generally, more trees make your classifier more robust and more accurate. However, when you reach a certain point, the accuracy does no longer increase with more trees. So you can test step wise:
10 > 25 > 50 > 100 > 200 > 400
and compare the resulting training accuracies. Once there is no longer a significant increase of accuracy, you don’t need more trees.
But: This all depends on the number of input features (rasters used for training) and training samples (polygons used for training). If these are too small, the random forest classifier becomes ineffective, because every tree is calculated based on the same training data. For example, if you only use 2 input rasters (VV and VH) and very small training polyons), it does not make a difference if you process 5 trees or 5000, they will all produce the same result, and therefore make the classifier ineffective.