How can I classify two different types of radar modulation from images obtained in the time-frequency analysis?

Hello, as in the example in "Radar Waveform Classification Using Deep Learning," I need to classify two forms of radar modulation under study. These waveforms do not fit into the FMCW, BARKER, RECT, FRANK, COSTAS and LFM image library. Note that the picture that I saved (.png) has different modulation between them and the conventional ones. I would like to classify them into two distinct categories, with an acceptable percentage, type A or B. In the example in "Radar Waveform Classification Using Deep Learning," I saw that it was possible to classify 3 different types of modulations (RECT LFM, BARKER), and I would like to do the same for these two waveforms that I am studying. The data matrix (image) of the time-frequency analysis that I am doing with STFT has 801 (bins frequency) x 1207 (time resolution). I did not use the Wigner-Ville due to the high computational cost, and as the modulations are very different, the time-frequency resolution of the STFT is sufficient. How could I succeed in this study? Could you help me please? I'm starting to study this part of Machine Learning, and I saw that this toolbox and your help are directly linked to my final goal. Thank you very much

image from modulation type 1

image from modulation type 2

Answer 1

[Bjorn Gustavsson]

The second modulation is obviously a chirp. How can I be so sure? Well I cannot because there is no way I can distinguish a point-object accelerating away from the radar during one pulse and a stationary point-object scattering back a chirped radar-pulse (You haven't even put labels on the figures so there is no way anyone can say what they represent...). You can try to de-chirp those signals as best you can.

[Maximiliano Barbosa]

Hi, Bjorn Gustavsson. I am sorry, I put labels on the figures now; the xlabel represents time (microseconds) and ylabel frequency (MHz) for two pulses transmitted signal intervals of observation. However, I am interested in trying to classify these two types of waveforms. The first one actually is not just a chirp, but a new class of waveform that i study that has an LFM component plus a random phase component modulation.

Could you help me?

Thank you

[Bjorn Gustavsson]

If you know the transmitted waveform then you know how to describe it? To me it is a bit unclear what you're asking for, if you know what's transmitted then describe that and you have classified the modulation, right. If you have a random phase-code then it is a random phase-coded pulse, if you have some additional modulation on top of that then add that to the description? What more do you want to do?

[Maximiliano Barbosa]

Bjorn Gustavsson, I want to give similar results as in the example ''Radar Waveform Classification Using Deep Learning'' using that CNN, but with my images waveform modulation that I study. So I need to obtain a matrix, as shown in the figure below, but not with BARKER, LFM, or REC modulation; instead it, I want to train a CNN with my time-frequency images and try to get a good rate of classification perfomance was showed in example ''Radar Waveform Classification Using Deep Learning''. Could you understand?

Thank you so much

[Bjorn Gustavsson]

Then you should be able to simply reproduce the entire process of the training/learning of the network but with your responses as input instead of the inputs used in the example, but you will most likely have to make a modified version of the function: helperGenerateCommsWaveforms to also produce your novel radar-pulses.

[Maximiliano Barbosa]

thank you