PulseSatellite 1

PulseSatellite

PulseSatellite is a collaborative web-based tool that combines cutting edge artificial intelligence with human expertise to extract the most relevant information from satellite imagery for use in humanitarian contexts.

About PulseSatellite

Every day, millions of images are captured from space by an ever-growing number of satellites. In humanitarian crises or conflict areas, accurate and timely satellite image analysis is key to supporting critical operations on the ground.

Use cases include: monitoring population displacement, settlement mapping, damage assessment,  flood assessment and identifying the direct impact of earthquakes, volcanoes, cyclones and landslides.

Until recently, this type of analysis was done by human analysts who spent hours in front of a map manually counting and classifying structures and other elements.

PulseSatellite is a tool to analyze satellite imagery assisted by neural networks that seeks to incorporate humans-in-the-loop at different stages in the model inference process to enable optimal results and expert validation.

Created by

Phase 1

We opened up the beta version of the tool to UN agencies and the interest we received was overwhelming. 

We’re currently evaluating submissions of UN partners whose projects we hope will yield interesting results and good lessons learned.

As we work to raise more funding, we plan to make the tool available to other humanitarian practitioners. 

PulseSatellite 4

See PulseSatellite in Action

PulseSatellite 5

Use Case: Mapping Structures

This video is a deep dive into how PulseSatellite can be used to count settlements in refugee camps. 

Current Capabilities

Settlement Mapping

Identifying and counting structures in refugee settlements

Roof Density

Inferring the size and proximity of roofs for slum mapping

Flood Mapping

Producing real-time maps of flooded areas for damage assessment

Settlement Mapping

Counting and classifying structures in a refugee settlement is a common analysis task for humanitarian agencies. In practice, this is currently manually done by human expert analysts using satellite imagery. A single settlement may have tens of thousands of structures, and identifying each of them can take several days.

For camp mapping, we have a Mask R-CNN model (He et al. 2017) trained on images from 12 settlements, where the image of each has been split into 300×300 pixel tiles and annotated by human experts. Once the model has been run on an unseen camp, the analyst can inspect the result in the tool and correct the outputs on a subset of tiles. An adaptation stage can then be performed to fine-tune the model to the unseen image and increase performance.

In recent independent tests, camp completion rates increased from 77.3% to 94.7% after adaptation, with a final user accuracy of 94.4% -in line with humanitarian performance requirements described in (Quinn et al. 2018).

Before After

Roof Density

Improvised shelters, often known as ‘slums’, found on the outskirts of cities around the world are largely unmapped and misunderstood.

We have developed a methodology similar to that used in flood detection whereby we train an image segmentation model to identify pixels in the image belonging to slum dwellings. Current models have been trained on slums in Uganda in collaboration with Pulse Lab Kampala; however, due to lack of similarity between slum regions, models for different slums will have to be developed.

With better mapping of slum dwellings, we hope to help UNOSAT, and other agencies, to more rapidly assess complex shelter coverage while enabling better and more timely care to be given to those in need.

Before After

Flood Mapping

Rapid analysis of regions affected by natural disasters is essential for response planning. Current methods for rapid flood mapping require an analyst to perform a series of manual pixel intensity thresholding and image processing operations s, taking several hours to be completed.

The PulseSatellite tool uses a U-Net model (Ronneberger, Fischer, and Brox 2015) to produce a flood map in the order of minutes. The model has been trained and tested on images from Bangladesh and Somalia achieving overall accuracy scores of over 90%, in line with current manual methods.

What Users Say

“PulseSatellite is an exciting example of a system combining AI and human experience to greatly improve the use of satellite imagery in the UN system. By employing artificial intelligence and a great and intuitive interface, the power of satellite imagery is much more accessible to UN colleagues.

Lars Bromley, Geospatial Specialist, UNOSAT

“PulseSatellite is a promising tool that will allow UN agencies to use their own models while including the analyst during on-the-fly training and evaluation. It will also help to scale the deployment phase providing a user-friendly interface.”

Edoardo Nemni, Machine Learning Researcher, UNOSAT

“We wanted to build a web-based tool that is tailored to the needs of UN agencies and that is easy to use. The collaborative nature of the system is what hopefully makes it invaluable, both because it allows teams across time zones and geographies to work on the same project, and because it encourages teams to upload and use their own models, as well as share them with others.

Tomaz Logar, Data Engineer, UN Global Pulse and PulseSatellite Project Lead

FAQ

UN Global Pulse and UNOSAT have collaborated on the project since 2017 to support the work of UNOSAT.
Detection in camps has different levels of accuracy varying from 65 to 99. A key aspect of the tool is the ability to incorporate user feedback to retrain AI models on-the-fly to increase accuracy over specific settlements until the quality of the detection is high enough. You can find more about evaluation in our paper, "Humanitarian applications of machine learning with remote-sensing data: review and case study in refugee settlement mapping". As the network of users grows, we will have models adapted to different situations.
Our plan is to open PulseSatellite up to as many UN Entities and humanitarian practitioners as possible. For now, the beta version is being tested with a select number of UN agencies. We will let you know once we open the tool up for further testing. If you have any questions, write to us using the Get in Touch form below.
PulseSatellite is a distributed cloud application, currently running on Amazon Web Services (AWS). The web-based user interface (UI) is written in JavaScript and HTML, and the web server, which communicates with the UI via HTTP and WS protocols, is written in NodeJS. The database is managed using PostgreSQL with the PostGIS extension and is currently also used as a message streaming service for Pub/Sub messaging between the web, GPU and tiling servers. All backend files are generated and stored on an Amazon Elastic File System (EFS). Multiple GPU servers can run ML processes across a range of frameworks including PyTorch, TensorFlow and Keras. GIS processes, including the tiling servers, use OSGeo libraries and programs, most notably GDAL. All used software is open source.
We love open source as much as the next person. We plan to open source the core engine of our tool and share the AI models with international and humanitarian organizations.
Yes we are. At this stage, we are looking for funding from the donor community (Member States, UN Entities, foundations) to help scale the technical solution and its adoption. Reach out to us at info@unglobalpulse.org if you are interested in supporting us.

Get in Touch

If you are from a UN Entity and would like to try the tool (once we open stage 2 of testing in the upcoming months), if you are a donor interested in taking the tool to the next level, or if you are an AI expert wanting to improve existing or add new models, please get in touch.

Thanks to Our Donors

PulseSatellite 8
PulseSatellite 9
Scroll to Top