As an alternative to actually filtering your data, have you considered just truncating your spectrograms? If your end goal is to feed this into a machine learning model, I think that this sounds like a lot of work just to give the model a dataset with a large region of "no data". It sounds like what you are hoping for is spectrograms that look similar to @B. Thomas' second image@B. Thomas' second image, but if that's the case then the entire blue region is not going to add any value to a machine learning model if all of your images have that same blue region. You should get equivalent results by just giving it spectrograms that all have been truncated at that same frequency value, since all the information in the image is below that frequency value. This also has the nice side effect of reducing the size of your dataset.
On the other hand, I think that by removing the harmonics from your spectrograms you are potentially losing a signal that a machine learning model could pick up on. You mention that you are worried about other signals in higher frequencies biasing your model results, but as long as your training dataset is large enough this shouldn't really be a problem. Unless you have a higher frequency sound that is consistently present with one type of call and consistently absent with other types of calls (which would be interesting to find out!) then it should all wash out as "noise" from the model's perspective.
