Ranking of mowing results based on confidence

I was wondering if there is some sort of ranking system to fill up the result table if there are more than 4 “mowings” detected. For example the analysis finds 3 mowing events:
|m1_dstart |m1_dend |m1_conf| m1_mis|
|2019-03-30 |2019-04-05 |0,228 | S1|
m2_dstart |m2_dend |m2_conf |m2_mis|
2019-04-28 |2019-05-04 |0,059 |S1
m3_dstart |m3_dend |m3_conf |m3_mis|
|2019-06-29 |2019-07-16 |0,75 |S2/S1|

But only the last one is actually a mowing event, the two first ones are probably misinterpretations. But what happens if the analysis fins 2 more mowing events with a high confidence at a later date. Will the early ones with a low confidence be replaced?

Hello Bastian,

yes, it is correct, in case of other mowing events detected with higher confidence, the early ones could be replaced.

Generally, for each parcel any new mowing detection (up to a maximum of 4 mowing events) is included if:

  • its confidence is in the top 4 most confidence detections
  • it has a temporal distance greater than 30 days with respect to the most confident detections

This means that the algorithm ranks according to a double criteria, confidence and temporal distance, eventually replacing mowing events early detected, but having a lower confidence.

I attached a set of slide explaining the fusion mechanism.


GrasslandMowing_Fusion.zip (939.3 KB)

Hi Laura,
Thanks a lot for the explanation and the power point! It sounds like a very good way of handling the mowing detection.

I have one more question regarding the interpretation of confidence levels. In your document ATBD for L4B grassland mowing detection product 1.2 on page 37 you describe the confidence levels for S1 and S2. S1 ranges between 0 - 0,5 and S2 between 0,5 -1. Does that mean that if I get a results only based on S2 with a confidence of 0,5, its actually a really low confidence but if I get 0,5 for an activity based on S1, its actually a high confidence? I am trying to filter the mowing results and take away those with low confidence, so it would be great to understand better what results I can trust.

Hi Bastian,
your interpretation is right. Based on these principles, you should separate S1 and S2 detections considering to divide detections in two sets the first, related to S1 detections, having confidence values within (0-0.5], the second one, related to S2, with confidences within (0.5-1). Then, you can apply 2 thresholds, one for each set to select the most confident detections.
At the moment, we keep separated the S1 and S2 detection confidence levels because currently we have no elements to assess a relation between the 2 confidences.
The detection algorithm based on S2 exploits almost completely the information of the S2 temporal series and it is currently more reliable of that one based on S1. The S1 algorithm requires further analysis to better exploit the whole information of the S1 amplitude and coherence temporal series and we are working in this direction.

Thank you for the explanation, that certainly helps in analysing our results.
Looking forward to the improvements on the S1 results since we have just too many clouds in Sweden :wink: