Procedural generation's history mirrors that of symbolic AI - hand tuned, interpretable procedures that parallel the conceptual models we have of a generating process.

The choices the author made reflect the way they conceptualize the generating process. This is in contrast to a data-informed approach.

A data-informed approach to island generation would be a spatially auto-correlated generating model whose parameters were derived from real life data on islands. You would then be able to generate more in-distribution islands that were statistically similar to islands in the corpus.

You don't often see this in the proc-gen world because part of the fun is adding epicycles and various tunable knobs.