Description: The "ambiguous" label was used to tag any miscellaneous sources of perturbance which may be confused as signal. The identifier hasn’t been used widely. Features include birds flying overhead; helicopter and aircraft; shadows of plumes on buildings (it’s not really a plume, and it’s not our typical representation of a shadow either).
Features / Attributes: The ambiguous identifier is mainly used as a catch-all for other sources of potential noise. If certain subclasses appear as false positives frequently, they may be revisited and further codified as what they actually represent (e.g. aircraft; shadow of a plume etc.)
Source of noise: These ambiguous identifiers will be used to provide extra information about miscellaneous signals that can't be captured by the existing labels. The example shown below in Figure \ref{643439} is the shadow of a plume on a building.
Issues / challenges: The ambiguous identifiers are quite sparse and will most likely have minimal effect on model training, however they have been included where the feature appear pronounced and similar to the plumes found either elsewhere in the image or previous images.
Summary of tags applied
For compilation of the training set, images were reviewed across several days, with a goal to tag 5,000 different examples of plumes; and 1,000 different examples of clouds, lights and shadows respectively. Figure \ref{812798} below show number of plumes, clouds, lights, shadows and other categories identified (represent count of bounding boxes).