Why is Majority so much slower than mean in Focal statistics?

Question

I am trying to use this solution to clean up small holes in my rasters. It uses focal statistics on the no data cells to replace them with a local average. The raster have a small number of unique values so using MEAN in focal statistics increases the number of values and gives dodgy results (I want a hole in a section where value = 100 to be 100 and not 98.9). The shouldn't be a problem because I can use MAJORITY in focal statistics.

What is stange though is that this python code takes 9 minutes to run:

filled = sa.Con(sa.IsNull(in_raster),sa.FocalStatistics(in_raster,
                     sa.NbrCircle(50, "CELL"),'MEAN'), in_raster)

still a little slow but manageable.

Then this code took 4 hours and 26 minutes to run (on the exact same in_raster):

filled = sa.Con(sa.IsNull(in_raster),sa.FocalStatistics(in_raster,
                    sa.NbrCircle(50, "CELL"),'MAJORITY'), in_raster

The only difference is mean vs Majority. Could anyone explain to me why there is such a difference in performance and if there is anything I can do about it?

I have a slightly laborious work-around of reclassifying the output to return to my small set of unique values, but I'd like to understand what seems like a simple solution is unworkably slow

Answer 1

I tested my idea in comments using this (perhaps not the most Pytonic and efficient) script:

import numpy,time
# create list of integers
N=int(1e6)
aList=[]
for i in range(N):
    v=int(numpy.random.random()*50)
    aList.append(v)
## calculate mean
t0 = time.time()
for repeat in range(100):
    total=0;count=0
    for v in aList:
        total+=v;count+=1
    mean=float(total)/count
print ('it took %i seconds to compute mean = %s' %((time.time()-t0),mean))
## calculate majority
t0 = time.time()
for repeat in range(100):
    aDict=dict();countMax=0
    for v in aList:
        count=aDict.get(v,0)+1
        aDict[v]=count
        if count<countMax:continue
        countMax=count
        majority=v
print ('it took %i seconds to compute majority = %i' %((time.time()-t0),majority))

RESULT:

• it took 27 seconds to compute mean = 24.510802
• it took 47 seconds to compute majority = 10

Does not explains massive difference observed.

NOTE:

I found that map algebra on rasters stored in geodatabase is much slower (and often unreliable) compare to algebra on humble grids.