Postprocessing: filter those too-close phase picks and save just one of them

[JUNZHU-SEIS]

Similar to PhaseNet, a minimal separation could be set to filter those picks too-close to each other and only save the most significant one.
More details are in lines 26-28 from https://github.com/AI4EPS/PhaseNet/blob/master/phasenet/detect_peaks.py

[yetinam]

Hey @JUNZHU-SEIS,

I agree, that would indeed be a good helper function. Maybe it could go into seisbench.util? There could also be a weighting factor between phase types because different phases tend to have different confidences.

Would you be interested in writing a PR for this?

[JUNZHU-SEIS]

Hi, @yetinam
Sorry for the delayed response.
Here I provide a function to achieve "removing too close picks". Hope it useful and can be adopted in the seisbench.util.

def remove_too_close_detections(detections, charfct, win=100):
# parameter description:
#               detections: An ndarry of indices of P picks or S picks. Do NOT use both the P and S picks.
#               charfct: Probability vector of the corresponding phase (i.e., model output). It is also an ndarry object.
#               win: the minimal separation between consecutive picks
# return:
#               uniqs: a subset of the detections, with close picks removed and highest picks saved
#               deleted_pool: a list of the removed picks

    deleted_pool = []
    uniqs = set()
    for i,win_s in enumerate(detections[:-1]):
        # if the current pick (win_s) is in the list of deleted_pool (see below for how the pool is defined), skip the loop
        if win_s in deleted_pool: continue
        win_e = win_s + win # the window following the current pick with a length of win
        densedetections = detections[i+1:][detections[i+1:]<win_e] # find picks in the window [win_s+1: win_s+win]
        if len(densedetections):
            # if there exist more than one pick in the window, only save the pick with the highest probability and delete the others
            all = np.insert(densedetections, 0, win_s)
            deleted_pool.extend(all)
            highestpeak = all[np.argmax(charfct[all])]
            uniqs.add(highestpeak)
            deleted_pool.remove(highestpeak)
        else:
            # otherwise, if no pick in the window, save the current pick 
            uniqs.add(win_s)
    deleted_pool = list(set(deleted_pool))
    uniqs = np.sort(np.array(list(uniqs)))
    return uniqs,deleted_pool