When the compiler code size is a fraction of the whole it would actually make sense to improve on it instead of rewriting everything. Like Facebook did for PHP and their HHVM. Obviously Fortran doesn't have a speed problem but it could be improved in other ways like static analysis and nicer syntax notations.
Converting to later versions of the Fortran standard would get you a long ways there. Plus with all the focus in 2003 & 2008 with C interoperability, you can plug in later open source libraries.
[edit: I wonder what a modern version of RATFOR would look like]
