Why is disaster-risk data difficult to collect?
MapAction works with governments and humanitarian organisations on disaster response, preparedness and risk analysis. A key part of that work is bringing together information on hazards, vulnerability and the ability of communities and institutions to respond.
For projects using the INFORM Subnational Risk Index, that information needs to be available below national level — for example, by region, province or district. INFORM combines information on hazards such as floods or earthquakes with indicators of vulnerability, including poverty and disease prevalence, and measures of coping capacity.
The difficulty is often not that the information does not exist. It is that it may be buried inside long PDF reports rather than stored in spreadsheets or geographic datasets.
Finding a table inside a report, copying the relevant values and checking which administrative area each number belongs to can take considerable time. When a risk model draws on many reports and indicators, that manual work can also limit how many sources a team is able to use.
Why does better data extraction matter for disaster planning?
A risk model is only as complete as the information that goes into it.
If useful local indicators remain difficult to access, analysts may have a less detailed picture of where hazards, vulnerability and coping capacity are concentrated. That can affect later decisions about preparedness, anticipatory action and where further analysis or resources may be needed.
The project therefore addresses an earlier part of the decision process: turning existing information into structured data that risk models and decision-makers can actually use. By reducing the effort required to extract data from reports, MapAction aimed to examine more sources and build more complete subnational risk datasets.
How did MapAction, ABW and Pipple approach the problem?
In 2024, MapAction, Analytics for a Better World and Pipple developed a pilot focused on information extraction from PDF reports.
The goal was practical: a user should be able to provide a PDF and receive one or more spreadsheets containing relevant indicators by administrative area, together with supporting metadata. The tool also needed to work across multiple languages and be available as a Python script.
The collaboration tested several open-source approaches for finding and extracting tables. Standard tools performed well on simple tables but struggled when reports used less consistent layouts.
The team ultimately selected GMFT, an open-source toolkit for extracting tables from PDFs, and adapted it for the types of reports MapAction works with.
Pipple data scientist Sanne van den Bogaart played a key role in developing the tool alongside MapAction and ABW.
MapAction at an INFORM Workshop
What did the pilot produce?
The project produced an open-source Python tool designed to convert information that would otherwise need to be collected manually from PDF reports into structured spreadsheets.
That creates three potential operational gains: reducing the time required for data collection, allowing teams to examine a larger number of reports and improving the completeness of the data used in risk models.
The pilot phase of the project focused on developing a working extraction tool and integrating it into MapAction's next stage of testing.
The scripts and user instructions were also made openly available so that other teams can test and adapt the approach.
What can this mean for future disaster-risk planning?
The first planned large-scale application was MapAction's work on the Southern African Development Community's regional INFORM Subnational Risk Index.
The SADC region includes 16 countries, more than 360 million people and over 200 first-level administrative areas. Building a risk index at that scale requires bringing together information from many national reports, making automated extraction particularly relevant.
That work involves the SADC Disaster Risk Reduction unit, MapAction, GIZ and UNDP, building on previous INFORM projects in countries including Eswatini and Madagascar.
More broadly, the approach could be useful wherever important public or humanitarian data exists but remains difficult to reuse because it is stored in long, inconsistently formatted documents. The next step is to test how reliably the tool performs across more reports, languages and table formats — and whether the time saved allows risk teams to build more complete evidence for preparedness decisions.
