Below are a few of the considerations and questions that have come up during the ramp planning, design, and development processes. Click on any question to learn more. Also, feel free to submit your own questions by clicking the button at the bottom of the page. We would love to hear from you!
Make machine learning methods usable by in-country users for rapid mapping of building footprints at scale.
Many teams at large companies have done large ML building footprint projects and have released building footprints datasets over large AOIs. In most cases this was done as an exploration of modern computer vision machine learning methods. These teams have written technical papers about their solutions, but for the most part have not documented their experiences or made their training data available.
Building footprint datasets also age quickly, even if they are high-accuracy to begin with, so the static building footprint datasets produced by large corporate projects have time-limited utility. Users of building footprints for microplanning are likely to find that highly accurate building model data is unavailable over their regions of interest. Good ML solutions to this problem are ‘out there’, but work needs to be done to make them usable by people who need the output data.
The question that RAMP set out to answer is this: Can we make data, code, and instructions available to end users with limited resources who want to use ML methods to create rough-and-ready building datasets at scale, to do microplanning or respond to humanitarian crises? And can we do it in such a way that users do not need to be extensively trained in machine learning and modern ML tooling?
We: (a) created materials to guide users through the process of setting up the project, creating and curating training datasets, training a building footprint extraction model, and using it to produce building footprints over an AOI; (b) created a building footprint extraction model, such that it can be specialized (i.e., ‘localized’) to perform well on the users’ AOI with a manageable amount of effort and training data; (c) documented the project, so that users understand what the process involves, what outputs to expect, and are guided through the many decisions they will need to make; (d) provided additional tools for common tasks, such as review and editing of georeferenced images and matching labels.
We recommend using imagery with .5 meter or better resolution.
Users would need additional imagery over area for finetuning, ability to generate high-quality labels using ramp label guidance and previously labeled image chips, ability to take ramp model and retrain with additional training data.
With WHO partnership, the focus for ramp so far has been digital microplanning for health outcomes. That said, there are many use cases that ramp can support beyond health and plenty of demand from those user communities too. Many governments, for example, are interested in urban flood resiliency; there’s been increasing demand for building footprints data for renewable energy projects. Census and more accurate population data remains a huge priority for many users too.
Providing access to high-resolution imagery is not in the scope of this project. However, ramp has leveraged the incredible existing repositories of open imagery and as a result, users will have access to the ramp labels along with the source image. There are many sources of open imagery (Maxar’s Open Data Program, for example) that are good sources of free and open high-resolution
The most-often used (and most restrictive) license for ramp outputs will be Creative Commons 4.0 BY NC since our primary imagery source, Maxar’s Open Data Program, carries that license. There is a smaller portion of ramp’s training data that is fully open, including commercial use, and that is noted with the specific dataset. Those labels are using SpaceNet and carry a broader use license.
Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use.
Non-Commercial — You may not use the material for commercial purposes.
No additional restrictions — You may not apply legal terms or technological measures that legally restrict others from doing anything the license permits.
In the future, we hope to create a more robust digital curriculum with offline content accompanied by an in-person training series. The training series would be designed according to the user type – data scientist, GIS analyst, manager/decisionmaker. There will also be technical assistance available via the ramp team. We plan to further tailor the skills development to our partners’ needs – to that end, we welcome any and all feedback about how your teams learn best. Even better if there are existing efforts we can integrate into (e.g., learning series, pilots, hackathons, etc) and build more adoption around the ramp data and tools.
The Replicable AI for Microplanning (ramp) project produced an open-source deep learning model that accurately digitizes buildings in low-and-middle-income countries using satellite imagery. It enables in-country users to build their own deep learning models for their regions of interest.
The model and resulting buildings data support a wide variety of humanitarian use cases. For this project, the team focused on responding to global health emergencies and designed our approach in partnership with the World Health Organization’s GIS Centre for Health and our Advisory Council.
The project ran from October 2021 to September 2022. Project contributions were released along the way, with everything being released by September.
The primary outputs are the model codebase and documentation. Beyond the model, the team released our data labels, label review tools, training data quality check methods, user personas, signed distance transform masks, and other geospatial tools.
The approach of using satellite imagery and artificial intelligence to extract building footprints is not new. Ramp’s contribution is to open and democratize access to the deep learning model itself. The project leverages many other open-source efforts (all of which will be documented in detail in our publications and releases) like Maxar’s Open Data Program and very high-resolution imagery and SpaceNet’s open building labels for various cities globally.
Our mission is all about democratizing access to technologies like artificial intelligence and high-resolution imagery. Since our business model is based on delivering best-in-class services with the most impact, open-source projects like ramp are right in our wheelhouse.