A strategic collaboration to help close the protection gap:
Since 2016, Rain++ has supported the Microinsurance Risk Organisation (MiCRO) in the design and operational implementation of index-based, inclusive insurance products to cover the underserved population against natural disasters, exacerbated by climate change. Targeted at small farmers and business owners, with this type of insurance pay-outs are automatically triggered when a predetermined index (e.g. rainfall accumulation during previous days) is exceeded, without the need to assess actual losses, as is required with traditional insurance products. The automatically triggered compensation helps the beneficiaries get back on their feet right after a natural disaster occurs, while also making the overall mechanism more objective and transparent.
As of 2021, MiCRO’s index-based insurance products against excess rainfall, drought and earthquake are available in El Salvador, Guatemala and Colombia. Work is currently underway to launch similar products in Mexico.
Our role:
Rain++ has primarily supported the design of the excess rainfall and drought insurance products. This has entailed extensive analysis of satellite rainfall and vegetation index products across the countries of interest, including:
- Quality assessment of multiple satellite products to inform product selection,
- Enhancement of rainfall estimates through combination of satellite estimates and data from ground stations, and
- Statistical analysis and modelling -including stochastic time series generation and extreme value analysis- to support MiCRO’s actuaries in defining a fair and sustainable pricing structure.
Aside from product design, Rain++ has also supported the implementation of the computational engine behind MiCRO’s calculation platform, which extracts data from relevant sources to detect and notify triggering events.
What was done?
We worked alongside actuaries from the client organisations on the design of much needed drought weather-index insurance products for Ethiopia and Madagascar, based on satellite-based, ground and climate re-analysis data. Targeted at small farmers and business owners, these simple and objective insurance products aim to cover the underserved population against natural disasters, exacerbated by climate change.
Context and local challenges:
A particular challenge faced in this project was the scarcity of ground information to assess satellite and reanalysis climatological data and to validate models. Limited data was available for Ethiopia, while in Madagascar ground information suitable for insurance product design was virtually non-existing.
Another challenge had to do with rocky soil conditions in Ethiopia, which impact the accuracy and usability of commonly used satellite-based drought indicators, such as EVI (Enhanced Vegetation Index).
Additionally, in the case of Madagascar, the local practice of planting seeds several times a year, to ‘hedge’ against increased weather volatility, made it harder to establish specific seeding periods to be used as reference for drought product design.
Our strategy to data assessment and selection:
To address the data scarcity challenge and identify climate variables and products which can well replicate local drought conditions and which can be used for subsequent insurance product design, we undertook a correlation and time lag analysis between different precipitation, soil moisture and vegetation index datasets. This was expected to provide insights into wet/dry patterns leading to drought conditions and into the ability of the different products to reproduce historical droughts.
Four precipitation products were included in the analysis, namely IMERG, ARC, TAMSAT and ECMWF’S ERA5. With regards to ground conditions, ESA’s CCI soil moisture product and MODIS’ Enhanced Vegetation Index (EVI) were considered in the analysis.
On the whole, the IMERG and ERA5 precipitation products displayed the best performance, with IMERG performing best at daily scale and ERA5 at monthly scale. These two datasets were then used for insurance product design
Insurance product design:
The design of index-based insurance products entails statistical analysis of relevant climatological variables -with a focus on extremes- to ultimately inform pricing structure. That is, the thresholds at which pay-outs are triggered.
In this phase and based on the selected precipitation products, we used Generalised Extreme Value (GEV) analysis and Standardised Precipitation Index (SPI) for drought characterisation. We worked closely alongside the clients’ actuaries to ensure that an effective, fair and sustainable pricing structure was obtained, accounting for the aforementioned local challenges.
Project overview:
The European FloodCitiSense project focused on improving cities’ resilience to flooding through co-development of innovative monitoring and early warning systems by citizens, researchers and local water managers. Part of the ERA-NET Smart Urban Futures funding initiative, the project included academic, industrial and governmental partners from the UK, Belgium, the Netherlands, Austria and Taiwan.
Urban living labs were setup in three pilot cities: Birmingham (UK), Brussels (Belgium) and Rotterdam (the Netherlands). Working with local citizen groups and following co-creation approaches, three main activities were carried out:
- Networks of citizen-operated low-cost rain sensors (piezoelectric disdrometers) were implemented,
- A mobile app and accompanying web platform were developed which allow citizens to report flooding in their local area, visualise rainfall and flood reports logged by other participants, and receive flood warnings, and
- Data-driven urban pluvial flood forecasting systems were developed using data and models currently available at each of the pilots. In the case of Birmingham, two modelling approaches were tested. The first approach assumes rainfall forecasts are available and initially categorises them as flood-inducing or not using eigenfaces. Afterwards, the spatial distribution of floods linked to a flood-inducing rainfall forecast is predicted using logistic regression. The second approach consists of a two-stage (weather + flood forecasting) analogue system. An overview of these approaches, including test results, and those implemented in other pilot cities is given in this presentation.
