Proba-V Symposium
Oostende
29-31 May 2018
Improved drought detection to
support crop insurance models
Kees de Bie
Ben Maathuis
Anton Vrieling
Higher basis risk
Lower
basis
risk
More
accurate / Less
error-propagation
Accumulated response to Agronomic Drought
Climatological Drought
INTRO: why NDVI and what is insured ?
Agronomic Drought
Response to Agronomic Drought
•Insures risks associated to “exposure to drought” •A specified money amount
(credit taken) is insured, NOT the value of suffered yield-loss!!
•Is area-specific (1x1km) •Is growing season specific •Is crop a-specific
•Is fully scaleable
•Suits only rainfed systems
Looking at
responses
Looking at
inputs
Never insure the output of a business model (=yield) !!
The key-problem to quantify NDVI-anomalies at pixel level and by dekad:
we always have
insufficient annual-repeats
to get robust frequency distribution estimates !
ANOMALY
mapping is mostly based
on comparing the present reading
with the ‘long-duration’ average
© Herman Eerens et al., 2014. DOI10.1016/j.envsoft.2013.10.021
Advised is to use ALL available
annual repeats BY dekad and BY
pixel.
Based on absolute deviation from
the long term average, calculated
for 1999-2010.
Insurance schemes require quantification of probabilities and not: poor, normal and good !!
Insurance companies deal in MONEY
Farmers will argue regarding COSTS and PAY-OUTS
With too few repeats, results are
often generalized, by using the
terms: poor, normal and good.
All these ‘anomalies’ refer to differences around the average. Insurance schemes require probabilities at the low-extreme !!
© FAO, JRC, ITC http://www.fao.org/eLearning/
ANOMALY
mapping is mostly based
on comparing the present reading
with the ‘long-duration’ average
The key-problem to quantify NDVI-anomalies at pixel level and by dekad:
we always have
insufficient annual-repeats
to get robust frequency distribution estimates !
How to calculate anomalies
BY dekad and BY pixel ??
The key-problem to quantify NDVI-anomalies at pixel level and by dekad:
we always have
insufficient annual-repeats
to get robust frequency distribution estimates !
AND
we have very few (or no) data at the frequency distribution
extremes
!
Any “relative” difference operator can show very unpredictable and sharply fluctuating results …(https://land.copernicus.eu/global/products/vci)
Based on 10 year data, the 20%
left-tail threshold is derived
from 2 readings only!
And the 5% left-tail is derived
from ……. none!
Do NOT assume,
for a pixel and
dekad, that it
has a normal
NDVI frequency
distribution
© FAO, JRC, ITC http://www.fao.org/eLearning/
The key-problem to quantify NDVI-anomalies at pixel level and by dekad:
we always have
insufficient annual-repeats
to get robust frequency distribution estimates !
AND
we have very few (or no) data at the frequency distribution
extremes
!
Do NOT assume,
for a pixel and
dekad, that it
has a normal
NDVI frequency
distribution
To address all specified issues, we need to use population statistics instead of sample statistics.
How do insurances
perceive DROUGHT
Is there ever a drought in the Sahara ? NO, the Sahara is perpetually dry
Is Tigray suffering from more droughts then Oromia ?
NO, the only differences relate to drought severity, otherwise, Tigray is just dryer!!
So: drought is a
negative anomaly
(deviation) from what is
normal
.
Anomalies can be measured through pre-defined percentile thresholds (by dekad):
15% 50% 5%
0%
100%
tr
ig
ge
r
ex
it
ND
VI
or a Normal Distribution ? Mean Mean Mode Mode Median Median Ri gh t sk ew ed di str ibu tio n HIGH Le ft sk ew ed di str ibu tio n LOW HIGH LOWND
VI
The key-problem to quantify NDVI-anomalies at pixel level and by dekad:
we always have
insufficient annual-repeats
to get robust frequency distribution estimates !
AND
we have very few (or no) data at the frequency distribution
extremes
!
Use of the 5 and 15 percentiles as thresholds, matches a Pure Premium rate of 10%
Pay-Out
probability
Insurance schemes require quantification of probabilities and not: poor, normal and good !!
Insurance companies deal in MONEY
Farmers will argue regarding COSTS and PAY-OUTS
All these ‘anomalies’ refer to differences around the average. Insurance schemes require probabilities at the low-extreme !!
Any “relative” difference operator can show very unpredictable and sharply fluctuating results …(https://land.copernicus.eu/global/products/vci)
To address all specified issues, we need to use population statistics instead of sample statistics.
Use of the 5 and 15 percentiles as thresholds, matches a Pure Premium rate of 10%
Once solved (space-time cube): • we do not require a long-duration
average of 30-years
• we do have enough degrees of freedom (pixel x annual repeats) to derive robust %-estimates for trigger and exit thresholds.
