Innovative Digital Citizen Science: Active Learning for Disaster Relief

Citizen science has been used for a wide variety of scientific discoveries, many of which have been enabled by the technical framework and expertise of the Zooniverse citizen science platform. The project team has been involved in applying the algorithms developed to make optimal use of inputs from citizen science volunteers in disciplines ranging from astronomy to zoology; among these are projects where the crowd assesses satellite and aerial imagery following natural disasters in order to rapidly provide responders and decision-makers with accurate situational information across wide geographical areas.

This project seeks to advance this work by building on a recent STFC-supported project to develop novel integrations of citizen science classifications and machine learning algorithms. Specifically, the project will develop and test an "Active Learning" system where the machine and citizen scientists collaborate in real time to provide the most accurate assessments of infrastructure damage and signs of life, as quickly as possible. The machine is first trained using the results of previous humanitarian crowdsourcing deployments, and then during a live deployment asks the crowd for further information where it deems this to be of the most use in providing high-quality data to responders and decision-makers on the ground in affected areas.

This system will be of obvious and immediate use for future disaster relief deployments. Additionally, the nature of the well-developed Zooniverse platform means that the software developed here can be applied to other disciplines where researchers need to be able to accurately detect changes in data. These include astrophysics, where researchers examine the changing sky for exploding stars and transiting exoplanets, and ecology and conservation, where researchers actively study animal migration and detect human activities such as poaching.

We propose to augment the change-detection abilities of the Zooniverse platform, specifically by building on the results of an ongoing STFC-funded programme to integrate machine learning techniques into the classification pipeline. The current programme is testing the use of transfer learning in our disaster relief citizen science project, the Planetary Response Network (PRN). A small additional extension to the effort on this project will enable us to integrate the results of the project into an Active Learning framework which we will develop and test during this Citizen Science Exploration project. We aim to use past PRN data for development and test the system on the Zooniverse either in a real-time deployment situation or a live simulation using data in hand, depending on whether a new disaster occurs during the project period and in either case clearly communicating the situation to our citizen science volunteers.

The current work is testing two different implementations of a Regional Convolutional Neural Network (Faster R-CNN and Mask R-CNN), using both expert-labelled and segmented images as well as crowdsourced labels from past deployments of the PRN to train object detection networks. During the proposed project we will choose an optimal means of uncertainty assessment so that the Active Learning framework can identify which images will be of most use when shown to citizen scientists for additional classification. We will then implement, test and deploy this framework. Integration of the algorithm with the Zooniverse classification platform will make use of the open-source Caesar back-end software package. Caesar is well-tested and has previously been used to allow other ML algorithms to incorporate citizen science classifications. Given the wide applicability of change-detection software across multiple disciplines and the generalisability of the Zooniverse software we expect the results of this project to have long-term multi-disciplinary impact.

The research described in this proposal seeks to directly improve the ability of citizen scientists to assist first responders through the Planetary Response Network. The increase in efficiency we expect from this work will allow us to produce more useful results, much faster than is currently possible. We therefore expect this work to lead to citizen science approaches to disaster response being adopted more widely, and to greater impact when they are used.

In addition, because the work proposed uses standard Zooniverse tools, it will be a pathfinder for other Zooniverse projects; at present, the work using active learning has not been generalised. Participation in Zooniverse citizen science projects has been shown to lead to volunteers seeking more scientific experiences within the domain they are working on. Participation in Planetary Response Network projects, even in a 'test mode' as envisaged in this proposal, will likely lead to an engaged volunteer base who are incentivised to learn more about issues around disaster preparation and relief.