141 lines
7.3 KiB
Markdown
141 lines
7.3 KiB
Markdown
---
|
|
date: 2023-01-26
|
|
author: Patrick Weygoldt
|
|
type: talk
|
|
speakers:
|
|
- name:
|
|
affiliation:
|
|
- name:
|
|
affiliation:
|
|
aliases:
|
|
tags:
|
|
---
|
|
# Chirp detection poster script
|
|
|
|
## 10 minute presentation
|
|
|
|
introduction:
|
|
- Project goal: Develop a chirp detection algorithm
|
|
- What are chirps?
|
|
- short frequency excursions in ms range of EOD (electric organ discharge) of weakly el. fish.
|
|
- Show plot:
|
|
- spectrogram of the EODf of two fish (two lines)
|
|
- frequency resolution 150Hz
|
|
- nfft: number of windows/datapoints over which the Fourier transform is performed
|
|
- frequency over time is shown
|
|
- color indicates power
|
|
- chirp = upper line, frequency increases shortly
|
|
- Problem:
|
|
- to resolve chirps on the time domain, frequency domain too coarse
|
|
- if lower fish chirps it becomes even harder
|
|
=> time-frequency uncertainty problem (general)
|
|
- Goal: Improve existing detection methods to detect and assign chirps for electric recordings with n fish
|
|
|
|
Chirp detection algorithm
|
|
Availabe data:
|
|
1. Raw electrical signal (EOD of multiple fish) over n electrodes (n = 11)
|
|
2. Tracked frequency bands on spectrogram (pre-tracked): just as in upper right plot, but with lower sampling rate (3Hz)
|
|
- for one frequency we want the electrode on which the power of the f is the greatest
|
|
- power of the strongest of 11 electrodes for each frequencypoint in time was used to track the frequency band
|
|
- we cannot track freq on spectrogram's time resolution is too low for detecting chirps if you wanna distinguish between the fish
|
|
Feature extraction (in 5s rolling window):
|
|
1. Bandpass filter around the tracked frequency band for one individual (+-5Hz)
|
|
- first subplot grey, red = envelope of filtered baseline
|
|
2. Dynamic bandpass filter above baseline (+-5Hz) = 2nd subplot, gray filtered search frequency, orange = envelope
|
|
- dynamic search window above the current fish of interest
|
|
- Why dynamic: if another fish has a higher frequency, we need to find a window without another fish to be able to detect the chirp
|
|
- window above fish: look if there's another fish (array stuff), True/false thing, find longest subarray
|
|
- chirps excursions always increase the frequency and decrease the amplitude
|
|
- to find chirp, we need to search above the fish and look for break down in amplitude
|
|
- no peaks in filtered above = no chirp
|
|
- amplitude break down of baseline can have multiple reasons (e.g. fish swims away, stone)
|
|
3. Instantaneous frequency of baseline = 3rd subplot, gray filtered inst., yellow = envelope
|
|
- calculated on filtered raw data
|
|
- get zero crossings of each period and calculate frequency manually
|
|
- we know that chirps are increases in frequency, here we look at the frequency feature of chirps
|
|
Peak detection:
|
|
1. Detect peaks on bandpass filtered and inverted baseline envelope (lower red line)
|
|
2. Detect peaks on bandpass filtered search frequency (lower orange line)
|
|
3. Detect peaks on absolute inst. freq (lower yellow line)
|
|
- Peak prominence: Minimal distance from highest peak to next peak
|
|
Peak classification:
|
|
- all three features have to be present at once in a 20ms window (appr. chirp length) in strongest electrodes
|
|
- mean of peak timestamps of features is saved as chirp timestamp
|
|
|
|
Chirps in dyadic competitions
|
|
1. Competition experiment by Til Raab:
|
|
- two fish compete for one shelter
|
|
- 6h recording, 3h light, 3h dark
|
|
- electrical and video recordings
|
|
- with video recordings, behavior was tracked and assigned to an antagonistic category: Chasing (on- and offset) and physical contacts
|
|
- we did behavioral analysis with the detected chirps of our algorithm
|
|
2. Plot: Contact an chasing event timepoints, chirps of both fish, tracked frequency bands
|
|
