This disparity means the query reserves 531MB of workspace memory where a much smaller grant would be sufficient. Improving the estimate would not necessarily make your query run faster (though it would if 531MB were not immediately available) but reserving that much memory can reduce concurrency and may mean that valuable index and data pages are uncached to make room.
It is difficult to suggest improvements from an anonymized plan, because "ScalarString18" could be anything. In principle though, you may be able to improve the estimate by providing filtered or multi-column statistics. A computed column on "ScalarString18" may also help.
The other alternative is to materialize the 15,660 rows into another temporary table before performing the aggregation.