,
2014. The potential and uptake of remote sensing in insurance: a review. Remote Sensing 6, 10888-10912.
THE DATA (1): NDVI A SPECTRAL INDEX
NIR
red
indicator of the presence of photosynthetically-active
green vegetation
THE DATA (2): COMPOSITING + SMOOTHING
17-24 May 2015
Compositing: select ‘best’ pixel
over time period
Smoothing: reduce remaining
atmospheric effects
use pixel time series of composites
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54
averageNDVI(-)
eMODIS dekad number
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
FORAGE SCARCITY INDEX (1): FROM NDVI TO INDEX
GOAL: indicator of seasonal forage availability within an insurance
unit, relative to ‘normal’ availability
normalization
temporal aggregation
1-10 June 2011 1-10 June
(2001-2014)
1-10 June 2011
Z-score
spatial aggregation
FORAGE SCARCITY INDEX (2): ORIGINAL APPROACH *
1-10 May 2011
NDVI image (10 day) Z-score (compare to 2001-2014) spatial averaging Z-score
Temporal averaging over season Seasonal index
* Chantarat, S., Mude, A.G., Barrett, C.B., Carter, M.R., 2013. Designing index-based livestock
insurance for managing asset risk in northern Kenya. Journal of Risk and Insurance 80, 205-237.
FORAGE SCARCITY INDEX (3): NEW APPROACH *
1-10 May 2011
NDVI image (10 day) NDVI aggregated Temporal averaging
Seasonal average NDVI Z-scoring to get seasonal index
* Vrieling, A., Meroni, M., Shee, A., Mude, A.G., Woodard, J., de Bie, C.A.J.M., Rembold, F., 2014. Historical extension of operational NDVI
products for livestock insurance in Kenya. International Journal of Applied Earth Observation and Geoinformation 28, 238-251.
OBJECTIVE
To better identify the temporal integration period for IBLI’s forage
scarcity index
phenological analysis
predictability of end-of-season variability
Asset replacement vs asset protection
early assessment early payout
Now:
LRLD: March – September
SRSD: October – February (payout ideally 1 month later)
Temporal averaging
ARE LRLD/SRSD GOOD SEASONAL DESCRIPTORS?
Index overall forage conditions for season
NDVI describes green vegetation ≠ all forage
But forage in dry season has been green!
focus on green biomass build-up only
Phenological analysis to determine end of season
PHENOLOGICAL ANALYSIS
For more detail, do not hesitate to ask.
Phenological analysis by Michele Meroni.
Adapted version as described in: Meroni M, MM Verstraete, F Rembold, F Urbano, and F Kayitakire. 2014. A phenology-based method to derive
biomass production anomaly for food security monitoring in the Horn of Africa. International Journal of Remote Sensing, 35: 2472-2492.
CAN WE PREDICT END-OF-SEASON VARIABILITY BEFORE?
Take as reference identified start/end
Calculate cross-validated R2
Example for Central Wajir (ID=96) but numbers are fictive
WHEN DO WE EXPLAIN 90% OF SEASON VARIABILITY?
end long long: 90%
end short short: 90%
CONCLUSIONS
Phenological analysis provides better seasonal definitions
Forage scarcity index relates to when forage is developing
Insurance payments can be made 1-3 month earlier
considering also season predictability
accounting for NDVI filtering (rainy season likely more clouds)
depends on insurance unit
Earlier payment may allow protection livestock
purchase of forage, water, medicines
CHALLENGES
divisions in Turkana with very limited variability
in-depth understanding of greenness variation on livestock:
Full assessment of reduction basis risk reference data
Plot biomass measurements / time lapse photography / crowdsourcing
Livestock mortality data / MUAC …
Drought recalls / weather stations / tree rings …
effect of previous season on
livestock mortality
within-season distribution:
importance?
Is this relationship location-
dependent (or livestock-type
dependent)? 0
0.1
0.2
0.3
0.4
0.5
0.6
1-10Mar
11-20Mar
21-31Mar
1-10Apr
11-20Apr
21-30Apr
1-10May
11-20May
21-31May
1-10Jun
11-20Jun
21-30Jun
1-10Jul
11-20Jul
21-31Jul
1-10Aug
11-20Aug
21-31Aug
1-10Sep
11-20Sep
21-30Sep
averageNDVI(-)
Yabello (Ethiopia)
average (2001-2014)
2011
2014
LEARN MORE ABOUT NDVI SERIES AND ANOMALIES?
New FAO E-learning course “Remotely Sensed Information for
Crop Monitoring and Food Security – Techniques and methods for
arid and semi-arid areas”
Lessons 4, 5, 6 by me
http://www.fao.org/elearning/#/elc/en/course/FRS
14-YEAR VARIABILITY PER DIVISION
evaluate when R2>0.90
use R2
cv instead of R2
fraction explained in prediction
Example for Central Wajir (ID=96)
Phenology retrieval method
Screening and retrieval of number of GS per year
Time series
> 60 % valid obs
flag missing
n
y
(95th - 5th)
percentile difference
> FAPAR uncertainty?
(0.1 units)
not vegetated
n
y
2 GS per year
Ratio of Lomb normalized
periodogram power spectrum
at 1 and 2 cycles
mono-modal bi-modal
≥6 <6
1 GS per year
Phenology retrieval method
Retrieval of pixel “climatology”: setting of the temporal breakpoints
that likely separate the periodic climatic cycles in the time-series
Time series
Find minima of
smoothed “median year” *
*iteratively smooth until n. of min = n. of
GS per year
Set the cycle and
sub-cycle breakpoints
Breakpoints are set independently on the actual existence of
a specific season allowing to detect complete season failure
Phenology retrieval method
Model to be fitted
𝑃𝐷𝐻𝑇 𝑡 = 𝑎0 +
1
2
𝑎1 𝑡𝑎𝑛ℎ 𝑡 − 𝑎2 𝑎3 + 1 +
1
2
𝑎4 𝑡𝑎𝑛ℎ 𝑡 − 𝑎5 𝑎6 + 1 − 𝑎4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 10 20 30
fAPAR
dekads
grow
dec
PDHT(t)
Parametric Double Hyperbolic Tangent Model: fusion of two ‘S-shaped’
function representative of a typical seasonal signal (7 parameters)
Phenology retrieval method
Fit the model on the upper envelope of observations
For every expected
cycle, fit the model
to observations if the
season is not failed
Phenology retrieval method
Extract phenology on the fitted model
GSL
Peak
CFAPAR
Metric Definition
SF Complete season failure if 95th - 5th percentile for that season < 0.05
SOS
Timing of the start of the
growth phase
when modelled season exceeds 20% of local growing amplitude
EOS
Timing of the end of the
decay phase
when modelled season drops below 80% of local decay amplitude
GSL
Length of the growing
season
EOS-SOS
MaxV
Maximum (peak) value of
FAPAR
Max(modelled season)
CFAPAR
Cumulative value of
FAPAR during the period
of plant activity
Integral of the fitted model between SOS and EOS. This
indicator is proportional to the total GPP.
Seasonal GPP proxy, 𝑪𝑭𝑨𝑷𝑨𝑹 =
𝒔𝒐𝒔
𝒆𝒐𝒔
𝑭𝑨𝑷𝑨𝑹 𝒅𝒕