It depends what kind of clusters, but if these clusters are not overlapping in XY, one way to identify these is to use a HullReplacer. The pointcloud need to be coerced to a single multipoint first using a PointCloudToPointCoercer, else the HullReplacer will not work.

Then use the PointCloudStatisticsCalculator to find the lowest z and a PointCloudFilter to find the points with this z.

Attached sample workspace demonstrating this.

It works great except for places where the clusters are intersecting at y (as you maybe mentioned).

So the clusters im interested in are the pink ones, and here i will only get three values which is the lowest one, is there a way to do this and get all values? 

The classic way to do this is to flip the data over one axis, so you can apply a 2d transformer also from the side. This way you will be able to split the single cluster in 3 seperate clusters. Not easy but doable. Can you share a sample of your data? Just clip a 30 meter buffer and use an offsetter with  random kilometers in x,y,z to anonimize it.

Here is a testfile containing that exact pole in the picture

Sample workspace. It first creates the clusters in 2d, then rotates the data 90 degrees over x and creates clusters again, grouped by the 2d cluster id’s.

This worked perfectly! So grateful!