search

DrCoffey / DeepSqueak

DeepSqueak v3: Using Machine Vision to Accelerate Bioacoustics Research
BSD 3-Clause "New" or "Revised" License
373 stars 89 forks source link

Call Feature (Slope) Explanation #204

Closed babywing closed 1 year ago

babywing commented 1 year ago

Hi! I am interested in analyzing call features, and I noticed the "Slope (kHz/s)" feature in the call stats file. I understand this feature is about the contour, but what information does Slope provide compared to Delta Freq? I checked the values, and the slope for some calls is -600 kHz/s -- thus, to my understanding, this is not the change in frequency of the call (and I'm assuming that's what delta freq is). Can you please define what "Slope" is as a USV property?

It also might be helpful to provide a glossary to interpret some of these features in the context of USVs. Thank you so much!

VoiceScientist commented 1 year ago

Delta frequency is simply the difference between the highest and lowest frequencies detected in the contour. The slope gives you the linear change in frequency over time of the contour. I don’t know exactly how it is calculated, but I think it’s a line of best fit (someone correct me on that). There is a minimal description of the output parameters on the wiki here: https://github.com/DrCoffey/DeepSqueak/wiki/export-to-excel Hope that helps!

DrCoffey commented 1 year ago

@VoiceScientist is correct. Slope is calculated via a linear best fit on the entire extracted contour.

babywing commented 1 year ago

Thank you! I definitely need a more detailed explanation for the slope. I have interesting data regarding that feature, and I want to interpret the results as best as possible. Is slope only used for detection/classification? Or is it a meaningful USV property?

DrCoffey commented 1 year ago

That is a great question, and one I don't really know the answer to. I'll say that slope is a relevant descriptor for simple calls (flat, up sweeps, etc.) but it doesn't do a good job capturing complex calls with multiple components. The attached example call has 2 sweeps, one up and one down, but the linear best fit line just goes up... image image

babywing commented 1 year ago

Thank you so much! This is actually very helpful. I appreciate your explanation and examples!