In the end I've managed to process the data following your recommendations but still, the allocated memory is to high.
For 60 seconds of data when I process one channel the memory goes up to 800 MB and on some cases even higher.
The problem is that besides the spectrogram which, I have to compute also order tracking, orders spectrum and frequency spectrum for entire set of data. I have notice another issue related with the order spectrum values computed with peak hold method. The data that is outside of the frequency range or close to it has some values higher than expected. I'm constrained to compute the order spectrum up to the 24 orders, but I've notice that the values that are close to the top end of the max frequency spectrum doesn't seems be real (e.g. rotor speed 2800 rps, Orders higher than 5).
Maybe now (when I'm processing first the colormap) I can look for the maximum speed and limit the number of orders accordantly.