Characterization

For extreme events Classification, the toolbox provides a post-processing step of data characterization. Characterization is a methodology for summarising and describing the characteristics of a priori aggregated locations that constitute the events. When activated, this post-processing stage provides a .txt file with the following statistics:

Extent or the number of locations (pixels) covered by the event.

Centroid and weighted Centroid.

Maximum, minimum and mean probabilities provided by the model in the event region.

Input Parameters

To use this funcionality, use the following code snippet in your configuration file:

characterization:
    activate: True
    params:
        time_aggregation: ...
        min_distance: ...
        remove_scant: ...
        min_area_holes: ...
        min_area_objects: ...
        threshold:
            Metric: {...}
        threshold_lower_is_best: ...

time_aggregation: Aggreggate through time, if test samples are taken chronologically ,type: bool
min_distance: Maximum Euclidean distance between centroids to be connect, to deactivate it, set to zero (default), type: int
remove_scant: Remove scant labels by filling small holes and small objects, type: bool
min_area_holes: Size of the holes to remove when remove_scant set to True, type: int
min_area_objects: Size of the objects to remove when remove_scant set to True, type: int
threshold > Metric: Metric to optimize the threshold, by default takes a value of 0.5, type: TorchMetric
threshold_lower_is_best: Sets the direction that means improvement, type: bool

Outputs

Example of .txt file generated after executing the Characterization module: