By Jackie Ratner
“Why don’t we know this already?” “Why hasn’t science figured this out yet?” “We have talking fridges but something as important as this is still unsolved?” These are some of the questions people ask every time there’s a disaster. (OK, maybe not the fridges specifically…)
We collectively have trouble understanding why disaster science, especially predictive science, is so far behind the tech curve, given its outsized importance to our lives. Part of the reason is that disaster science is incredibly complex! Whether we’re trying to harness atmospheric physics to understand the inner workings of a hurricane, or trying to understand the sociological complexities of keeping people safe in crises, there’s just no easy way to “predict” disasters.
But what about that old saying, “It takes a village to raise a child”? It occurred to me, as early as the 2004 Boxing Day Tsunami (also known as the Indian Ocean Tsunami), that a village could share the load of figuring this mess out instead of wholly relying on scientists and experts, especially in regions where scientists and experts might seldom be found. After all, when it comes to disasters, we’re definitely all in this together, like it or not.
Fifteen years later, along with co-authors Jonathan Sury, Mike James, Tamsin Mather and David Pyle, I’m proud to announce the publication of my first paper presenting evidence of what I’ve suspected all along: the “village” can indeed help predict disasters.
To find this out, we blended a few research methods in computer science, geography, and geohazards, and applied them to a hypothetical disaster situation to see if citizen scientists or the general public could help collect valuable terrain data, and whether that data would be any good for the purposes of disaster prediction.
Terrain models (3D digital representations of the Earth’s surface) are important for predicting potential disaster scenarios. Without terrain models, we can’t know where it might flood, how severe a landslide might be, or how quickly an avalanche can move.
The historic issue has been that terrain models are costly to make. Data for even a small study area can cost thousands, or likely tens of thousands, of dollars. Sometimes you can get terrain data for free, but often the tradeoff is that it’s lower quality and usually it’s not precise enough for predicting a disaster scenario. The commonly used free datasets derived from satellite missions would need to be 3 to 9 times more precise to give a good prediction for many disasters, such as landslides, avalanches, most volcanic phenomena, and flash floods. Some kinds of floods over large, flat areas can still be predicted reasonably well from lower resolution datasets, and little of the discussion here applies to disasters that happen below the Earth’s surface, such as earthquakes, or in its atmosphere, such as hurricanes. That said, it does apply to hazards that are related to and caused by earthquakes (landslides, rockfalls, etc.) or hurricanes (floods). The important distinction is to recognize where the hazard occurs: if it occurs at terrain level, then our article is relevant.
Typically as scientists, we operate under the assumption that we want the highest quality data we can get. However, for disasters, that’s not always the case. Sometimes we just have to make do with what’s available, or if time is of the essence then we need to just get moving even if everything’s not quite perfect yet. That’s where citizen science or crowdsourcing comes in: you utilize other people to help accomplish an otherwise laborious task in less time, even if they’re not highly trained in how to do it.
For our study we crowdsourced the task of data collection — in this case, collecting digital photos of terrain. Our experiment took place at Agios Georgios volcanic crater in Greece, part of the Santorini caldera volcano. Our citizen scientists were a group of Oxford undergraduates visiting the site to learn and improve field techniques, and thanks to a prior Oxford study on deformation at Santorini, we already had access to Lidar for the area to compare with the crowdsourced data.
Our crowdsourcing would not have been possible even a few years ago, because this work depends entirely on a relatively new technology called structure from motion, or SFM (which I did not develop, I only opportunistically co-opted it for the benefit of humankind, which seemed fair). It’s a kind of computer science that uses digital photographs as input data to yield 3D digital models of whatever was in the photos. It does this by treating the pixels in the photos as data points, and looking for similar clumps of pixels across different photos. It makes best guesses as to which pixel clumps represent the same thing in different photos, and uses it to reconstruct that object from multiple perspectives. The basic procedure is called photogrammetry and there are a number of other ways to do it, including pen and paper, but you can see how computers aid in getting it done much faster!
Depending on the algorithms employed in the structure from motion process, it can be incredibly good at matching pixel clumps to each other and throwing out false matches (95% confidence or better). With such high fidelity, it didn’t seem entirely necessary for the photos to be expertly sourced. But although this had already been shown to work well for things like architecture, it hadn’t yet been applied in the geosciences.
Terrain presents unique challenges when we’re talking photography: among 10 pictures of a bunch of rocks and dirt, could you spot the same rock from a different angle? You could probably do that pretty quickly if somebody showed you 10 pictures of a building from different perspectives. Because of these unique challenges, we couldn’t assume that structure from motion with crowdsourced data would work in geoscience the same way it works in architecture. We had to test it to be sure.
The 17 undergraduates, with no prior experience in photogrammetry or structure from motion, were split into 3 groups with different directives and given one hour to collect photos. Group A wasn’t informed about the project at all; they were asked to role play as people who would possess “incidental” photos of terrain that, in a real world scenario, could be mined from the internet. Such roles included tourists, travel bloggers, and photographers. Group B was given a one-sentence directive on how to capture quality photos for the project, and were asked to approach it from the perspective of a concerned citizen, observatory intern, or photogrammetry hobbyist — people who, in the real world, might have interest in disaster risk reduction or community engagement. Group C, the “experts,” was given a four-page manual explaining the project, its objectives, and the nitty gritty details of structure from motion data collection, as well as a mapped route for taking photos. We wanted to find out whether the data sets would progressively improve in quality if people were more informed before taking photos.
Once the photos were collected and randomly downsampled to data sets of equivalent size, we ran the photos through the free structure from motion software Visual SFM, and refined the resultant point clouds using the free software Cloud Compare. We were able to use the previously collected Lidar as reference data, and mapped our SFM point clouds onto it. We also created another SFM data set combining photos from all three groups into Group ALL, a mixture of the photos.
We were mildly surprised to see that with the exception of Group A, the other three SFM data sets actually outperformed the Lidar in terms of data density for the study area, by up to a factor of 10. There were some differences in the density distribution of data points, which is a quirk of structure from motion that makes the result look uneven when compared to Lidar. But each data set was accurate to within a meter of the Lidar, which is more than adequate for disaster modeling. With the low cost, speed of processing, and low barriers to accessibility (computers and camera phones are ubiquitous; Lidar rigs are not), it was easy to conclude that crowdsourced SFM indeed produces a viable alternative to traditional terrain modeling techniques.
And now we’re pretty reasonably sure* that there’s a better (more precise), cheaper, and faster way to create digital models of Earth’s terrain, which is important because we need terrain models to predict disasters. All you need are a basic computer set up and digital photos collected via citizen science or crowdsourcing.
Check out the study in Progress in Physical Geography!
The next step is to compare Lidar and SFM point clouds again, but in a less ideal environment. A valley in Cornwall, UK, that is prone to flash floods provides a smallish (1 kilometer) field area with vegetation and human activity that would be expected in a real-life disaster area. After we compare Lidar to SFM for point clouds and flood models in this location, our next study will enlarge the field area to several kilometers, to a hazard-prone valley in rural Ecuador that doesn’t have Lidar or other high quality reference data available. We’ll assess how well SFM-derived terrain models work to predict the flow of volcanic mudslides (called lahars) in this area. With these more realistic scenarios, we hope to bring this technique closer to being utilized in the real world, to build terrain models and save lives.
*Classic scientist/academic hedging, right?
Jackie Ratner is senior project manager at Columbia University’s National Center for Disaster Preparedness.