Performance assessment and implementation plan for a new DMQC method based on machine learning (for temperature and salinity)
|Garcia Juan Andrea1
|Maze Guillaume, Balem Kevin, Cabanes Cecile, Klein Birgit, Cancouet Romain
This deliverable provides a performance assessment and implementation plan for a new quality control method based on machine learning. This method uses a statistical classifier (a PCM: Profile Classification Model) to organise and select more appropriately reference data for the quality control of an Argo float. The PCM based selection is able to distinguish profiles from different dynamical regimes of the ocean (e.g. eddies, fronts, quiescent water masses). Thus, selecting reference data out of the same dynamical regimes as the Argo float data to be quality controlled ensures more robust and relevant reference statistics.
We show that this new DMQC-PCM method is able to improve the detection of salinity drift and temperature or salinity outliers. We further show that the new method, when combined with the standard salinity calibration method, is able to reduce the error on the correction while preserving confidence in this correction amplitude.
We further provide an implementation plan for the DMQC-PCM method. The goal of this implementation plan is to make the method easily accessible to Argo QC operators. It is based on a collection of fully documented Jupyter notebooks demonstrating the use of the method with only open source software already available online freely. We also provide guidelines for the full integration of the new method in existing salinity calibration software.