I2UD
The International Institute for Urban Development (I2UD) AI Climate Platform
Vision: Use machine learning (ML) to map vulnerability to potential flooding hazards, landslides, and land values for communities in the Global South through the AI Climate platform.
Project Overview
Data Challenge
Optimize data gathering processes with partner consortium, Data cleaning and harmonization
Data needs
Access to high quality satellite imagery, Tests for predictive accuracy of models, community validation of insights generated
Data practices
Predictive Modeling, GIS Mapping, Citizen Science/Crowdsourcing
target users
Policy makers in cities lacking localized hazard risk data
tech team
I2UD Project TEAM
I2UD had a team of 2 on the project:
Technical partner Dymaxion Labs (Argentina) had a team of 4 working on the project:
hypothesis
Spatial data’s increasing abundance and decreasing processing costs will allow the team to use ML to cut the time and expense involved in hazard mapping, assessment, and prediction — subsequently freeing funds to build climate resilience in cities and communities in a scalable way.
What is spatial or ‘geospatial’ data? >
results:
I2UD incorporated four layers of predictive results and carried out local validation during the season of highest flooding and landslide risk. This work generally confirmed hypothesis that higher resolution would yield improved results.
What are data layers? >
“The purpose of the tool is to be agile and inexpensive, not highly accurate. Its underlying approach is the 80/20 Pareto Principle, aiming at the best possible result using the least necessary information, given that the last mile to high levels of accuracy can be very costly. All software involved is open source, as a means to ensure that the platform is widely accessible to the large number of cities and communities that lack the resources to carry out regular and comprehensive hazard identification. The resulting layers are presented in an easy-to-use viewer.”
Introduction
Identifying the Problem . . . FLOODS, Displacement, Destruction . . .
As extreme weather events become more frequent, the risk of flooding, displacement, and destruction increases in communities around the globe. Particularly vulnerable are marginalized communities in low and lower-middle income countries that are not formally recognized by governments, often settled in areas on tenuous land, on unstable hillsides, or too close to flood-prone riverbanks. These risks are dynamic in space and time as communities move and change the built environment around them. Local government officials need to be able to monitor and predict potential threats to communities quickly, to understand how land values are impacted, so they are ultimately better equipped to intervene and protect the public in the event of a disaster in time to mitigate its worst impacts.
Gathering Data to Predict Hazards . . .
Gathering the data needed to understand and predict hazards in vulnerable areas is not straightforward. Earth observation techniques from modern satellite technology can provide a significant boost to decision makers via high resolution geospatial data, but this data is not sufficient. Ground truth data on what is actually happening in communities where people live and how people relate to the space around them is necessary for improving risk maps and enhancing predictive accuracy of when and where disaster might strike.
Assembling the Team . . .
The team at the International Institute for Urban Development, along with a consortium of partners from international aid groups and civil society enlisted a technical partner to create a dynamic tool, AI Climate, to help the local government in the Sula Valley, Honduras understand and adapt its urban planning and build resilient communities in the face of escalating climate risks.
READ More:
HOW IT Started:
“When this idea hatched in 2018, AI was still an unproven technology”
The Institute has always been technology oriented, there has always been someone inside the system who’s a computer instruction person, and it has always thought about how to look at technology and its potential use for urban planning.
“When we started the institute, Google Earth didn’t exist. We only had access to imagery from the US government. A lot of new technology has emerged in the past decade, particularly from the proliferation of satellites monitoring the earth’s surface by taking images with regularity, via private companies. We began to gain a better understanding of how we might use the tech for climate and urban settlements. It used to be very expensive to prepare base maps, and learn what’s going on.”
“We took a course on coursera on deep learning and wondered, how can we use this to advance our work? …We determined that we could use it to start mapping places where we couldn’t get in.”
The vice president of the institute brought forward the idea for this work that built on the use of satellite imagery data. She enlisted partners at Dymaxion Labs because they had been on the leading edge of work in AI and GIS, using remote sensing to detect informal settlements in other countries in Latin America. Dymaxion were positioned to be ideal partners in Honduras given their prior work and would benefit from speaking a common language: Spanish.
The first difficulty was to convince the founders of I2UD that it was a worthwhile project. The institute had been losing traditional consulting projects, and it was generally accepted that they needed to move toward more startup and entrepreneurial work.
“We were motivated to lean into innovative processes, and this seemed like a perfect opportunity to do so in the supportive environment of the Patrick J. McGovern Foundation’s Data Practice Accelerator program”
Data Journey
From a technical point of view, the team characterized the project as a data experiment to test three elements of the overall hypothesis of AI Climate.
1. Construction
First, the ability to construct a machine learning process for the desired features of AI Climate (The name of their platform)
2. Feature development
Second, the feasibility of using machine learning to develop these features effectively, and generate the desired information
3. assessment
Third, the usefulness of the machine learning results for people on the ground
I2UD/DYMAXION Tactical Roadmap

Data Collection Requirements
Translate planning outcomes into specific data needs.
The most critical need was ground truth data, which is necessary to construct a supervised machine learning process we envisaged.
