K-means clustering of bivariate polar plots

Function for identifying clusters in bivariate polar plots (polarPlot()); identifying clusters in the original data for subsequent processing.

Usage

polar_cluster(
  data,
  pollutant,
  x = "ws",
  wd = "wd",
  n_clusters = 6,
  data_after = NA,
  return = "plot",
  ...
)

Arguments

data

A data frame containing wind direction, wind speed, and pollutant concentrations.

pollutant

One or more column names identifying pollutant concentrations. When multiple pollutants are specified for a single-pollutant statistic (e.g., "mean"), a faceted plot will be returned. Two pollutants must be provided for certain statistic options (e.g., "Pearson" in polar_plot()).

x

Name of variable to plot against wind direction in polar coordinates, the default is wind speed, “ws”.

wd

Name of wind direction field.

n_clusters

Number of clusters to use. If n_clusters is more than length 1, then a faceted plot will be output showing the clusters identified for each one of n_clusters.

data_after

Optional. Data representing the "after" case; see polar_diff() for more information.

return

"plot" (the default) or "data". "plot" will return plotted clusters for visual analysis so that an appropriate value for n_clusters can be selected. When such a value has been chosen, "data" will return the original data frame appended with a cluster column for use in, for example, trend_prop().

...

Arguments passed on to openair::polarPlot

statistic

The statistic that should be applied to each wind speed/direction bin. Because of the smoothing involved, the colour scale for some of these statistics is only to provide an indication of overall pattern and should not be interpreted in concentration units e.g. for statistic = "weighted.mean" where the bin mean is multiplied by the bin frequency and divided by the total frequency. In many cases using polarFreq will be better. Setting statistic = "weighted.mean" can be useful because it provides an indication of the concentration * frequency of occurrence and will highlight the wind speed/direction conditions that dominate the overall mean.Can be:

“mean” (default), “median”, “max” (maximum), “frequency”. “stdev” (standard deviation), “weighted.mean”.
statistic = "nwr" Implements the Non-parametric Wind Regression approach of Henry et al. (2009) that uses kernel smoothers. The openair implementation is not identical because Gaussian kernels are used for both wind direction and speed. The smoothing is controlled by ws_spread and wd_spread.
statistic = "cpf" the conditional probability function (CPF) is plotted and a single (usually high) percentile level is supplied. The CPF is defined as CPF = my/ny, where my is the number of samples in the y bin (by default a wind direction, wind speed interval) with mixing ratios greater than the overall percentile concentration, and ny is the total number of samples in the same wind sector (see Ashbaugh et al., 1985). Note that percentile intervals can also be considered; see percentile for details.
When statistic = "r" or statistic = "Pearson", the Pearson correlation coefficient is calculated for two pollutants. The calculation involves a weighted Pearson correlation coefficient, which is weighted by Gaussian kernels for wind direction an the radial variable (by default wind speed). More weight is assigned to values close to a wind speed-direction interval. Kernel weighting is used to ensure that all data are used rather than relying on the potentially small number of values in a wind speed-direction interval.
When statistic = "Spearman", the Spearman correlation coefficient is calculated for two pollutants. The calculation involves a weighted Spearman correlation coefficient, which is weighted by Gaussian kernels for wind direction an the radial variable (by default wind speed). More weight is assigned to values close to a wind speed-direction interval. Kernel weighting is used to ensure that all data are used rather than relying on the potentially small number of values in a wind speed-direction interval.
"robust_slope" is another option for pair-wise statistics and "quantile.slope", which uses quantile regression to estimate the slope for a particular quantile level (see also tau for setting the quantile level).
"york_slope" is another option for pair-wise statistics which uses the York regression method to estimate the slope. In this method the uncertainties in x and y are used in the determination of the slope. The uncertainties are provided by x_error and y_error --- see below.

exclude.missing

Setting this option to TRUE (the default) removes points from the plot that are too far from the original data. The smoothing routines will produce predictions at points where no data exist i.e. they predict. By removing the points too far from the original data produces a plot where it is clear where the original data lie. If set to FALSE missing data will be interpolated.

uncertainty

Should the uncertainty in the calculated surface be shown? If TRUE three plots are produced on the same scale showing the predicted surface together with the estimated lower and upper uncertainties at the 95% confidence interval. Calculating the uncertainties is useful to understand whether features are real or not. For example, at high wind speeds where there are few data there is greater uncertainty over the predicted values. The uncertainties are calculated using the GAM and weighting is done by the frequency of measurements in each wind speed-direction bin. Note that if uncertainties are calculated then the type is set to "default".

percentile

If statistic = "percentile" then percentile is used, expressed from 0 to 100. Note that the percentile value is calculated in the wind speed, wind direction ‘bins’. For this reason it can also be useful to set min.bin to ensure there are a sufficient number of points available to estimate a percentile. See quantile for more details of how percentiles are calculated.

percentile is also used for the Conditional Probability Function (CPF) plots. percentile can be of length two, in which case the percentile interval is considered for use with CPF. For example, percentile = c(90, 100) will plot the CPF for concentrations between the 90 and 100th percentiles. Percentile intervals can be useful for identifying specific sources. In addition, percentile can also be of length 3. The third value is the ‘trim’ value to be applied. When calculating percentile intervals many can cover very low values where there is no useful information. The trim value ensures that values greater than or equal to the trim * mean value are considered before the percentile intervals are calculated. The effect is to extract more detail from many source signatures. See the manual for examples. Finally, if the trim value is less than zero the percentile range is interpreted as absolute concentration values and subsetting is carried out directly.

cols

Colours to be used for plotting. Options include “default”, “increment”, “heat”, “jet” and RColorBrewer colours --- see the openair openColours function for more details. For user defined the user can supply a list of colour names recognised by R (type colours() to see the full list). An example would be cols = c("yellow", "green", "blue"). cols can also take the values "viridis", "magma", "inferno", or "plasma" which are the viridis colour maps ported from Python's Matplotlib library.

weights

At the edges of the plot there may only be a few data points in each wind speed-direction interval, which could in some situations distort the plot if the concentrations are high. weights applies a weighting to reduce their influence. For example and by default if only a single data point exists then the weighting factor is 0.25 and for two points 0.5. To not apply any weighting and use the data as is, use weights = c(1, 1, 1).

