I believe the reason is that graphical (2d) notations are harder to communicate and therefore not standardized. If they are standardized, they are usually very application specific and not really general enough. For example lambda calculus has a nice graphical representation as string diagrams, since it corresponds to calculations in a cartesian closed category. Typed lambda calculus has a similar representation, but with multicolored strings.
Pretty much any problem easily modeled as data flow has a nice a nice representation in an appropriate (traced, braided) monoidal category, examples would be electrical circuits, chemical reaction networks (simple atmosphere models, so called "Box models" are an example), hardware control software (LabVIEW), general ODE/Statemachine simulation (SIMULINK), "business processes" (forgot the name of the vendor), not to mention the various video/audio/shader programming environments.
Probably the intersection between the people capable of building a "generic" programming language for graphical programming and people who know enough category theory to inform the design is zero, even though software that solves subproblems of this kind is very successful. Ultimately you really want at least a 3d graphical language, because you want to express relations between 2d processes (for example the equation a+b = b+a, should be really understood as such a 3d process). Coincidentally relations between functions is something even programming languages like Haskell fail to address, it has type classes, but no way of specifying the equations that are required to hold.
The only language I'm aware of that does that is Magma, which unsurprisingly is based on Operads (essentially multicolored monoidal category with a fixed set of monoidal functors and natural transformations) and was designed by mathematicians and to a lesser extend Mathematica and Prolog.