The key-problem to quantify NDVI-anomalies at pixel level and by dekad:
Actuaries: we cannot create
sound insurance models, based
on ≤ 30 historical records (annual
repeats) ?
- How do we define/generate a
POPULATION (a group of pixels),
from which to extract the
required percentile statistics ?
- Is using a 30 years window
representative for present-day
situations ?
Medium NDVI-Values by dekad of all 18 annual-repeats
(the reference to assess anomalies)
ETHIOPIA: from a
space-time cube to
population statistics !
Solution:
ZONATION
Results of unsupervised classification: • Pixels Zones (= Classes, Clusters) • Time Legend (with NDVI-info)
We used the ISODATA tool:
About 2,000,000 pixels were grouped into 160 clusters; each with 2,500 to 12,500 pixels.
Medium NDVI-Values by dekad of all 18 annual-repeats
(the reference to assess anomalies)
ETHIOPIA: from a
space-time cube to
population statistics !
Solution:
ZONATION
A small part of the generated legend (NDVI-Information by zone/class/cluster)
Differences between
zones relate to
differences in:
•
growing seasons,
•
land use and mix of
crops grown,
•
cropping calendars,
•
soils and terrain
•
land cover, species
composition, and
abundance (density)
•
weather and climate
•
farming systems,
•
etc.
Shown legend-items
depict various actual
gradients !
Dekadal assessments during a growing season
Pay-out starts once the actual NDVI-value drops below the preset Trigger-value.
Pay-out is 100% if the actual NDVI-value drops below the preset Exit-value.
actual NDVI-values
for 1 pixel Zonal thresholdsby dekad
The pay-out amount will be the averaged value (%) for the specific season
x
insured amount.
Indemnity amounts (0-100%)
… with full drought assessment by pixel by season: 2015 (
Early+Long
)
These maps show where ‘land’ was exposed in 2015
to agronomic drought *
… with full drought assessment by pixel by season: 2015 (
Late+Long
)
2ndmajor drought
in ONE year!!
These maps show where ‘land’ was exposed in 2015
to agronomic drought *
… with full drought assessment by pixel by season: 2016 (
Early+Long
)
2015 “after-drought” effects
These maps show where ‘land’ was exposed in 2016
to agronomic drought *
… with full drought assessment by pixel by season: 2016 (
Late+Long
)
This drought (that lasted well into 2017) caused lots
of nomadic displacements ! These maps show
where ‘land’ was exposed in 2016
to agronomic drought *
Farmer claimed yield-losses % 0 0-20 20-40 40-60 60-80 = Payout % 0 43% 6% 3% 0% 0% 0-20 20% 12% 0% 0% 0% 20-40 1% 4% 7% 0% 0% 40-60 0% 0% 1% 2% 0% 60-80 0% 0% 0% 0% 1%
Frequency counts of pay-outs (%)
versus farmer claimed yield-losses (%)
[n=190; North Wollo zone]
Only 3+1 = 4% of the farmers claims
deviated between 20-40% from our
pay-out estimates.
Most remain happy (some got money,
while not needed), and for the
remaining 1% a payout was expected
(but not received).
The acceptable accuracy
(< 20% deviation) for the surveyed
area and farmers is thus 96% !
Field validation of Insurance accuracy: Basis risk
Density plot
[many 0-0 combinations; R2=60%]
Modified from: Fetene Zerihun Minale, 2017
Two additional, independent studies, provided
… and the
highest
pay-out by pixel by season in 18 years (
Early+Long
)
The maps show where agronomic
drought was severe and potential impacts
high *
… and the
highest
pay-out by pixel by season in 18 years (
Late+Long
)
The maps show where agronomic
drought was severe and potential impacts
high *
Not all pixels in a zone experienced a drought, while some experienced several droughts;
outliers based on past facts: Red+Green (
Early+Long
)
The maps show where customer
satisfaction would have been
high (red) or low (green)*
Not all pixels in a zone experienced a drought, while some experienced several droughts;
outliers based on past facts: Red+Green (
Late+Long
)
The maps show where customer
satisfaction would have been
high (red) or low (green)*
Solved the “space-time cube”
• We can now focus on using only long-duration data that: • represents the present-day climatic status (moving window) • represents the present land conditions, land use, and farming
systems (that follows developments/trends/changes)
• We have now tested logic, concepts and tools to group pixels into clusters (classes), which allows us to derive robust trigger- and exit-threshold estimates.
• During that process, we did re-obtained (lost!) logic on ‘how to act like an actuary, representing the insurance world’
• The product proved fully scaleable, low cost, practical and accurate • Basis-risks, moral hazards, and fraud-opportunities are all minimized • Fields never need to be visited once their location is known (1km2grid)