Filtering observations
======================
We filter observations by wind speed and absolute deviation between model prior and observations, very similar to :footcite:t:`Bruch2025`.
Observation data with low predicted wind speed are discarded.
Similarly, extreme outliers in the concentrations are also masked before the inversion.

Filtering is implemented in :py:mod:`dubfi.fluxes.readobs`.
All filtering parameters must be provided in the configuration.
Use the option :code:`return_flags` to obtain an array of flags that inform why some observations were not used.


Coordinate filtering
--------------------

#. Select observation sites and sampling heights. You can explicitly exclude stations and/or station sampling heights in the configuration (:code:`coordinate_filter.exclude_stations`). You can also set the maximum number of sampling heights used per station (:code:`coordinate_filter.num_samplingh_heights`).

#. Select a time range (:code:`coordinate_filter.start` and :code:`coordinate_filter.end`, usually start to end of the inversion). Some stations may be restricted to certain time periods (:code:`coordinate_filter.station_seasons`).

#. Filter by time of day (:code:`coordinate_filter.start_window` and :code:`coordinate_filter.end_window`). A daily start and end time are used to select observations based on the local mean time at the station. The time window can be specified per station. Time of day must be provided as positive time interval since midnight (e.g., :code:`5 h` is equivalent to :code:`18000 s` and refers to 05:00 local mean time). For example, many inversions use night time observations for mountain sites and afternoon hours for other stations. :footcite:p:`Bergamaschi2015`


Data-based filtering
--------------------

#. Ignore observations that lie sufficiently below the model-predicted far-field concentration (contribution of lateral boundaries).
   The model-observation mismatch is averaged before checking the difference.

#. Ignore outliers for which the absolute model-observation mismatch exceeds a threshold.

#. Ignore when 24h time series of prior and observations is anti-correlated but shows finite standard deviation.

#. Ignore observations that are strongly influenced by the Nord Stream pipeline leaks.
   Strong influence of the pipeline leaks can lead to large uncertainties. :footcite:p:`Harris2025`

#. Ignore observations that are strongly influenced by wildfires.
   Emissions from wildfires are uncertain and vary strongly in space and time, leading to a large modeling uncertainty.
   The method is not optimal to estimate wildfire emissions.

#. Ignore observations with too low model-predicted wind speed.
   Low wind speed is considered an indicator that the observation may not be representative for a large area. :footcite:p:`Bergamaschi2022,Bruch2025`

.. footbibliography::