Skip to content

Latest commit

 

History

History
165 lines (131 loc) · 7.49 KB

README_statistics.md

File metadata and controls

165 lines (131 loc) · 7.49 KB
Raw

Statistics in RCAT

How do you specify which statistics to do?

In the STATISTICS section in config_main.ini you specify the statistics to be done, in a Python dictionary structure.

stats = {
	'annual cycle': 'default',
	'seasonal cycle': {'time stat': 'mean', 'thr': {'pr': 1.0}},
	'percentile': {'resample resolution': ['day', 'max'], 'pctls': [90, 95, 99, 99.9]},
	'pdf': {'thr': {'pr': 1.0}, 'normalized': True, 'bins': {'pr': (0, 50, 1), 'tas': (265, 290, 5)}}
	'diurnal cycle': {'dcycle stat': 'amount', 'time stat': 'percentile 95'}
	'moments': {'moment stat': {'pr': ['Y', 'max'], 'tas': ['D', 'mean']}},
    }

The keys in stats are the statistical measures and values provides the settings/configurations to be applied to the specific statistical calculation. The statistics that are currently available in RCAT and their default settings are given in the rcat_statistics.py module. In particular, the default_stats_config function in that module specifies the statistics possible to calculate along with their properties. Many of the properties (or settings) are common for each of the measures, for example resample resolution, thr or chunk dimension, while others may be specific for the kind of statistics.

If you set default as the key value in stats, as is the case for annual cycle in the code snippet above, then (obviously) the default settings will be used. To modify the settings, the key value should be set to a dictionary containing the particular properties to be modified as keys and values with the modified item values.

Common settings and properties

  • vars

String 'variable_name' or list of strings [*variable_name_1*, *variable_name_2*, ...] representing which variables (as defined in variables under SETTINGS in the main configuration file). If set to empty list [] the calculation will be performed for all defined variables. Some statistics are specifically for a certain variable and then this variable is set as default.

  • resample resolution

If you want to resample input data (model or obs) to another time resolution before statistical computation, you can set this property to a list with two items; the first defines the temporal resolution (e.g. 3 hours, day, month, etc) and the second the method resampling used (taking the mean or sum for example). The xarray resample function is applied here which builds on the similar function in the pandas package. For example, resampling to 6 hourly data, taking the sum over intervening time steps would be defined as follows in the configuration file:

'resample resolution': ['6H', 'sum']

The documentation of and available options for the resampling function can be found here (xarray) and here (pandas) (see DateOffset Objects section for frequency string options)

  • chunk dimension

An important feature behind Dask parallelization is the use of blocked algorithms; to divide the data into chunks and then run computations on each block of data in parallel (see Dask documentation for more information). When working with multiple dimension data arrays in Dask you can chunk the data along different dimensions, depending on what kind of calculations you may want to do. For example, when computing seasonal means of data with dimensions (time, lat, lon) it doesn't really matter along which dimension the data is chunked along. However, if you want to calculate time series percentiles for each grid point then chunking should be done in space ('lat', 'lon'). The chunk dimension property has two options; time/space. For example to chunk along time, set 'chunk dimension': 'time'.

  • pool data

Boolean switch (True/False). Set to True if statistics should be done on pooled data, i.e. assembling all the grid points and time steps and then perform calculations. If you want to the pooling to be done over certain sub-regions, then you need to specify these in the regions property in config_main.ini.

  • thr

Thresholding of data. The value of this (None is the default) should be a dictionary with keys defining variables and values an integer of float; e.g.

'thr': {'pr': 0.1, 'tas': 273}

Specific settings and properties

  • stat method

In many of the available statistical calculations, computations can be done using various methods or moments (e.g. mean, sum, std, etc). For example, if calculating the diurnal cycle, one could compute the mean of all values for each time unit in the cycle or another measure such as a percentile value. This can be specified with this property. Default value is mean. To use a percentile, set (for 95th percentile); 'stat method': 'percentile 95'.

  • dcycle stat

In the computation of the diurnal cycle (including harmonic fit) the dcycle stat defines whether to compute magnitudes or frequency of occurrences. For the former set it to 'amount', for the latter to 'frequency'. When calculating frequencies you must also set the 'thr' option, so for each unit of time in the cycle the occurrence above this threshold is calculated.

  • hours (in diurnal cycle)

The value of this property is a list of hours that should be used in the diurnal cycle computation. It might be changed if you want to compare data sets with different temporal resolution (this can also be achieved with the resample resolution option).

  • normalized (in pdf)

Boolean switch. With normalization, the normalized contribution (by the total mean) from each bin interval in the pdf (or frequency intensity distribution) is computed.

  • normalized (in Rxx)

In the Rxx function (see statfuncs.py module) the counts above the threshold is normalized by the total number of values if this property is set to True.

  • moment stat

The moment statistical calculation involve a basic calculation on the data, such as means, sums or standard deviations. It is basically the same as the resample resolution property and the moment stat is set the same way. For example, if you want to calculate the annual maximum of the input data set

'moment stat': ['Y', 'max']

How do you add new statistical methods to RCAT?

The code in RCAT is heavily based on xarray as well as dask. Xarray has been interfaced closely with dask applications so much of the things that can be done in xarray, like many (basic) statistical calculations, are already dask compliant and therefore relatively easy to implement in RCAT. If you would like to include any new such feature, have a look in the rcat_statistics.py script, for example how the implementation of 'seasonal cycle' has been done.

For more elaborate statistics, using for example functions created by the user (using standard numpy/python code), it may be a bit more complex. Xarray has a function called apply_ufunc which allows repeatedly applying a user function to xarray objects containing Dask arrays in an automatic way. See here for some more information.