Our role:
Besides taking active part in the co-creation activities at the UK pilot, Rain++ staff were responsible for:
- Providing technical support in the installation and operation of low-cost rainfall sensors,
- Deploying a network of LoRa nodes for data transmission across the Birmingham pilot,
- Implementing an IstSOS database for storing rainfall data from low-cost sensors and other local rain gauge networks across the three pilot cities,
- Supporting the development of the web platform, and
- Providing scientific inputs for the development of Birmingham’s data-driven weather and flood forecasting system.
Lessons learnt:
Through the FloodCitiSense project we learned about the benefits of co-creation approaches and citizen observatories (e.g. improved understanding, acceptability and ownership of solutions) and the challenges that these approaches entail (e.g. data reliability, technical problems with low-cost sensors, difficulty of long-term stakeholder engagement).
Furthermore, we were able to explore the feasibility of implementing data-driven flood forecasting models based upon currently available data and models. Amongst other things, we learned that typically available flood records may be insufficient for training data-driven models (records are generally short, sparse, and often of inconsistent quality). Instead, where available, hydraulic models can be used to generate flooding scenarios which can then be used to train data-driven models.
Relevant links: A project infographic is available from this link and the full project website is available here.
What was done?
An analogue rainfall forecasting model which integrates numerical weather prediction (NWP) and satellite data was developed and deployed operationally to support the operation of the Nam Ngiep 1 hydropower plant in Laos. The forecasting system provides 1-hour rainfall forecasts during the first 48h, and 3-hour rainfall forecasts thereafter and up to 7 days into the future. This work was undertaken as sub-contractors for Artelia Group.
Modelling and implementation details:
In the proposed precipitation model, global NWP is the driving force and post-processing techniques are used to improve the quality of and downscale NWP outputs for catchment-scale applications. The NWP product adopted in the precipitation model was the GFS, which was found to outperform NVGEM.
Two post-processing techniques were considered and evaluated, namely frequency distribution matching and analogue downscaling, with the former displaying an overall superior performance. Different satellite products were tested for use as reference for real-time calibration of the post-processing routines. These included IMERG Early, GSMaP and GSMaP now. Ultimately, IMERG was adopted on grounds of quality.
The Delft-FEWS platform was adopted for operational deployment of the rainfall forecasting system in 2019. Training on the use of the operational system was delivered to staff members of the Nam Ngiep 1 Power Company, who are now in charge of its operation.
Why stochastic rainfall generators?
Hydrological consultants and water companies require sub-hourly rainfall data for operational use as well as for the design of components of urban drainage networks. Where no measurements are available, plausible time-series of rainfall depths can be generated using stochastic rainfall generators. These time-series can then be used as inputs to hydrological and hydraulic models, either in continuous mode or by picking out events of interest.
The TSRSim Tool:
TSRSim is a stochastic rainfall generator, which produces rainfall time series suitable for urban drainage assessments. It has been successfully validated using historical observations covering the whole of the UK. It comprises two components: an hourly generator and a disaggregation methodology that further downscales the hourly rainfall depths to the scale of five-minutes.
What’s new?
This project addressed the need to update the hourly generator in the light of the latest research at Imperial College London and Rain++. Specifically, the generator is now able to produce more realistic cells whose intensities are typically inversely proportional to their duration, while avoiding the generation of unrealistically intense cells. Additionally, by drawing upon the latest research in the Department of Statistical Science at University College London, the calibration of the model was made more effective.
References:
A detailed description and validation of the improvements made to the TSRSim tool can be found in this paper: ONOF, C. & WANG L. P. (2020). Modelling rainfall with a Bartlett–Lewis process: new developments. Hydrology and Earth System Sciences, 24(5), 2791-2815.
Other relevant publications:
- Kaczmarska, J., Isham., V., Onof, C. (2014) Point process models for fine-resolution rainfall, Hydrological Sciences Journal, 59(11), 1972-1991.
- Verhoest, N.E.C., Vandenberghe, S., Cabus, P., Onof, C., Meca-Figueras, T., Jameleddine, S. (2010) Are stochastic point rainfall models able to preserve extreme flood statistics? Hydrological Processes, 24(23), 3439-3445.
- Cockcroft, J. (2006) Using TSRSim in Flooding Solution Development. In WaPUG Meeting, Blackpool, UK.
What was needed?
The ultimate aim of this project was to develop a localised and computationally simple radar adjustment methodology to improve radar accuracy for the real-time rainfall-based services provided by Meniscus across the UK, such as rainfall-sensitive cycling route optimisation.
How did we do it?
The re-calibration strategy was informed by a preliminary radar QPE (quantitative precipitation estimates) assessment which entailed comparison against ground rain gauge records for selected winter and summer storm events. This assessment revealed a tendency of radar QPEs to underestimate high rainfall rates, such as those observed during convective events. Moreover, the discrepancies between radar and rain gauge records were found to be non-linear and to vary across events, rendering a simple 1st statistical order model, such as mean field bias correction, insufficient for adjusting radar records.
In light of this and considering application requirements and operational constraints, different radar re-calibration strategies were tested, including dynamic and static Z-R (radar reflectivity to rain rate) function re-calibration. Both re-calibration strategies were shown to improve the quality of radar QPEs, with the dynamic adjustment generally leading to best results. However, dynamic re-calibration is conditional on good quality rain gauge records being available in near real-time. The two strategies can be used inter-changeably, depending on operational conditions.