An alternative to down-weighting these points they can be removed altogether using min.bin.

min.bin

The minimum number of points allowed in a wind speed/wind direction bin. The default is 1. A value of two requires at least 2 valid records in each bin an so on; bins with less than 2 valid records are set to NA. Care should be taken when using a value > 1 because of the risk of removing real data points. It is recommended to consider your data with care. Also, the polarFreq function can be of use in such circumstances.

force.positive

The default is TRUE. Sometimes if smoothing data with steep gradients it is possible for predicted values to be negative. force.positive = TRUE ensures that predictions remain positive. This is useful for several reasons. First, with lots of missing data more interpolation is needed and this can result in artefacts because the predictions are too far from the original data. Second, if it is known beforehand that the data are all positive, then this option carries that assumption through to the prediction. The only likely time where setting force.positive = FALSE would be if background concentrations were first subtracted resulting in data that is legitimately negative. For the vast majority of situations it is expected that the user will not need to alter the default option.

k

This is the smoothing parameter used by the gam function in package mgcv. Typically, value of around 100 (the default) seems to be suitable and will resolve important features in the plot. The most appropriate choice of k is problem-dependent; but extensive testing of polar plots for many different problems suggests a value of k of about 100 is suitable. Setting k to higher values will not tend to affect the surface predictions by much but will add to the computation time. Lower values of k will increase smoothing. Sometimes with few data to plot polarPlot will fail. Under these circumstances it can be worth lowering the value of k.

normalise

If TRUE concentrations are normalised by dividing by their mean value. This is done after fitting the smooth surface. This option is particularly useful if one is interested in the patterns of concentrations for several pollutants on different scales e.g. NOx and CO. Often useful if more than one pollutant is chosen.

key.footer

see key.footer.

key.position

Location where the scale key is to plotted. Allowed arguments currently include "top", "right", "bottom" and "left".

ws_spread

The value of sigma used for Gaussian kernel weighting of wind speed when statistic = "nwr" or when correlation and regression statistics are used such as r. Default is 0.5.

wd_spread

The value of sigma used for Gaussian kernel weighting of wind direction when statistic = "nwr" or when correlation and regression statistics are used such as r. Default is 4.

x_error

The x error / uncertainty used when statistic = "york_slope".

y_error

The y error / uncertainty used when statistic = "york_slope".

kernel

Type of kernel used for the weighting procedure for when correlation or regression techniques are used. Only "gaussian" is supported but this may be enhanced in the future.

tau

The quantile to be estimated when statistic is set to "quantile.slope". Default is 0.5 which is equal to the median and will be ignored if "quantile.slope" is not used.

Details

Bivariate polar plots generated using the polarPlot() function provide a very useful graphical technique for identifying and characterising different air pollution sources. While bivariate polar plots provide a useful graphical indication of potential sources, their location and wind-speed or other variable dependence, they do have several limitations. Often, a `feature' will be detected in a plot but the subsequent analysis of data meeting particular wind speed/direction criteria will be based only on the judgement of the investigator concerning the wind speed-direction intervals of interest. Furthermore, the identification of a feature can depend on the choice of the colour scale used, making the process somewhat arbitrary.

polarCluster applies Partition Around Medoids (PAM) clustering techniques to polarPlot() surfaces to help identify potentially interesting features for further analysis. Details of PAM can be found in the cluster package (a core R package that will be pre-installed on all R systems). PAM clustering is similar to k-means but has several advantages e.g. is more robust to outliers. The clustering is based on the equal contribution assumed from the u and v wind components and the associated concentration. The data are standardized before clustering takes place.

The function works best by first trying different numbers of clusters and plotting them. This is achieved by setting n_clusters to be of length more than 1. For example, if n_clusters = 2:10 then a plot will be output showing the 9 cluster levels 2 to 10.

The clustering can also be applied to differences in polar plot surfaces (see polarDiff()). On this case a second data frame (after) should be supplied.

Note that clustering is computationally intensive and the function can take a long time to run --- particularly when the number of clusters is increased. For this reason it can be a good idea to run a few clusters first to get a feel for it, e.g., n_clusters = 2:5.

Once the number of clusters has been decided, the user can then run polar_cluster() to return the original data frame together with a new column cluster, which gives the cluster number as a character (see example). Note that any rows where the value of pollutant is NA are ignored so that the returned data frame may have fewer rows than the original.

Note that there are no automatic ways in ensuring the most appropriate number of clusters as this is application dependent. However, there is often a-priori information available on what different features in polar plots correspond to. Nevertheless, the appropriateness of different clusters is best determined by post-processing the data. The Carslaw and Beevers (2012) paper discusses these issues in more detail.

References

Carslaw, D.C., Beevers, S.D, Ropkins, K and M.C. Bell (2006). Detecting and quantifying aircraft and other on-airport contributions to ambient nitrogen oxides in the vicinity of a large international airport. Atmospheric Environment. 40/28 pp 5424-5434.

Carslaw, D.C., & Beevers, S.D. (2013). Characterising and understanding emission sources using bivariate polar plots and k-means clustering. Environmental Modelling & Software, 40, 325-329. doi:10.1016/j.envsoft.2012.09.005

Usage

Arguments

Details

References

See also