Geospatial Data in Agriculture Risk Management | The IRI Index Insurance Experience

Dr. Daniel Osgood, Lead Scientist, Financial Instruments Sector Team 

FARMD (March 2012) | Until recently, many have doubted that index insurance could scale to the large numbers of farmers needed to meaningfully address poverty.  There has been worry about low demand as well as concern that supply is blocked by technology, data, and infrastructure limitations.

New projects have been overcoming many of these problems. To make it worthwhile for farmers to pay premiums, they are engineering insurance so that it leads to increased profit by making it safe for a farmer to invest in more productive farming methods.  These projects address liquidity constraints, use farmer driven design, and carefully craft insurance to compliment other risk management options.  Logistical delivery challenges are being addressed by leveraging existing distribution systems, such as microfinance networks, seed distributers, work for food safety net programs, and farmer cooperatives that are already delivering related products to large numbers of farmers. 

These insurance projects have overcome many barriers.  A year or so ago, the largest African index insurance projects had around a thousand farmers.  This year, multiple projects in Africa have tens of thousands of farmers buying insurance (e.g. Syngenta Foundation & the HARITA Project).

Despite these significant advances, the absence of comprehensive rainfall and crop data remains a key constraint in scaling insurance, since data is needed for index design and determining payouts.  Index insurance is not scalable if it only works in areas covered by existing rain gauges with long histories. 

Many are hoping satellites could be the solution to this “data poverty” problem.

However, satellite estimates are inaccurate and have many known failings. They cannot be applied blindly across all landscapes. For example, satellites face different problems interpreting a landscape with a lot of rocks, versus one covered by trees, or blanketed by grasses, or with a lot of water.  One side of a mountain range may be fundamentally different than the other.  Developing and validating the appropriate strategy for each location is essential so that farmers are not hurt by indexes based on bad data. This requires a more sophisticated use of geospatial data than ever before to link on-the-ground data, experiences, needs, and concerns with advanced satellite and climate model information.

At IRI, we have been working with partners on these problems.  Our insurance related efforts began in 2006, when we worked with the World Bank ARMT led (formerly CRMG) insurance project in Malawi, and performed exploratory analysis in Kenya and Tanzania (for details, see the report here) In these projects, indexes were based on data from ground measurements.  Satellite information, such as climatological evapotranspiration, was used to aid in insurance design. In addition, we mapped the ground based rainfall measurements in order to know which farms were close enough to a rain gauge to get insurance, and to look for spatial diversification based on climate processes (such as ENSO).

Later, IRI helped design indexes using satellite vegetation measurements. These products were offered by SwissRe and bought by the Millennium Villages Project (MVP) in 2007. We were uncertain if satellite vegetation measures would be adequate for indexes because satellites had well-documented problems when used to estimate yields.   In addition to the errors due to dust, clouds, solar angle and satellite angle, we also had a resolution problem: any individual pixel we looked at was a mix of farmland, bare earth, trees and grasses.  Fortunately, we were working in MVP sites, which had a relatively large amount of data for product design and validation, and teams of researchers stationed at each site, so we had the information necessary to move forward.

Ironically, the solution to our problem was not to focus on the details.  Instead it was to look at the bigger picture.  We found that the satellite indexes were not reliable when we tried to work with fine scales using high resolution images and detailed modeling.  Instead, satellite data worked well when we aggregated over a large region and applied common-sense strategies (that we validated using the detailed modeling). 

We ended up using the satellites to observe the entire landscape to figure out if rainfall had not fallen during critical times for crops.  In areas comprised mostly of grasses and crops (with few trees), we could target the drought years by tagging when everything turned “brown” a month after the crops needed water.  In hindsight, this was the strategy we should have started with, since remote sensing estimates of agricultural production are often much more reliable across large regions than at small scales.

For each village, we relied upon the on-site researchers and large amount of ground-based data to design and validate each index.  As we worked to improve accuracy, we found that indexes based only on ground measurements of rainfall missed some important loss events, while indexes based solely on satellites missed some other important loss events.  In the end, we were able to get the best targeting of loss through a regional index that combined satellite vegetation sensing with ground based rainfall measurements, so that each measure could fix the problems of the other (see p56 of the CSP2 at iri.columbia.edu/csp2 for more details). At this point, we were not comfortable with the accuracy of results for a stand-alone farmer level product. But, we were confident in our ability to target losses at larger scales, partly because since the MVP had complimentary mechanisms that would come into effect if the index did not pay out.

Next, IRI helped build a satellite rainfall indexed insurance product first purchased in 2009 by individual low-income farmers in Ethiopia. This was for the Horn of Africa Risk Transfer (HARITA) project (led by Oxfam America) – a project that has since expanded beyond East Africa and is now called R4 – Rural Resilience Initiative. For this project, instead of using vegetative satellite products, which look at what is on the ground, we used satellites that look at the tops of clouds to estimate rainfall.  As you might imagine, these rainfall estimates are educated guesses at best, so it was important for us to figure out what they could do reliably that was useful for a farmer.

