Writing it out helped clarify the problem. Since the MODIS comes with
LAT and LON rasters, I contoured them at whole degrees to clearly
identify the crossing points at (41, 90), (42, 90),..., (45, -99), (46, -99),
then noted down the associated raster (line, sample), still in the unprojected space.
Then, in another window, I took the UTM-projected TM data and displayed
the whole degree lat/lon grid. This gave me a set of corresponding (x,y) points.
From there I could throw away the actual latitude and longitude values
and build a table of MODIS_x, MODIS_y, TM_x, TM_y.
Should the TM be expressed as pixel-locations or UTM-meters?
Should all projection info for the TM be scrapped, treating it as a
raw raster?
A five-minute MODIS section covers a much greater area (at lower spatial
resolution), so I needed work from a smaller clip that fully contains the TM,
to minimize distortion, adjusting the numbers accordingly.
Mapping the TM straight to the MODIS projection is a problem, as the TM is 30m
and the MODIS is around 1000m. It won't do to merely warp TM to MODIS,
since the result would select one TM pixel out of a thousand to represent the value.
I need to fudge a raster 33.33 times the size of the original, increasing the
resolution, but keeping the same extents. The new MODIS would look like a
checkerboard, with one value covering a large square and the TM filling the block
at normal res. The goal is to count the pixels within each block, comparing the
MODIS fractional snow-covered area with the calculated FCSA from the TM.
Set snow as 1, bare-ground as 0 and cloud as NO_DATA, then a simple mean
calculates the FSCA.
I'll try fishnet to build a grid surrounding each big MODIS pixel, then
Zonal stats to find the mean. The first attempt made a polyline grid
and only computed stats along the lines. Need to convert the grid to polygon squares,
but it just runs for hours. A smaller section completed in 34 hours. I may have to
chop this into much smaller pieces and merge.