I am trying to refine a raster dataset of annual rainfall from a 5km grid down to a 100m grid. Before I continue, I am aware that this could result in some uncertainty regarding the data. However, the data is only available in the 5km grids, and this is of no use to me, as I am looking to compare rainfall to landuse, and the landuse map is to a 100m grid.
To get around this, I thought I would be able to infill the data through the use of contours (as rainfall data is generally gradient based, rather than discrete isolated data points)
I contoured the 5km raster file using Contour (Spatial Analyst), and then tried to use the contour file to create a new raster file (Feature to Raster) at a 100m resolution. However, because the contour feature is a polyline feature, there isn't any data between the contours. This means that anywhere the contour lines fall more than 100m apart, the raster file leaves a blank pixel.
As I see it, there are three options that I may be able to look into (no idea how, or if, the first two would work)
Option 1: Directly refining the resolution of the original raster file, by increasing the resolution from 5km to 100m pixels and infilling the new pixels with values interpolated from the original file.
Option 2: Some sort of interpolation tool, (I saw several in the Geostats toolbox, but don't understand how they work) that could transform the contour polyline feature into a continuous dataset, which could then be converted to raster.
Option 3: Create a contour file with ridiculously fine contours (I have already tried 0.5 mm, over a range from ~ 500mm to 3000mm, and this didn't do the trick in places)
Best Answer
I know you mention this in your question, but I'd be very wary of down-scaling 5km precipitation grids to 100m. Precipitation is notoriously difficult to interpolate, as so many factors (local topography, temperature etc.) affect its spatial distribution. The two methods described by Aaron and Martin will work, but I wouldn't have much faith in the output if you go all the way down to 100m.
Two more sophisticated methods, both used a lot in climate science and meteorology, are statistical and dynamical down-scaling. This paper (not necessarily a recommendation - just one of the first hits on Google) compares the two approaches. However, even these methods are very unlikely to get you down to 100m resolution with meaningful results. They also require a lot of work, especially in the case of dynamical down-scaling.
Bear in mind that the 5km grids which you're currently using were probably themselves derived from coarser data in the first place (either by spatial interpolation of rain gauge measurements or from RCM/GCM output). It's likely that the data providers have already interpolated the "raw" data to the highest resolution reasonable given the source.
Quite a lot of my work tries to link climate data to land use and I often have the same problem - high resolution land use data together with coarser climate information. My current approach involves performing as many of the (non-climate related) land use calculations as possible at 100m resolution, and then aggregating the land use data to 1km using, for example, some kind of area weighted averaging technique. I've tried using statistical down-scaling to get my climate data from 5km down to 1km, but whenever I mention this to climate scientists they tend to inhale sharply and shake their heads at me.
I'm currently applying the 5km climate data directly across the 1km land use grid (so that the 25 1km2 cells within each 5km cell all receive the same rainfall).