dschwilk
commented
8 years ago Aren't the predictors the cause? There are missing data in the topo vars. The input grids have bad data. See https://github.com/schwilklab/skyisland-climate/issues/19
I don't trust the topographic variables right now at all and think we must solve that first.
I'm not sure why, but the RF model outputs have holes in them with many no data values. That's why they don't look good when projected in ArcScene for presentation purposes. This may be due to two things: 1) The use of NHD+ data, although I've used this before for my prior projects and did not have those issues. 2) The fact that much of the DM grid prediction surface is really far from where the actual sensors are located. Using the watershed approach seems logical, but the ends of the grid surface are really far from any of the sensors, meaning that those predictions likely are outside the envelope of tmin and tmax values.
My suggestion is that we see what the grids look like when we actually have PCs by mountain range and reevaluate. I see nothing on the web about holes in DEM data from NHD+ and the original input grids for all of the layers are hole free after looking arcMap at each layer, individually. NHD+ has values in cm (which I'm converting to M for the figures), and it has the streams burned in (-highly negative values for the streams), but like I said before, I have not encountered this problem before and that should not matter.