Identify the annotations on satellite imagery or other input data, so that they can presume to be correlated with the items to be recognized.
Approach
data tools
before:
Google suite, GIS data layers on flooding, landslides, and land degradation built using ML
After:
Cloudera, Supervised ML, Python, Tensorflow, QGIS, Cesium, AWS (satprog, unetseg)
data Fluency
The team worked alongside a trusted technical partner, Dymaxion Labs, for the technical aspects of the project, and learned a significant amount about standardized data collection and data cleaning within a broad consortium of partners
“They recognized a need – to make data more accessible, available, consumable for decision makers. Network of collaborators, advisors, implementers covered everything from communities, municipalities, etc. Subject matter expertise. The approach that they had was so highly collaborative. Their data journey was not just about technical aspects.”
— Chelsey Walden-Schreiner, Data Scientist, Data Practice at PJMF and foundation technical lead for this project
Lessons Learned
Developing a data field guide and catalog
Partners confronted the difficulty of defining and explaining the data needs adequately, the reluctance of public officials to share data, and the absence of different kinds of data.”
- In categorizing informal settlements, there is a whole spectrum. The I2UD team had to define many things in writing, had intensive conversations, and shared the documentation so that actions were clear to all.
- From images, they tried to optimize so that the imagery were not too dark or too light to identify important features in the landscape and built environment
- I2UD spent time reformatting training data for models due to mixed GIS experience/skill”
The Data Field Guide
The team, together with Dymaxion, defined a set of guidelines for the preparation and transmission of data by the local partners:
Placing emphasis on data validation & ground truthing
“This aspect is very critical and was embedded in design from beginning”
- Validation not only improves results, but it also builds trust. “Working collaboratively, from the initiation of project, we wanted to overcome the black box issue – not just saying that ‘the machine says you’re in harm’s way so you need to move’”
- The team showed the results to people from the community and asked ‘does this make sense? Does it resonate with your experience on the ground?’, if not, they would work to improve results.
- Conceptually, there was a two way flow of information. Dymaxion sent model outputs, followed by validation by experts or stakeholders.
- Strong tropical storms required partners GOAL and Habitat to support communities in Tegucigalpa and Sula Valley. This emergency delayed local validation, and as a result not all of the local validation feedback could not be incorporated into the final recalculation of algorithms.
Non-profits can and should ask for satellite imagery donations, as well as considering needs for spatiotemporal resolution
- Publicly available Sentinel 2 satellite imagery isn’t highly resolved enough for the work that the team was doing
Learn about the resolution of satellite imagery for different applications from Maxar > - The Patrick J. McGovern Foundation team helped broker an introduction to Planet labs, a commercial high-resolution satellite imagery provider, which helped I2UD secure satellite imagery at a more granular resolution.
Learn more about Planet Labs >
“The team weighed tradeoffs and balances between data sizes. They had already trained a model on sentinel data with a 10m resolution, but needed this to be closer to 1m. The temporal needs (1 image per year or 1 per month) and the computation requirements are important questions to consider in building a model like this”
— Chelsey Walden-Schreiner, Data Scientist, Data Practice at PJMF and foundation technical lead for this project
Relying on network of established, trusted local partners, and maintaining regular touchpoints with these partners
“This process knit together a variety of organizations, from suppliers of satellite images to residents of informal settlements. Our approach was to build AI Climate in close partnership with Dymaxion Labs and in collaboration with technical experts and three local partners”
- Amid lots of partner consortium complexity, it helped that all of them shared a common language: Spanish
- The most important element of the project was a series of periodic meetings across the partnership, including McGovern Foundation technical leads via Zoom, and sometimes on WhatsApp. When a quick response was necessary, this was how to do it.
- The organizational structure of the project integrated all participants along a vertical axis of co-production, from the sourcing of images to validation of results with local communities in Honduras.
Partner Groups and Roles:
- IHCIT, with strong technical knowledge of weather and climate patterns, flooding, and landslides, particularly in the Tegucigalpa region.
- GOAL Honduras, with a focus on informal urbanization and risk reduction, and a strong presence in informal communities in both Tegucigalpa and Sula Valley.
- Habitat for Humanity Honduras, with a strong presence across the country, and a growing strategic interest in understanding the risks faced by low-income urban communities.
- Dymaxion had a specific person to handle the information differences who played a critically important role in data cleaning
The variety of models and scenarios generated made overall interpretation of results more complex, this necessitates intentional user education
The partner consortium hosted a forum in the Sula valley, Honduras with mayors and private sector, brought people together to introduce the AI Climate platform to show its potential, and took advantage of being there to have a workshop where dymaxion helped to explain the technical aspects of the platform
project
SustainabilitY
Critical issues remain which the project could not completely address given the nine-month Accelerator grant period. The main challenge is the lack of institutional arrangement that could ensure the continuation, feeding of information, and operation and maintenance of AI Climate.
Contact The Project Teams:
Future Direction
The team continues to work on fundraising for AI Climate, as I2UD is looking to apply this platform in other parts of the world – regions like sub-Saharan Africa, Central America, Central Asia, and the country Brazil.