There's a few options, you could try a RasterExpressionEvaluator to find pixels with RGB values in certain (green) ranges. Alternatively a RasterCellCoercer to turn them into polygons (with the RGB values as attributes) and then the regular testing/filtering transformers.

I've tried it myself for just a little bit and I must say I'm not completely satisfied with the results so far. I used a RasterPaletteGenerator before the RasterCellCoercer to bring everything down to a fairly small number of discrete colors, hoping to save myself the trouble of selecting ranges of values.

However, on the sample photo I used there is not a whole lot of difference between a tree crown, a grassy field and the shadow of a tree crown on a grassy field.

Calculating an NDVI could be of help. The formula is (NIR - red) / (NIR + red) and it returns a value between -1 and 1.

1 generally indicates (healthy and dense) vegetation, 0 indicates sand/rock and -1 indicates bodies of water. Grassland and sparse shrubs should give values of about 0.5.

You could do this with the RasterCellExpressionEvaluator I think, but I haven't done this in FME yet.

It should be noted that this is a rough indication and there are many alternatives to this index, including ones that adjust for soil color, so be picky!

Oh, also take the temporal aspect of your imagery into account. Imagery from the growing season is more suitable than e.g. winter imagery.

thanks @joepk​ and @Hans van der Maarel​ . I calculate an NDVI and afterwards I set the values that are not required to nodata. With the help of the AreaCalculator I filter out the larger areas.

Interesting is also greenroof detection. I will try to intersect with building boundarys and than identify green areas with the help of NDVI.