I've never used the PLTS tools, and the help doesn't really give much info on how the Merge Raster Datasets works, but the Mosaic to New Raster geoprocessing tool, in the Data Management Toolbox under Raster Dataset, should work fine, and doesn't require an ArcInfo license or Spatial Analyst.
To clip your resulting mosaic without Spatial Analyst, use the Clip geoprocessing tool, also in Data Management under Raster Processing, and use the county line feature as the Output Extent. If you have the county feature selected and enable the "Use Input Features for Clipping Geometery" option, it will clip to the actual border. If you don't select the feature, it will clip the raster to the bounding box of the feature.
Stay in raster format.
If ArcGIS would provide a zonal percentile function you could do this in a couple easy steps. Since it doesn't, you have to work around that limitation. There are many ways to proceed. The key ideas are:
Determine the proportion of cells in each class represented by the desired sample of 300.
By thresholding a grid of uniform random values, you can identify slightly more than the number of points needed in each class.
Extract these points, then post-process them by retaining only those having the lowest random values within each class.
This procedure obtains a simple random sample without replacement, and independently, of each land class.
Here are some details. Let there be N classes (n = 3 in the question), each with n(i) cells (i = 1, 2, ..., N). Suppose you intend to select k(i) randomly from each group (k(i) = 300 for the class indexes i = 1, 2, 3 in the question). You will do that by generating a uniform random grid and comparing it to a threshold grid [T]. It has the constant value p(i) at each cell in class i, with p(i) somewhere (to be chosen appropriately) between 0 and 1.
The result of this comparison in class i is a random count X(i) having a Binomial(n(i), p(i)) distribution. Its expectation therefore is m(i) = n(i)*p(i), which needs to be close to k(i). However, X(i) can randomly be less than that. Its variance is v(i) = n(i)*p(i)*(1-p(i)). To be really sure that X(i) will be at least k(i), we need a very small lower percentile of the distribution of X(i) to exceed k(i). You can work out what the resulting value of p(i) has to be by means of a statistical calculator, but an effective rule of thumb (for moderately small numbers of classes N) is to ensure that k(i) is less than several standard deviations less than m(i). Letting a reasonable starting value for "several" be L, we obtain the quadratic inequality
k(i) <= m(i) - L*sqrt(n(i)*p(i)*(1-p(i)))
whose approximate solution is
p(i) = (k(i)/n(i)) * (1 + z(i) + z(i)^2/2)
where
z(i) = L/sqrt(k(i)).
In the question, k(i) = 300 and it looks like the n(i) are in the millions. Taking L = 3, say, gives z(i) = 3 / sqrt(300), which is about 1/6 (and z(i)^2/2 is small enough to neglect) This tells us to ask for 1/6 more points than the desired proportion k(i)/n(i).
With this (simple) calculation out of the way,
Reclassify the land classes i into the values of the p(i) you computed. Call the resulting grid [P].
Create a uniform random grid [R].
Set the values of [R] to null wherever [R] > [P].
This last grid will have slightly more than X = k(1) + k(2) + ... + k(N) (= 900 in the question) cells that are not null. At this juncture use your favorite method (there are many that will work) to
combine it with the land class grid and
output all tuples (i, R(i), x coordinate, y coordinate) as a point shapefile.
Post-process the shapefile by retaining only the lowest k(i) values for each class i: that's the desired sample. If you are so extremely unlucky as to not obtain enough points in some class, just repeat the whole process from the beginning. (Consider increasing the value of L when you do so.)
Alternatively, doing the work in R would be simple, because it offers functions to find the lowest k(i) random values in each class directly, rather than having to take several steps and convert things into a point shapefile format. The same conceptual workflow is involved.
Best Answer
Start by converting them to raster formats (assuming they are in DEM format to start with):
then use the 'Mosaic to New Raster' tool to combine the different rasters into a single one:
Make sure that they are in the same units before mosaicing them or they will look really odd when you stitch them together.
That's the basic approach I use, I'm assuming the DEMs you have are all in the same coordinate system/projection/units/cell size/etc. If not, make sure you get all the rasters into identical formats before you start, it will make your life a lot easier.