Relying heavily on the strength of partners such as REST, DECSI, Nyala, AIC, the Ethiopian National Met Agency, SwissRe, and many others, we were able to invert our model for insurance design--since farmer focus groups in previous projects had provided the key insights in insurance design, we made a new process that was formally led by the farmers. We developed contracts based on the parameters farmers specified, providing them with the related satellite information and indexes and giving them formal authority to make the key decisions on moving forward. 

Insurance was only one small part of the farmer discussions.  Farmers developed a comprehensive set of risk management activities, and figured out how to better take productive risks to increase their production.  Therefore, instead of trying to make the insurance do everything, the insurance was specifically designed to provide coverage for key, important hazards that could not be dealt with otherwise.  This holistic approach made the insurance technologically feasible, because it simplified and focused the insurance problem.  We were optimistic that the satellite images of clouds could target the problem identified--flagging the worst drought years in terms of the years when the rainfall season started or ended badly.  If bad years could be accurately flagged when there were few dark clouds but the rainfall season should be underway, the satellites might be able to do the job adequately.

In the initial villages we did a lot of work to see if the satellites would be up to snuff.  In addition to obtaining as much production and rainfall data as we could, we also ran crop models, interviewed farmers and local experts, and strategically installed new automated raingauges.  Furthermore, remote sensing experts performed groundtruthing surveys and dozens of farmers recorded daily rainfall at their farms to build a combined understanding of rainfall patterns (click here for an animation of farmer rainfall data).

Of course, this level of intense study would not be possible at large scales.  So, we used these initial villages, where we were able to invest in an intensive data-gathering exercise, to develop a process to scale the project to places where we would have less information. Protocols were formed for the farmer design process in each village and software scripts were written to process the data.  We harnessed ongoing food security geospatial data efforts, such as WFP LEAP and FEWSNET to provide a starting point for our discussions in new villages. We developed a set of interacting geospatial datasets that included ground-based rainfall measurements, farmer and expert reports of agricultural timing, crops and bad years, available agricultural information, and the satellite data.  For each village, the different types of geospatial information were packaged in different forms so that the farmers, local experts, and international partners could all participate in an iterative design and data-update process. 

Because satellite vegetation products are based on completely different information than satellite rainfall products, and have very different strengths and weaknesses, we thought satellite vegetation might be a good crosscheck of how well satellite estimates of rainfall catch the  bad years.  We tested several different kinds of satellite vegetation products.  Although the satellite vegetation measures showed promise, none of the products performed very well.  Before they could be relied upon to identify major drought years they need to be improved, and further tested (see publication). 

By 2011, indexes had been purchased in 43 villages, and the design process had been performed in an additional 40 villages.  Based on this experience, the aforementioned R4 Rural Resilience Initiative, an Oxfam America, United Nations World Food Programme and SwissRe effort is planning to scale this model much more broadly (see press release). Demand has been high, and initial impact assessments are finding significant production increases in some of the villages (for more information see HARITA's quarterly report Q2 2011).

It is critical that we can get the satellite vegetation-based validation working well as we reach larger scales.  Otherwise, we will not know where the satellite rainfall index needs to be fixed.  Based on our experience so far, we have identified potential solutions, and we are moving forward with partners to get them working.  We have active projects with the UN International Labour Organization (ILO) (see website) and USAID to solve these problems, and are building upon our relationships with NASA, other space agencies and several index insurance implementation groups to tackle these challenges together.

One especially exciting breakthrough is the Google.org-funded work that the Ethiopian National Meteorology Agency has done in partnership with Reading University and IRI.  Using their expertise, knowledge of local rainfall patterns, and weather station network, the Ethiopian NMA has built a blended satellite/rain gauge gridded rainfall product of unprecedented coverage and accuracy.  The launch of this product has provided a very important source of information for the R4/HARITA insurance efforts.  Anyone online can access this advanced product through the Ethiopian NMA’s geospatial webserver.

Because of the high level of ground validation in the Ethiopian NMA product, it can make estimates using satellites that have been in orbit for decades that are more accurate than what has been done with the newest, most advanced satellites.   This is extremely valuable for insurance because estimates using old satellites may be the only option to have the decades of uniform historical data necessary for index design, pricing, and studying climate trends.  Ongoing efforts to bring this approach to other parts of Africa are likely to yield valuable geospatial data for a wide range of projects.

Moving forward, as insurance projects scale, it will be increasingly valuable to improve our geospatial capabilities. It is important that we take the care to get it right.  If we are careless, large numbers of farmers could be hurt.  From our experience, technology alone is not enough.  For success, we need data solutions that not only link raingauge and satellite datasets, but ones that allow the poorest farmers in the world to interact with the community of global researchers, and leverage in-country expertise to go beyond todays most advanced remote sensing products.  

Additional Resources from IRI