examples_precipitation
A typical workflow for the quality control of precipitation might be: first apply a buddy event test or a sct dual test. Both tests answer the question "what is the probability that it is raining, or not raining here?" Then, combine this test to a buddy test and an isolation test using "further test only unflagged values". The isolation test aims to identify the stations that have too few neighboring stations, where spatial statistical quality control may be limited.
When using "Apply test (no combination / first test)" twice with different tests or different parameters, it is possible to visually compare the results. The observations that get a different flag from the previous test get an orange circle. Also flagging statistics are indicated in the menu (Stations: total | removed | new flagged | new unflag).
SCT dual is meant to recognize if there is an organized spatial pattern for the "no rain" versus "rain" events, while the buddy event does not take into account the spatial organization of the data in a given neighborhood. Thus SCT dual may reduce the risk of erroneously removing small-scale intense precipitation.
The precipitation distribution can be more normalized by applying a Box-Cox transformation (with special treatment of the 0 values). This reduces the risk of erroneously removing small-scale intense precipitation. The Box-Cox power can be chosen in the interface (a value around 0.5 is advised).
Clicking the [BoxCoxObs] button before running the test will also display the transformed Box-Cox values below the original observation values.
Note: the label for the Box-Cox power slider was changed, as this slider can now be used both to setup the Box-Cox power or the started Box-Cox power (see below).
A variant of the Box-Cox transformation is proposed, called "started Box-Cox" transformation. Values smaller than the scaling value are scaled linearly, and values larger than the scaling value are transformed with a scaled box-cox transformation (definition described here). To use a "started Box-Cox" transformation instead of a Box-Cox transformation, one has to set a scaling that is not 0, and choose a power value that is above or equal to 0. For instance power 0.5 and scaling 1. Just like with the Box-Cox transformation, the transformed values by the "started Box-Cox" can be visualized by clicking the [BoxCoxObs] button before running the test.
When suspecting large errors in the dataset it can be beneficial to run a buddy test twice. During the first run the statistics will be influenced by erroneous values, while the second run will flag smaller variations as shown below. One can also run several buddy check with different parameters using "further test only unflagged values".
- Would you like to flag these precipitation observations 21 mm and 26 mm?
- If yes, why?
- If no, why?
- Can you find parameters for tests that would flag these observations?
- Which additional data or additional weather information would you need to determine if they are "good" or "bad" observations?
- Add fake precipitation data and rerun the tests!