- from literature: Chirps assumed as submissive signal by loser fish
|
|
4. Winner Loser boxplot
|
|
- chirps counts for winner and loser (n=22 recordings with winner and loser)
|
|
- loser tends to chirp more (Wilcoxon not significant, but trend with 0.054)
|
|
- white lines are paired fish for competition
|
|
5. Size difference plot
|
|
- Literature: Larger wish usually wins. (Larger resource holding potential theory, Till with rises)
|
|
- The smaller the size difference between fish, the more chirps are emitted taken winner and loser together
|
|
- correlation within winners and losers are not significant
|
|
- n = 21 because one recording with equal fish size was excluded
|
|
6. Frequency plot
|
|
- Literature: Males are more aggressive and chirp more, males have a lower EODf
|
|
- EODf has no effect on the competition outcome
|
|
|
|
Chirps emitted by loser fish might stop chasing events
|
|
- Chirps were centered around the timestamps of each event in a +-60s time window (for each category and each recording)
|
|
- kernel density estimation of centered chirps (gaussian kernel with 2s width and 10ms resolution)
|
|
- We show some example plots
|
|
1. First plot: No correlation case for chasing offset
|
|
- no correlation between chirping and the offset
|
|
- this was the case for most recordings and all events
|
|
2. Second plot: Correlation case for chirping and offset
|
|
- For some few dyads/individuals, chirp rate increases drastically before the chasing offset
|
|
- also slightly visible before chasing onsets
|
|
- no correlation for physical contacts
|
|
3. Third plot: Time of chasing events in the night VS the chirps during the chasing events and during night
|
|
- fraction of chirps during chasings is not increased relative to the fraction of chasing events overall
|
|
- Chirps do not seem to have an increased significance for chasing events
|
|
- only for some few dyads the chirp rate increased during chasings
|
|
- Gray/black areas:
|
|
- bootstrapped data (n = 50)
|
|
- all chirps for one recording during the night (because there more chirps)
|
|
- all shuffled chirps again centered around event and convolved
|
|
|
|
Conclusion:
|
|
- First tests indicate that our algorithm is able to detect chirps in recordings of multiple fish
|
|
- Algorithm results were applied on behavioral data for further analysis
|
|
|
|
|
|
## 2 Minute presentation
|
|
|
|
Introduction:
|
|
- Project goal: Develop a chirp detection algorithm
|
|
- What are chirps?
|
|
- Short frequency excursions in ms range of EOD (electric organ discharge) of weakly el. fish.
|
|
- to resolve chirps on the time domain, frequency domain too coarse, especially for multiple fish
|
|
- Goal: Improve existing detection methods
|
|
|
|
Detection algorithm
|
|
- Improved existing detection methods by extracting 3 features that change during a chirp but are not limited by the time frequency uncertainty
|
|
1. Amplitude drop of EOD (show trough)
|
|
2. Peaks of instantaneous frequency of EOD
|
|
3. Peaks in the dynamically adjusted frequency band above the fish's baseline EODf (special).
|
|
- Detected and classified peaks are chirp times
|
|
|
|
Application: Chirps during competition
|
|
- Detected over 10000 chirps in real data from a competition experiment
|
|
- Analysis of the relationship of chirps and competition events
|
|
- Fish competed for a shelter
|
|
- Were able to replicate some findings from literature
|
|
- e.g. loser fish tend to chirp more
|
|
- Other findings are not that clear and require the consideration of more factors, e.g. sex
|
|
- We explored how the chirp rate changes during onsets and offsets of chasing events
|
|
- For some recordings, chirping increased strongly before the offset of a chasing, for some nothing happens
|
|
- The number of chirps during chasings is only elevated for some dyads
|
|
|
|
Conclusion
|
|
- Algorithm can be used to detect chirps
|
|
- We could replicate some literature findings and motivate further examination |