Simplifying the Identification of School Infrastructure Vulnerability at Scale
Using AI algorithms and photographic images from school buildings in Nepal and Kyrgyzstan, two international public universities collaborated with a team from the World Bank to develop a technical solution to the long-standing problem of identifying the most vulnerable school building infrastructures in hard-to-reach areas of developing countries. With this solution, an estimated 875 million children and teachers at risk of being injured can be better protected from natural disaster harms.
The California Polytechnic State University, San Luis Obispo (Cal Poly DxHub) and Munich University of Applied Sciences (MUAS DTLab), both supported by Amazon Web Services (AWS), collaborated with the World Bank’s Global Program for Safer Schools (GPSS) to design a process that could save time and money in determining the structural type of school buildings to assess their vulnerability. Worldwide, natural disasters like earthquakes and cyclones put more than 1,000,000 school buildings in low- and middle-income countries at risk of collapsing, and an estimated 875 million children and educators at risk of being injured. Yet gathering the necessary baseline information to prioritize risk mitigation investments is a time-consuming, expensive task that requires experts to travel and inspect school infrastructure in remote areas.
One of the biggest challenges in identifying at-risk building structures is the lack of high-quality data about school building inventory, and the absence of efficient mechanisms to update and manage this information. The data collection process to assess school infrastructure is commonly done through field inspections conducted by engineers, which are usually costly and time-consuming. Therefore, innovative and more efficient approaches to collect baseline data are essential to strengthen the capacity of developing countries to scale up safer school activities
GPSS seeks to improve the safety of schools by strategically prioritizing investments based on factors such as building vulnerability and the region’s earthquake hazard level to which they are exposed. However, there’s a lack of high-quality data about building structural characteristics and no efficient way to analyse and manage this data once it’s collected. The process of labelling, managing, and assessing the photographic data is also time-consuming, expensive, and hard to do across tens of thousands of schools.
The California Polytechnic State University’s DxHub and Munich University of Applied Science’s DTLab work directly with governments and other public sector organizations, designing free, open-source solutions. Their goal is to utilize the deep subject matter expertise of the public sector, the technical and innovative expertise of a cloud technology company, and the diverse disciplinary knowledge existing across universities to bring innovative solutions to challenging public sector problems. Mirroring the real-world, these solutions typically require the involvement of experts with different subject matter expertise and perspectives.
In collaboration, these public sector organizations designed and demonstrated a mobile application that guides school administrators and other community members through photographic data collection. The photos are uploaded to the cloud where an algorithm determines the building category, height range and main structural system. The results are remotely reviewed by a trained engineer for accuracy. From there, the data are aggregated and provided to planners and decision-makers to prioritize risk mitigation investments quickly.
The goal of this data collection effort is to establish a set of standardized photos that are compatible with AI algorithm inputs. The teams worked together to demonstrate the use of AI and ML methods towards a more effective way of assessing school buildings in areas difficult to access by the World Bank evaluators.
While the teams worked on these solutions in their courses, the Cal Poly DxHub supported the collaboration with cloud technology credits and resources, as well as technical and project management support. In addition, the DxHub, in collaboration with the faculty member teaching these courses, developed an international student exchange opportunity with MUAS, a partner university in Europe. The purpose of the exchange was to provide students with the additional opportunity to collaborate internationally and consider how varying international contexts and perspectives could improve the development of this international solution. U.S.-based students were to travel to Europe and work with a group of international peers to further develop the initial implementations by the HCI and AI/ML teams towards a deployable solution within a Software Development class. Unfortunately, days before a delegation of four students and one advisor was to leave for a trip to Europe to facilitate the handover of the project, COVID-19 restrictions led to a cancellation of the trip. Fortunately, this collaboration remains intact, and this next step in the work will be resumed once travel restrictions are fully lifted.
What Makes Your Project Innovative?
Prior to this innovative response, engineers were deployed into remote areas to assess school building infrastructure. Even before the travel restrictions imposed by the global COVID-19 pandemic, this led to gaps in response time and put children and their educators at great and prolonged risk. This project empowers communities by providing them with the tools to share their own school building imagery, so that they may receive the financial and structural resources needed to ensure school buildings are safe for children and educators.
This project is innovative both technologically and socially/ethically as it relies on advanced technologies to simplify the identification of structural abnormalities that put children and educators at risk, while empowering the people served through technological training and community agency.
