Multi-template detections

[mmesim]

This discussion might also be related to issue #21.

I was able to run a multi-template detection on the same stream using the following json

[{
  "detectorId": "detector-01",
  "createArrivals": true,
  "createTemplateArrivals": false,
  "gapInterpolation": true,
  "gapThreshold": 0.1,
  "gapTolerance": 1.5,
  "triggerDuration": 0.5,
  "triggerOnThreshold": 0.5,
  "triggerOffThreshold": 0.0,
  "arrivalOffsetThreshold":1.0,
  "originId": "smi:ch.ethz.sed/sc3a/origin/NLL.20191105125505.255283.1897990",
  "filter": "BW_BP(2,1.5,15)",
  "templateFilter": "BW_BP(2,1.5,15)",
  "streams": [
    {
      "templateId": "template-01",
      "initTime": 10,
      "templateWaveformStart": -0.5,
      "templateWaveformEnd": 2,
      "waveformId": "8D.RAW2..HHZ",
      "templateWaveformId": "8D.RAW2..HHZ",
      "templatePhase": "Pg"
   }
 ]
},
{
  "detectorId": "detector-02",
  "createArrivals": true,
  "createTemplateArrivals": false,
  "gapInterpolation": true,
  "gapThreshold": 0.1,
  "gapTolerance": 1.5,
  "triggerDuration": 0.5,
  "triggerOnThreshold": 0.5,
  "triggerOffThreshold": 0.0,
  "arrivalOffsetThreshold":1.0,
  "originId": "smi:ch.ethz.sed/sc3a/origin/NLL.20191105115351.959446.1893532",
  "filter": "BW_BP(2,1.5,15)",
  "templateFilter": "BW_BP(2,1.5,15)",
  "streams": [
    {
      "templateId": "template-02",
      "initTime": 10,
      "templateWaveformStart": -0.5,
      "templateWaveformEnd": 2,
      "waveformId": "8D.RAW2..HHZ",
      "templateWaveformId": "8D.RAW2..HHZ",
      "templatePhase": "Pg"
   }
 ]
}]

The resulting .scml file contains detections from both detectors. Specifically, 75 detections from detector-01, which is the same number of detections when the detection is performed on single stream-single template, and 93 detections from detector-02

A question is how can we combine these detections and keep the one from the best detector (template), for example based on Cross-Correlation Coefficient (CCC).

I tried to tune scevent, and particularly minimumMatchingArrivals and maximumMatchingArrivals. However, I'm not sure if scevent works for picks on the same station and chanel as in this case.

I think that an association based on CCC, perhaps combined with event score, might be useful for real time operations.

[damb]

A question is how can we combine these detections and keep the one from the best detector (template), for example based on Cross-Correlation Coefficient (CCC).

@mmesim, in the example above the detector configurations are based on two different template origins. Are these template origins somehow related? Do they belong to a common cluster?

• In case they belong to a single cluster AFAIK the general approach is for performance reasons not to run the actual cross-correlation for all but the best template from a cluster (you wouldn't gain any additional information from using all templates, anyway). Though, this approach requires performing a cluster analysis in advance.
• What metrics do you suggest to use except of the CCC? Simply using the CCC is most probably not enough, since the detections need to be related somehow in time if grouping detections in real-time is required.
• Associating and filtering detections form different detectors would require an additional layer within scdetect. I'd rather prefer a solution where these kind of grouping is performed as part of an additional module.

[mmesim]

@mmesim, in the example above the detector configurations are based on two different template origins. Are these template origins somehow related? Do they belong to a common cluster?

Yes, these are from the same sequence, and probably the same cluster.

What metrics do you suggest to use except of the CCC? Simply using the CCC is most probably not enough, since the detections need to be related somehow in time if grouping detections in real-time is required.

Yes, I agree. I was thinking that first we group the detections based on the origin time. Then if multiple templates detect the same earthquake we could keep the detections made by the best template, meaning the template that has the highest CCC. Sorry if that wasn't clear before.

Associating and filtering detections form different detectors would require an additional layer within scdetect. I'd rather prefer a solution where these kind of grouping is performed as part of an additional module.

I'm sorry but I still do not follow much the technical part/structure of the code. My suggestion would be to try and incorporate the detections' association and filtering in scevent. Still I'm not sure how, and thinking about it.