This problem is a kind of closest pair between two disjoint set. Upperside picture
This problem is usually called the closest bichromatic pair problem. Here are a couple approaches.
Delaunay triangulation. (This certainly works with L2 (= Euclidean) distances; I think the steps generalize to L1.) For every Delaunay triangulation (there can be more than one in degenerate cases), there exists a minimum spanning tree whose edges all belong to the triangulation. In turn, this minimum spanning tree contains a shortest edge crossing the cut between the color classes.
Nearest neighbor data structures.
If it is given that the red points are separated in x from the blue points, then you may be able to adapt the O(n) merge step of the Shamos–Hoey divide-and-conquer algorithm for the closest (monochromatic) pair problem, described here.