What is the current status of your innovation?
Since the development of this initial prototype, Cal Poly faculty and students have continued to further develop the machine-learning models by adding structural taxonomy classifications to the models, which will improve the resolution of the assessment, and by focusing upon improving the accuracy of the models overall. Furthermore, the GPSS team will advance the students’ work to create a data management, annotation, and analytics tool that will leverage machine learning. The solution will ingest data collected from the field and apply classification models for validation by a trained engineer. The engineer will act as a ‘human in the loop’ by further training the machine learning models to improve the algorithms and classifications incrementally. By enabling data analysis on a large scale, rapid and cost-effective assessments can be made. This solution supports a data-driven prioritization of investment and will improve student, teacher and staff safety.
Collaborations & Partnerships
While this solution was first deployed in the contexts of Nepal and Kyrgyz Republic, the goal is to scale globally. The Cal Poly DxHub, MUAS DTLab and World Bank will continue to improve this solution through additional cross-governmental and institutional collaborations, and external and internal funding mechanisms. An added benefit of this work is the international workforce development opportunities for students, but also the governments and public sector organizations.
Users, Stakeholders & Beneficiaries
• School infrastructure managers
• Local structural/earthquake engineers
End users are governments/agencies such as:
• Ministry of Education
• Ministry of Emergency Situations
• State Agency for Architecture, Construction, and Communal Services
• Ministry of Construction
Results, Outcomes & Impacts
The tool described here requires two steps: first, collecting sets of key photos of school buildings, and second, implementing the AI solution to process the collected information and develop the classification. It is envisioned that community members without an engineering background are involved in the data collection phase by taking key photos of the school buildings. Subsequently, the data would be processed by the AI with supervision of local trained technical teams to identify the structural typologies and assess their vulnerabilities. Utilizing this framework could enable governments to characterize large sets of data in a more efficient and sustainable way, as a baseline to make informed decisions on infrastructure investments by maximizing available resources where they are most needed. The next step is to assess the effectiveness of these tools and processes. Due to COVID-19 related restrictions, this was delayed, though student outcome data were gathered.
Challenges and Failures
The development of such AI empowered solutions is a multidimensional problem that requires expertise in the computer/data science field, knowledge of the architectural engineering domain, large volumes of relevant existing data, and experience on the ground with local school communities. The goal of our collaborative partnership was to set up a process for the development of components that can be incorporated across governmental contexts, while utilizing a mobile app for the collection of additional images through a “community science” effort. These components were to be handed over to a Software Development class at the partner university to take them from a proof-of-concept stage to a usable system. Due to the COVID-19 situation, this handover had to be postponed. In the meantime, another team in an AI class refined and expanded the computational model, and we are planning a similar collaboration between the two educational partner institutions for the 2021-22 school year.
Conditions for Success
• A reliable algorithm that automatically classifies structural typologies in line with GLOSI based on field collected photos
• A platform to allow easy retraining and updating of the classification algorithm when more information is available (e.g., photos of different school building types collected in other countries)
• A sustainable and efficient cloud data management system to hold the baseline database.
This solutions was first constructed in the context of Nepal, and then replicated in Kyrgyz Republic. The goal is to scale this solution globally.
Students face unfamiliar situations and obstacles such as signing non-disclosure agreements to get access to resources. Coordination of teams from different time zones and continents requires meetings at inconvenient times, and consideration of constraints related to COVID-19. It is also becoming clear that a continuation of this effort with teams of students that change every term is getting more and more difficult as the complexity of the existing software base and data repository increases. We are exploring funding opportunities to establish a core team of student developers that can maintain consistency over longer periods of time. But the fact that this activity is still ongoing indicates that the individuals and organizations involved consider it worthwhile: for all the people involved in the project it has been a great experience to be working on a meaningful, real-life project with the potential to make a difference at scale for people in need.
The work described here represents the initial stage of an ongoing collaboration between a large public university in the United States, a large international organization, and a partner university in Europe, all focused on developing an innovative tool to improve school safety globally. Both universities have established digital innovation centres in partnership with a cloud services provider. Their role is to act as an intermediary between university personnel and students, and external collaborators, with an emphasis on projects that offer benefits to the public. The goal of this particular project is to improve the structural safety of school buildings in developing countries, especially with regard to natural hazards such as earthquakes. The cross-national, multidisciplinary collaboration demonstrated that the approach is feasible, with promising results.