While I agree with boftx that a real script would be more appropriate for something like this in a production setting, I appreciate that pushing one-liners to their limits can make for good learning exercises.

As for your splice based solution, I think that's actually pretty good. Here are two small suggestions to tweak it further:

1. You can get rid of the  BEGIN { $"=":" }  by using  $_  instead of  "@F"  to refer to the original line.
   
   
2. You can get rid of the separate  %line  hash by adding that information directly to  %h  whose contents would then look like:

   (
     ZM => [ [20470, "20470:ZM:Samfya:Africa"],
             [20149, "20149:ZM:Sesheke:Africa"],
             [18638, "18638:ZM:Siavonga:Africa"] ],
     ZW => [ ... ],
     ...
   )
   [download]

Here's the one-liner with those changes, in a "scriptified" representation (which I find easier to work with; it's trivial to convert it back to the one-liner format by removing the lines with comments after them):

use warnings;       # just for debugging
use strict;         # just for debugging
my (%h, $k, @F);    # just for debugging

while (<>) {        # -n
  chomp;            # -l
  $\ = "\n";        # -l
  @F = split(':');  # -F":" -a
  
  push @{$h{ $F[1] }}, [$F[0], $_];
}
for $k (sort keys %h) {
  print $_->[1] for splice [sort {$b->[0]<=>$a->[0]} @{$h{$k}}], 0, 4
}                   # -n
[download]