Welcome to the Machine Learning Challenge

" Learn multipulsed clicks to estimate the range of a cetacean from only one hydrophone "

Detecting and localizing the echolocation clicks of cetacean provides insight into its behavior, but usual methods are limited in range, or costly. In summer 2018, DYNI LIS CNRS Toulon University, demonstrated the first 3D tracking of deep diving cetacean from a five-channel small-aperture hydrophone array fixed under a moving autonomous surface vehicle (the ASV Sphyrna, SeaProven SA). The resulting 3D tracks depict the behavior of the cetacean in the abyss (at 1.2 km deep) during 3 hours, with one position at each pulsed emission of the cetacean (Poupard et al ICASSP 2019 in the challenge URL). Based on this unique data set, we propose to train original mono-channel biosonar transient analysis model.

Then the tasks are to predict with only one channel (sampling rate is 300 kHz, 16 bits),
if the signal is a click,
and/or the range of the whale, 
and/or its  depth.
Any kind of machine learning is allowed. The training set has around 3600 clicks plus 440 non clicks. 
The test set has around 2012 samples. 

All the clicks (train and test sets) are pre-detected and given in wave form in a matrix (a row = a click). 

The annotation for each click of the train set, from line 1 to 4024 = [ X, Y, Z ] (in meter, Z is depth, it is positive (oriented downward), 0 is the surface).
but if it is not a click it is annotated [ -1,-1,-1 ]. 


TASKS :
The task is to estimate the range of the whale = sqrt( X^2 + y^2 + Z^2 ), 
and/or its depth (Z) in meter.
Because some clicks in the test are not direct clicks, a third task is to predict for each signal of the test data 
if it is a click (if yes, answer = 1, if not output = -1).
One can answer only Z or range, in that case organizers compute scores only on the true clicks, removing the predictions computed from non clicks.
You can submit up to 3 runs for each of the task. You can participate to 1, 2 or all the tasks. There are 3 independant rankings.


FORMAT : 
Your prediction will be sent by email to glotin@univ-tln.fr into a .csv matrix named 'monochan_biosonar_tracking_task_runId_yourname.csv', where 'task'= 'detect' or 'range' or 'depth' and runId = 1, 2 or 3. 

Thus each raw is depending of the task :   
[ number the processed line in the data matrix, detection (1 or -1 if not a click), and/or estimated range of the whale (m), and/or estimated Z (m) ].

Your scripts will be sent with your prediction (in Python, Matlab, C...). This will allow to check that the model is full automatic.
Up to three submissions are allowed per challenger.
Ranking of the runs will be based on MSE of the ranges, or of the depth, or the AUC of the click detector.

Deadline: 
15th of May, 18:00, Paris time.

Data, details, updates : 
http://sabiod.univ-tln.fr/workspace/challenges/learn_from_clicks

Content :
- the acoustic data i.e. the click waveform, one per line, centered, high pass filter : waveform_of_the_clicks_TRAIN_TEST_sets.mat (matlab v7 format, 6036 clicks, 1.5 Go).
- this README.txt.
- the  ICASSP paper 2019 describing the experiment. 
- the annotations.txt gives for each click of the training set :  x, y, z.
- private files : rando.mat (shuffle) and the annotation of the test set.

Contact & Submission : glotin@univ-tln.fr
Organizers : Herve Glotin, Maxence Ferrari, Marion Poupard, Pascale Giraudet, DYNI LIS CNRS Toulon university, FR.
We thank SeaProven, CNRS EADM and SABIOD groups, and NanosSpike and SMILES ANRs.