The GO platform aims to alert the IFRC membership to imminent events and new emergencies. Using a mix of field reports and disaster assessment information from the Red Cross and Red Crescent network, as well as a curated set of external data sources, we aim to enhance situational awareness and provide actionable and trusted information for our network.
Currently, GO displays alerts and key disaster impact estimates from the Global Disaster Alert and Coordination System (GDACS), UN World Food Programme (WFP), and the Pacific Disaster Centre (PDC) on risk watch pages at country, regional and global levels. Receiving and displaying information from these three sources has posed a number of questions including:
- Are the sources all referring to the same event?
- If so, why are the estimated impacts (e.g., number of people affected) different?
- Can we trust the estimated figures, and which source has the most accurate estimates? What are our means of verification?
- Is it better to have a single or multiple estimates? How do we account for different estimates from national authorities? Are they different? If so, why?
This blog describes our approach to answer some of these questions, and the way forward to continue to improve the alerting service we provide to our network.
Are the sources all referring to the same event?
The short answer to this, is that we don’t always know.
Up to now, there has been no one comprehensive source and there is no unique ID to tie estimates from different sources and databases together. GLIDE was set up to provide this, however it is manually generated and has not been applied systematically or comprehensively by the humanitarian community. EM-DAT provides a common reference ID for major emergencies, however this is only applied later, and it captures a relatively small percentage of disasters due to its rigorous inclusion criteria. The Common Alerting Protocol (CAP) is gaining traction but is not global in coverage and does not apply to all types of events. DesInventar, despite capturing data from a huge number of events, has been discontinued in most countries where it was implemented.
Retrospectively, we’ve been trying to address this issue through what is called ‘event pairing’. Essentially, we’ve applied rules to understand if different records refer to the same event in the million plus records in Montandon — the Global Crisis Data Bank. First we applied these rules manually to 7,000 records, scouring through the available meta-data and additional contextual information and research and iteratively improving our logic. We’re now scaling the application of these rules, taking into account various biases encountered, through the use of machine learning. The technical details behind this approach were shared in the Montandon Technical Working Group meeting in September.
It isn’t (yet? ever?) perfect. In order to continue improving we’re collaborating with MapAction and the OCHA Centre for Humanitarian Data to improve the event pairing methodology. We are also learning and adapting as we receive and integrate more national disaster databases into the Montandon.
If so, why are the estimated figures (for impact and populations exposed) different?
Sometimes, we do know that alerts are referring to the same event. However the estimated figures from the different sources are different.
Let’s take an ‘easier’ example; that of an alert for an incoming tropical storm. Typhoon Yinxing, known in the locally as Typhoon Marce, was a powerful tropical cyclone that impacted the Philippines before later affecting Vietnam in early November 2024 (link). As described on the Q3 GO briefing, we received alerts from our partners PDC and GDACS, displaying the following information differently:
- Cone of uncertainty
- Alert rating
- Estimated figures for people affected or exposed
Some of the reasons for these discrepancies are known, rational and explainable. They are related to where the data originates from, the use of different models to analyse and estimate impacts, and the different ways chosen to present the information for different audiences.
Some are unknown, due to lack of methodological transparency and appropriate technical infrastructure. Alerting sources do not share all the code and assumptions behind their calculations, including for reasons of protecting their intellectual property. Others do not share these details as there is no technical means to do so easily.
We aim to address the lack of transparency and infrastructure, potentially allowing us to eventually harmonise the information shared. First, we are building an infrastructure which will allow methods and data to be shared and widely understood in a transparent manner. This will all be enabled, with the help of our partners Development Seed, Data Friendly Space and ToggleCorp, through the use of the Spatio-Temporal Asset Catalogue (STAC) specification, a common language to describe geospatial information, so it can more easily be worked with, indexed, and discovered.
By doing so, we will be able to know whether we are able to compare apples with apples.
Can we trust the estimated figures? What are our means of verification? Are they different for different hazards and locations?
In short, we don’t know.
We do know that it is better to be generally correct, than precisely wrong. By providing transparency around the limitations, errors, omissions and bias for each estimate, we can account for the uncertainty in downstream decision-making systems and processes.
However, we’ve been stuck in a position where we were unable to improve, since we lacked a system to record and evaluate performance over time. Up to now, there haven’t even been any universally accepted means of evaluating the estimates made, and by so doing understanding which estimate has proven to be most reliable in which context, for which hazard and locations.
This is a recognised problem, and one that has held back improvements in early warning systems across the globe — but it’s one IFRC and our partners have begun to address.
Is it better to have a single or multiple estimates? How do we account for different estimates from national authorities? Are they different? If so, why?
We are looking for more signal, less noise. However, we’re not aiming for a one-number-fits-all approach.
Access to multiple sources of information can be useful. Comparing sources helps us to triangulate; combining sources to produce ensemble forecasts allows us to estimate a range of uncertainty; and we should always be open to disruption from a novel technique, perhaps emerging from rapidly advancing Large Language Model and/or Machine Learning based approaches. However, in an emergency setting, it helps to have consensus and clarity to allow for quick decision-making.
We are aiming to complement, not replace, authoritative sources. Indeed, since our National Societies are auxiliary to their national governments, we truly understand the key trusted role of duty-bearing authorities to issue alerts. Our National Red Cross and Red Crescent Societies are often tasked with communicating these messages, as well as adding key information so that communities can act appropriately.
Our interest is in improving the alerts issued by pooling our collective approaches and intelligence. We can only do this at scale by using modern computing power, methodological transparency and an iterative, performance-driven mindset. That way we can begin to understand the reasons behind different estimates, in order to improve those over time, so that those responsible, and at-risk, have access to the best available information.
How would we use this information?
This information would potentially help to trigger early action, enable rapid funding, prioritise interventions, initiate readiness checks, and deploy surge personnel and assets where required.
Humanitarian information requirements include a common means to produce estimates of affected population and their humanitarian needs, including likely preferences for interventions based on sector (e.g. shelter) and modality (e.g. cash), system and services disruption and disaggregated impacts on specific populations. This common analysis should be served to as detailed a spatial granularity as possible, and updated in a predictable manner.
So what next? And what role will the IFRC play?
IFRC’s vision is to host the largest archive of structured data about current and historical disasters worldwide. This will enable analysis to reveal patterns at various spatial and temporal resolutions. Montandon — the Global Crisis Data Bank (‘Monty’, for short) will be the foundation for forecast models and systems and a dynamic database constantly reflecting the humanitarian community’s approaches.
Over the last year, the IFRC have built several components of the Montandon, including a data schema, data processing, transformation scripts, as well as a proof-of-concept API. Now, the IFRC GO team will improve upon Montandon components to operationalize the project for use within IFRC and the wider humanitarian community. Find more detail on our plans to operationalise the Montandon here.
In discussions with partners in the past months, we are starting to find consensus that this is a core set of problems to address. Indeed, IFRC are playing a leading role in many related initiatives, including Early Warnings for All, and the INFORM Warning technical working group. We have also been working with scientific partners, such as those gathered at the workshop hosted by EconAI in Barcelona in May 2024, under the title “Growing Together: Prediction, Prevention and Preparedness”, from which an academic paper will shortly be shared.
One of the next steps will be to build on a recent UNDAC Assessment and Analysis Cell workshop to convene the key partners providing early warning alerts for the humanitarian community (including WMO, PDC, WFP, and GDACS), alongside those who use that information for action (UN OCHA, FAO, MapAction etc) to tackle these, and many more detailed questions. Please get in touch if you want to collaborate — im@ifrc.org
