Let’s talk about these words: Downsampling, downscaling, upscaling, upsampling, large scale, & small scale.
“You keep using that word. I do not think it means what you think it means.”
Inigo Montoya, The Princess Bride.
When I work in climate resilience and spatial analysis, downscaling is the process of taking coarser resolution data and estimating a finer resolution for use at local scales. The rationale, going from a larger cell size down to a smaller size, makes sense. In some recent GIS reading, it became clear that downscaling can take on an opposite meaning for some GIS uses. This had me scratching my head a bit and led me to do a bit of research on the topic.
First, let’s do some defining:
Large Scale: A representation of the earth where objects appear large on a map compared to a small scale map.
Example: A local park map is a large scale when compared to a map of the country. If we express the ratio scale 1:200 in the form of a fraction, this is 0.005.
Small Scale: A representation of the earth where objects appear small on a map compared to a large scale map.
Example: A map of a state is a small scale when compared to a map of a town. If we express the ratio 1:200,000 in the form of a fraction, this is 0.000005.
Note: there is no quantitative dividing line between small and large scale, these are qualitative terms.
0.005 is a LARGER scale than 0.000005
Next, some downsampling explanations I pulled from multiple GIS sources:
Downsampling & Downscaling: These are interchangeable terms. When going down in scale, the ratio gets smaller. The file size gets smaller. The resulting output is less sample points, or pixels in the raster. This allows for the representation of the earth over a larger area.
Example: 1-meter lidar of a county can be merged with all other county data in the state and DOWNSAMPLED to 30-meters for ease of processing, faster upload and download speeds, and a smaller file size.
Upsampling & Upscaling: These are interchangeable terms. When going up in scale, the ratio gets larger. The file size also gets larger. The number of sample points, or pixels in the raster, gets larger. The representation of the earth is therefore usually reduced to a smaller area.
Example: If the best data one can find for a county project is 30-meter resolution, the data can be UPSAMPLED to 1-meter for estimated 5-foot horizontal measurements. The data will still be 30 meter resolution with the level of accuracy it was collected with but 5-foot measurements can be estimated.
What does the Dictionary say?
In the dictionary definitions, something that is UPSCALED has a higher level of quality or more desirable features. It doesn’t mean bigger, it means BETTER. In terms of data and rasters that can only mean that more refinement is happening to the data. These definitions agree with the dictionary as well. So what is going on?
Why are there two definitions?
As stated earlier, the rationale for “downscaling” climate data is the data is going from a larger cell size down to a smaller cell size. The difference in definitions has to do with raster imagery versus spatial analysis. The definitions where upscaling is the same as climate downscaling all discuss images and pixels.
Pixels are the smallest discernible piece of data but are not always cell. Cells are always pixels. But pixels are not always cells because pixels can be round and cells are always rectangular. Any image at the finest resolution will be a collection of small dots. Most spatial analysis rasters are not images, they are grid cell data. This means the process for dividing these units is different. Square cells neatly go down in size when split and interpolated into smaller areas. Round pixels are more complex, so while their size does go down it’s easier to describe the process as going up in the number of pixels. Pixel imagery is therefore upsampled and spatial data grid cells are downsampled to achieve the same goal. There are enough people in GIS collecting imagery and enough doing spatial analysis to make this confusing when working in related areas.
What can we do?
I would caution that whenever using these terms that people should do a content check for what is really being discussed before doing the work, especially with new clients or colleagues. When it comes to terminology where we think there is a standard agreement in meaning, this step is usually skipped. Think of downscaling like “inconceivable” in the Princess Bride, it does mean what you think it means… but does the other person understand what you mean?
Best,
Michelle
Michelle Toner 