Granularity Matters

The finer the resolution, the better the risk management or resilience decision. A quick flood risk analysis of South Florida Starbucks shows why.

Something Could be Finer: Making GCM Data Actionable

The climate crisis is unfolding as I write these words. Recently a climate change-induced “heat dome” smashed temperature records in historically temperate areas all over North America. Climate-crisis-related disasters are crossing some very interesting and very scary thresholds. For example, during the recent heatwave, portions of Canada exceeded the all-time record high temperature recorded in well-known hotspot Las Vegas, Nevada. Yet even with these record-breaking events, some sites will make it through relatively unscathed. In another example, in New York City, heavy rains are flooding streets in places that don’t typically flood. But not all of NYC was negatively affected. Some parts of the city are quite capable of coping with the recent floodwaters. These differences are absolutely a social justice problem. Yet, before we address societal root causes, we must understand the geography of the risks. And therein lies the crux of this article. If our goal is to assess the risk of the climate crisis on specific regions, or specific assets, then granularity matters.

The United Nations International Panel on Climate Change (IPCC) collects and freely distributes multiple “Global Climate Models” or GCMs. These data are freely available. In recent years, many companies have been founded to collect and rescale these GCM data to be more granular and user-friendly. The publicly available GCM data are the combined output from hundreds of scientists from all over the globe. Given the amount of effort already invested in producing these datasets, why would private companies need to spend years and tens of millions of dollars refining the output?

The answer is that weather is what policymakers need to understand, and climate is what the GCM data are designed to study. The climate is not the weather. Climate is a description of a long-run average over a large area, and the weather is the realization of climate in a small geographic and time scale. Questions such as What is going to happen to my town? and What is going to happen to my house? are questions about the weather. Understanding how climate and weather interact is complicated. To understand how difficult this is, consider that companies (for example, the one at which I work) spend millions on massively complex cloud computing resources to take what the GCMs say about the climate and transform it into something we can say about the weather. This practice is called “downscaling,” and it is no small task.

A Relatable Example: Assets and Decisions Look Different at 1.5-Million-X Higher Resolution

Consider the risk faced by Starbucks locations in Florida. Starbucks makes delicious caffeinated beverages that increase my productivity (and therefore, I do not mind giving them some free analysis of their climate risk). The choice of Florida is obvious; “The Sunshine State” is subject to significant levels of flood risk, and flood risk among all climate perils is the most spatially variable. I picked Starbucks locations for this example because I like coffee, they are everywhere, usually similarly sized, and as a playful nod to the below image that spawned a viral photoshop battle a few years ago.

As noted above, GCM models need some help to be decision-useful. They typically simulate the Earth’s climate at one degree (or greater) spacing in latitude/longitude (approximately 110 km). That level of aggregation has important implications for the usefulness (or lack thereof) of these data. To quote a recent article in the journal Nature from Alice C. Hill:

Without information about where and how damaging events are likely to unfold, choosing the right adaptations to invest in is little more than guesswork. The smallest resolution of climate models is generally at a scale of 100–150 square kilometres. That sort of area can span several towns and extreme differences.

The exact extent of 110km grid cells may be hard to grasp. To get an idea of just how large an area the raw GCM outputs cover, I plot a grid of red one-degree latitude/longitude lines over Florida in the figure below. Note that only two of these large grid cells (and thus two risk levels) exist to describe all 6,297 km² (2,431 square miles) of Miami-Dade County. Anyone seriously concerned about local conditions in Miami will need better data than what’s offered by 1-degree grid-cell resolution. The figure callouts demonstrate how climate service providers fill this gap. Combining the GCM output with expert science and local ground conditions, they refine (downscale) the GCM outputs to the 90m grid cell level. For context, within a single 1-degree latitude/longitude cell, there are approximately 1.5 million 90m grid cells.

In my analysis, l use Florida counties to represent aggregate level data instead of 1-degree grid cells. A county is usually slightly smaller than a 1-degree grid cell, but it makes for a more naturally understood aggregation. Plus, even with this somewhat smaller aggregation, it is still very apparent that the granularity of the data matters.

Both high and low-risk regions are well represented in the sample. Of the 302 Starbucks sites sampled, roughly 30% are located in the very-high-risk Miami-Dade and Broward counties. The next most Starbucks-dense county is the relatively much safer Orange County. The figure below summarizes the risk level at each location and county. The green color indicates a lower risk for both counties and points, and the red color indicates higher risk.

Aggregate Information Masks Realities “On the Ground”

The main takeaway from this analysis is easy to see from a simple visual inspection of the figure. Not all the locations in very risky counties are at high levels of risk, and not all places in ostensibly safe regions are safe. Therefore, focusing only on the aggregate information would often lead to inefficient or — worse — flat-out wrong conclusions. Furthermore, errors from aggregation are often correlated with things we are interested in measuring — for example, the level of damage to structures. This can lead to biased estimates when the climate projection data are used in downstream models.¹

If we zoom in further to Miami-Dade County, the pattern is even more obvious. Multiple Starbucks locations are colored green, indicating relatively lower risk, and many others are dark red, indicating higher risk relative to the county median. Even in Miami, the region widely regarded as being one of the riskiest, there is broad variance in outcomes at a hyper-local level. For any asset or property manager, this is crucial information to understand.

Taking another point of view, consider the scatter plot below showing the risk score of all our Starbucks locations plotted against the average risk in each county. The color in this chart indicates whether each site is above or below the 45-degree line, indicating higher or lower risk on average than the surrounding area. The plotted circles nearly form a square, indicating that the total range of the risk within each county is only loosely correlated with the average risk in that county. In other words, the maximum and minimum risk scores don’t change much from the safest to the riskiest counties, and just knowing the risk of a county is a potentially poor predictor of a given location in that county.

A Resolution to the Granularity Challenge

In this playful example, I show that the scale at which we view climate change data, a.k.a its granularity, really does matter. This blog highlights a source of measurement error in climate analytics that is very likely to introduce unwanted bias in any models using these data. When we know the location of assets at a granular level, we should evaluate climate perils at that same level. Fortunately, the granularity problem is easy for analysts to solve; simply contact your friendly neighborhood climate service provider. There are tools in the marketplace available today that provide this level of resolution. So why not start there?

Ryan Vaughn, PhD is a Technical Product Manager at Jupiter Intelligence. Learn more about Jupiter at jupiterintel.com.

Footnotes

¹ As noted by Solomon Hsiang in Climate Econometics, “Coarse climate measures introduce large measurement errors that will cause attenuation bias, leading to under-rejection of the null hypothesis”