Apparent problem with ncell

[laurapaccini]

Hi,
I've been trying to apply tobac to my simulations of precipitation, but I have some doubts about "ncells" output.
I only use one threshold with target 'maximum', which are the same in the feature detection and segmentation functions. I also set 'n_min_threshold' to be relatively large (e.g 200) so it only detects features with a minimum of ncells. However, after performing the segmentation function I get many values with ncells << than my 'n_min_threshold, and even with zero values. The only option to avoid the ncell=0 is by setting position_threshold='extreme', but in that case the number of cases with ncells >200 is less than with the other position_threshold options (e.g weighted_diff).

Here is an example of such case (ncell =0 ) with every option except with position_threshold='extreme':

The other issue is that the mask output from segmentation does not always coincide with the ncells. I saw this especially when the ncells are < 200. For example, for a given timestep the dataframe shows features with ncells = 21. However, when looking at the mask, some are absent, others have sizes much larger than those correspondent to 21.

Here the plot correspond to features 30 and 31:

I created a repository in my account with the scripts and subsets of my data in case you want to reproduce these results. (https://github.com/laurapaccini/test_tobac)

[JuliaKukulies]

Thank you for the careful report of this, @laurapaccini!

The first issue is a behavior of the segmentation in tobac that we have recognized before and described in more detail on our documentation page: https://tobac.readthedocs.io/en/latest/features_without_segmented_area.html As you found out correctly, this behavior can be avoided when setting position_threshold='extreme'. Alternatively, you could set the segmentation threshold slightly higher than your feature detection threshold. This is because the feature detection will detect features that are equal to or larger than your threshold, while the segmentation procedure only includes areas larger than the segmentation threshold.

The second issue looks like a potential bug to me. I will have a closer look into your example notebook and move this question to our issues !

[JuliaKukulies]

OK @laurapaccini, I could reproduce your results and found that the second issue is related to a bug that assigns the value for ncells of the last feature in the timestep to all features in that timestep. I raised an issue for that (see #246) and we will fix this bug asap. Thanks again for reporting this!

[laurapaccini]

Changing the threshold value sounds like a good option! Thank you as well for looking into the segmentation issue!