A friend recently told me about a tool called climate.you that shows “temperature change, over land and sea”, at all points on the earth’s surface in a bid “to show how warming is already affecting people everywhere”. You can enter the name of your city or town and find out how the local conditions have changed. Based on interactions with some scientists who have written on climate modelling for The Wire and The Hindu, however, I’d also come to be wary of projections for scales smaller than a whole region, especially for a data-poor country like India. But after the chat, I also wondered if my position was outdated — and learnt that it was. So here goes an update.
Climate change is fundamentally global in its drivers but its effects operate across all scales, from continental changes in rainfall patterns down to local phenomena like coastal erosion. This said, a confusion about the phenomenon’s ability to operate at different scales often arises from the way scientists model it.
Global climate models collect data on atmospheric and oceanic parameters and simulate them on a grid whose cells are 50-200 km across, maybe more, which is a very coarse spatial resolution. When you render this grid on a screen, there’s a value for every pixel and, given the cell size, that pixel represents a regional average rather than a precise local forecast for the place at that pixel. But this doesn’t mean the model is wrong at that pixel, it just means it’s not designed to predict the consequences at that level.
(For example, if the RMC Chennai station says it’s 32 C right now, it’s harder to know how much the relative contributions of land-use, radiation from built structures, heat transported by local winds, and regional warming to that figure are. The temperature may also be sensitive to other factors we’ve deemed inconsequential, such as the amount of dust in the air around the station or traffic outside. A common way out of this seeming intractability, beyond quality control measures at the station itself, is to collect data for several years and check which temperature trends hold up and which ones fall away.)
The scale question is reminiscent of the coastline paradox: no well-defined landmass has a coast of well-defined length, yet the coast exists at all points along the edge of the landmass. This weirdness arises because the length depends on the scale at which you measure it. If you look at the India map zoomed out to 1:10,000,000 — like on Google Maps on your laptop screen — the coast shows some features but smooths over others because your laptop’s screen doesn’t have enough pixels to capture those smaller than a particular size. If you zoomed in further, say to 1:1,000,000, you’d find more features because there are more pixels now for a certain number of features, and smaller variations in the shape of the coast show up. If you zoomed in further, even smaller variations would show up, and so on.
Hat-tip to Sambavi P. for the fillip.
The climate signals that models are sensitive to are similarly, but partly, a function of the scale at which various instruments record those signals. In both cases, aggregating data from different scales to prepare a region-wide projection actually smooths over complexity rather than capturing it. There’s also no one ‘correct’ resolution and it’s not reasonable to expect a model prepared for one resolution to be equally accurate or certain at another.
This said, the analogy is only partly apt because it diverges over how geography and climate change behave at ever smaller scales. Specifically, while features on the coast become more numerous, and thus its length ever greater, as you keep zooming in, there are no signals relevant to climate change beyond a particular floor. Which means if climate change manifests as, say, a higher local tide level, for that parameter there’s no ‘zooming in’ beyond that point. Second, as you zoom in, climate signals become less messy whereas geographic signals become messier.
In fact, scientists have developed a technique called downscaling whereby they use a combination of statistical and dynamical methods to translate a model’s coarser outputs into finer projections. Obviously this isn’t a lossless exercise — you can’t get more information without paying a cost — and downscaling from one scale to the immediately next one ‘below’ adds some uncertainty. Which means a downscaled local projection carries the errors implicit in the global model plus the errors introduced by the downscaling method. Ultimately, the projection for a particular pixel exists: it’s just laden with uncertainty.
Now, while a model need not have a good projection for a particular pixel, that’s not synonymous with the data collected from that pixel being irrelevant or a non-signal for climate change. Local measurements like tide gauge records, weather station temperatures, regional snow measurements, etc. are all contain climate signals at very fine spatial scales. In other words being sceptical of a hyperlocal projection is reasonable but being sceptical of a local observation demands a higher bar.
For example, say a modeller feeds data about a city’s population density, road networks, and growth trends into a model of a city and tries to predict the congestion in your neighbourhood in one year. This effort is only going to be as good as the model’s assumptions. Now, say your neighbour leaves for work at 8 am every day for five years and tracks her commute time. This data has its own limitations — perhaps foremost that its patterns can’t be easily generalised — but while the model might excel at predicting how the city as a whole will change, your neighbour’s experience is a better predictor of how your neighbourhood in particular will.
In effect, it makes sense to be wary of sub-regional climate indicators derived purely from global model outputs without proper downscaling or local validation but to extend that cynicism towards observed data or even properly validated regional models would be to throw the baby out with the bathwater.