Image courtesy of Danielle de Haerne.
The Himalayan region—Nepal, Bhutan, and the India Himalaya—hosts not only intrepid mountaineers but also seventy-four million regular inhabitants. Thirty-six million of them live in areas susceptible to multiple natural hazards.
Yale researcher Jack Rusk, a graduate student at the Yale School of the Environment and the Yale School of Architecture who works at the Karen Seto lab, led a project whose machine learning model produced that last statistic. When considering wildfires, floods, and landslides as “hazards,” forty-nine percent of people in the Himalayas live in areas susceptible to more than one of them—despite those areas only encompassing thirty-one percent of the region’s land. The model produced by Rusk and his colleagues provides data on the susceptibility of the Himalayas and allows geographers to consider hazard management in a novel way, treating different hazards not individually but all at once.
Achievement of the models
Rusk’s study is a part of the Urban Himalaya project, a NASA-sponsored collaboration between Yale, the University of British Columbia, Kumaun University in India, and the International Centre for Integrated Mountain Development (ICIMOD), an intergovernmental agency for the Himalayan region. The project seeks to understand two dimensions of the connections between Himalayan urbanization and natural hazards: how natural hazards affect urbanization and how urbanization processes can induce or prevent further hazards. Rusk’s model attempted to answer the first half of that issue.
It took three years of adjustments to get the final model, which considers the risks of floods, wildfires, and landslides. Using historical records of the hazards and known environmental characteristics, the model would produce a map of hazard susceptibility for each variable, then combine those into a single map for overall multi-hazard risk.
An initial difficulty for Rusk was that the three hazards were tracked with different parameters. For example, he had to compare hazard intensity data in forms as different as flood depth at a particular location and the total volume of a landslide.
There was also the issue of non-reported data. If a hazard happens in a less populated area, it is less likely to be reported, skewing the distribution to favor more populated areas. Additionally, not all the factors had data for the same number of years. Furthermore, to have a single model incorporating different hazards, one would have to consider the same environmental factors.
The final paper used ten environmental variables, some of the more important ones being elevation, distance to permanent water, type of land cover, precipitation, slope, and soil type. Certain variables may be correlated for some hazards but not others, and not all of these would be relevant for every hazard. For example, slope has little to do with wildfires but is more associated with landslides.
Hence, a multiple-hazard informed model seemed implausible at first. Yet, these considerations are necessary to understand and improve life in the Himalayas, where multi-hazard risk is present over the long term and in the short term. Floods, wildfires, landslides, and earthquakes commonly cause each other, so hazard mitigation teams need to be prepared to handle these hazards simultaneously.
“You need to develop a framework for hazard mitigation that describes the overlaps and interactions between hazards,” Rusk said. “Practices that are good for managing the risk of one hazard might exacerbate the risk of another.” For example, clear-cutting land is a common hazard prevention method for wildfires—creating a fire break. But since trees stabilize soil with their roots, denuding a piece of land also makes it more prone to landsliding, potentially leading to disastrous outcomes.
Understanding the results
Rusk ultimately found that maximum entropy modeling would be the best for his data. Maximum entropy modeling works by finding the uniform hazard distribution for the entire region while accounting for the environmental variables. This methodology has several advantages. For one, it works without knowing where hazards did not happen, which negated the issue of inconsistent reporting. Additionally, it can handle both categorical and continuous environmental factors—for example, specific types of soil and total precipitation are both variables in the final model. It also does not lose accuracy when fed irrelevant or correlated factors, allowing for a consistent set of factors to be used for all three hazards. Finally, it outputs a single probability for each hazard at each location. This simple output allowed Rusk’s team to use a consistent methodology for defining “risk” for all three hazards, allowing them to combine the three hazard maps into one. When the model was constructed using a subset of the historical data, it was able to predict patterns in the rest, an early indication of success.
Rusk’s results must be placed in the context of Himalayan urbanization patterns to make sense. Himalayan urbanization often occurs as micro-urbanization, a term coined by Seto to describe the growth of settlements that are small, scattered, and removed from existing cities. Tzu-Hsin Karen Chen, a postdoctoral fellow at the Seto lab who collaborated on Rusk’s study, attributes micro-urbanization in the Himalayas to a feedback loop initiated by road construction. “Villages [near a road] will have a lot of new products that are transported to the market in the urban area, and therefore they have more capital to expand,” Chen said. Thus, people flock to settlements in thin, fertile valleys that are convenient places for expanding existing cities and optimal places to lay roads leading to them.
However, these valleys are also the most hazardous parts of the Himalayas. Their moist, fertile soils take less water to saturate in a flood. Their steep hillsides and low elevation make them prone to landsliding, especially as settlement on the valley bottoms forces people to move up the hills and cut terraces for arable land. These valleys also have hotter temperatures than higher elevations do, making them more prone to wildfires during droughts.
And yet, millions still inhabit these hazardous areas. “There are reasons to be near these urban agglomerations that aren’t related directly to the presence of hazards—access to education, access to healthcare, access to the money economy,” Rusk said. People choose these opportunities for socioeconomic mobility, despite the hazards, in the hopes of connecting with a wider world.
Where to go from here?
Rusk is the first to admit that his work would have been impossible without his fellow researchers physically located in the Himalayas. “I’ve been humbled by the opportunity to work with such an amazing group of collaborators,” Rusk said. Truly understanding the impacts of hazards requires talking to people where they happen; machine learning models can only go so far since they don’t explain why hazards happen in certain patterns or how they affect people. “In all of this work, you just have to shuttle between large-scale patterns and everyday life on the ground,” Rusk said.
The Yale team’s next project will zero in on how urbanization changes the landscape locally and affects hazards—the second part of the Urban Himalaya project’s overall goal. “We have one map that assesses overall hazard patterns across the past three decades,” Chen said, referring to the output of the current model. “But now we want to have a map for every year, from 1992 to the present.” These maps will allow the researchers to see both hazards and urbanization change together.
Humans are not only changing the environment on a local scale but also on a global scale. As climate change increases extreme precipitation and lengthens droughts, existing hazards will also grow in frequency and destructiveness.
Managing multi-hazard risks requires the coordination of normally independent national governments, local agencies managing separate hazards, and individuals alike. Rusk’s group has helped illustrate that progress can be made with an integrated, multi-talented team looking at the big picture. Now it is time to do the same back on the ground.
Further Reading
Rusk, J., Maharjan, A., Tiwari, P., Chen, T.-H. K., Shneiderman, S., Turin, M., & Seto, K. C. (2022). Multi-hazard susceptibility and exposure assessment of the Hindu Kush Himalaya. Science of The Total Environment, 804, 150039. https://doi.org/10.1016/j.scitotenv.2021.150039
Grainger, C., Tiwari, P. C., Joshi, B., Reba, M., & Seto, K. C. (2021). Who is vulnerable and where do they live? Case study of three districts in the Uttarakhand region of India Himalaya. Mountain Research and Development, 41(2). https://doi.org/10.1659/mrd-journal-d-19-00041.1