A Physics-based Approach to Temperature Upsampling
August 19th, 2025
geospatial modeling, data engineering, forecasting
The industry that I work in is influenced heavily by the temperature. As such, accurate measurement of temperature is crucial for driving business decisions. Below I will discuss my approach and methods to upsampling a cooling-degree-days dataset by NOAA from state-level metric, down to the 50 meter level, allowing for much more confident decision-making.
What the problem was
State-level data would be fine, if the weather didn't play such a key role in our business. It is crucial to get a fine of data as possible, but without spending money for commercial licenses, this data is very hard to come by. Frugal as I am, I figured I would take a stab at using a free dataset from the National Oceanic and Atmospheric Administration (NOAA) that measures historical monthly cooling degree days (how many the daily mean temperature rises above 65 degrees Fahrenheit) across all the states, and try to squeeze out as much as I can from it.
What the obvious approach was
The first, naive approach I had was to simply create a temperature gradient between the states. So if the CDD for state A was 300 CDD for a given month, and state B (let say it's directly adjacent) was 400 CDD for the same month, then we can just create a gradient between 300 and 400 for all the points between the two states.
This raise several questions. The first is how do you pick where to put the 300 in state A, is it the centroid of the state? The measure is an assumably an aggregate, so do you locate all the measuring stations and calculate the centroid from those? No matter how you choose, it will be more or less arbitrary, but we will take steps later to account for that.
Next, what do we do about corners? Consider, for instance, New England. There are many states that are at tight angles from each other. Do we distort the gradient to account for this, or do we just take it as is. Moreover, in New England, the geography is better represented by the many more datapoints present as compared to, say, Alaska or Texas.
So we see simple linear gradient between state centroids causes distortions and bias in measurements, but can we devise a better approach?
Yes, of course we can
Let's take a step back and study how temperature actually behaves.