1 - Preparing data for Segmentation/Clustering with segclust2d

R. Patin

2024-04-24

library(segclust2d)
data(simulshift)
data(simulmode)
simulmode$abs_spatial_angle <- abs(simulmode$spatial_angle)
simulmode <- simulmode[!is.na(simulmode$abs_spatial_angle), ]

This summary provides information on:

Type of data accepted and content required to run segmentation() or segclust()

Right now, the function in segclust2d package accept three different kind of input data:

Future version may provide support for sftraj objects as well.

data.frame

data.frame is the format natively supported by segmentation() and segclust(). If x_data.frame is a data frame, the syntax is simply:

segmentation(x_data.frame, lmin = 5, Kmax = 25)

Move

Move object can alternatively be provided to the function. If using segmentation(), the user may omit seg.var argument and the algorithm will use the movement coordinates as segmentation variables. Alternatively if the user specifies the segmented variable with argument seg.var, those variables must be present in the data associated to the Move object x_move@data If x_move is a Move object, the syntax is simply:

segmentation(x_move, lmin = 5, Kmax = 25)

ltraj

ltraj object can alternatively be provided to the function. If using segmentation(), the user may omit seg.var argument and the algorithm will use the movement coordinates as segmentation variables. Alternatively if the user specifies the segmented variable with argument seg.var, those variables must be present in the data associated to the ltraj object x_ltraj@data If x_ltraj is a ltraj object, the syntax is simply:

segmentation(x_ltraj, lmin = 5, Kmax = 25)

sftraj

sftraj objects are not supported for the moment.

Limitations on data size and subsampling

Computation cost for the algorithm scales non-linearly and can be both memory and time-consuming. Performance depends on computer, but from what we have tested, a segmentation on data of size > 10000 can be quite memory intensive (more than 10Go of RAM) and segmentation-clustering can be quite long for data > 1000 (few minutes to hours). For such dataset we recommend either subsampling if loosing resolution is not a big deal (looking for home-range changes over a year with hourly points might be a lost of time when daily points are sufficient) or splitting the dataset for very long data. Although for segmentation-clustering, clusters will not be easily comparable between the different part of the dataset, if one provides parts where all cluster are present for sure, there should be no problem.

Subsampling options

Disabling automatic subsampling

Subsampling is automatically enabled in the function to avoid unwanted memory saturation or very long computation time. By default argument subsample is set to TRUE. In order to totally disable subsampling you have to provide argument subsample :

shiftseg <- segmentation(simulshift,
                         Kmax = 30, lmin=5, 
                         seg.var = c("x","y"), 
                         subsample = FALSE)

mode_segclust <- segclust(simulmode, 
                          Kmax = 30, lmin=5, ncluster = c(2,3,4),
                          seg.var = c("speed","abs_spatial_angle"),
                          subsample = FALSE)

Automatic subsampling

By default subsampling is allowed (subsample = TRUE) and subsampling will occur if the number of data exceed a threshold (10000 for segmentation, 1000 for segmentation-clustering). The function will subsample by the lower factor (by 2, 3, 4…) for which the dataset will fall below the threshold once subsampled. For instance a 2500 rows dataset for segmentation-clustering would be subsampled by 3 to fall below 1000 rows. The threshold can be changed through argument subsample_over.

shiftseg <- segmentation(simulshift,
                         Kmax = 30, lmin=5, 
                         seg.var = c("x","y"), 
                         subsample_over = 2000)

mode_segclust <- segclust(simulmode, 
                          Kmax = 30, lmin=5, ncluster = c(2,3,4),
                          seg.var = c("speed","abs_spatial_angle"),
                          subsample_over = 500)

Manual subsampling

One can also override this automatic subsampling by selecting directly the subsampling factor through argument subsample_by.

shiftseg <- segmentation(simulshift,
                         Kmax = 30, lmin=5, 
                         seg.var = c("x","y"), 
                         subsample_by = 60)

mode_segclust <- segclust(simulmode, 
                          Kmax = 30, lmin=5, ncluster = c(2,3,4),
                          seg.var = c("speed","abs_spatial_angle"),
                          subsample_by = 2)

Consequences of subsampling on lmin

Beware that subsampling will also affect your lmin argument. If subsampling by 2, lmin will be divided by 2. The function will tell about the value of lmin and its adjustment with subsampling with different messages:

#> ✔ Using lmin = 240
#> ✔ Adjusting lmin to subsampling. 
#> Dividing lmin by 60, with a minimum of 5
#> → After subsampling, lmin = 5. 
#> Corresponding to lmin = 300 on the original time scale

Best practice with subsampling

In addition to reducing computation time, subsampling may also help the algorithm. Considering movement at the scale of hours when looking for home-ranges at the scale of months may blur the signal and for such analysis, one data per day may be sufficient. For all analyses, the user should think about the appropriate temporal resolution, with the idea that the finest temporal resolution may not always be appropriate.

Note that subsampling has been implemented in such way that outputs will show all points but segmentation is calculated only on subsampled points. Points used in segmentation can be retrieved through augment in data column subsample_ind (The subsample indices for kept points and NA for ignored points).

Outputs may be more easily explored if subsampling is done before providing the data to segclust2d functions

Covariate calculations

The package also includes functions in order to calculate unusual covariates, such as the turning angle at constant step length (here called spatial_angle, see Patin et al. 2020 for more details). For the latter, a radius have to be chosen and can be specified through argument radius. If no radius is specified, the default one will be the median of the step length distribution. Other covariates calculated are : persistence and turning speed (v_p and v_r) from Gurarie et al (2009), distance travelled between points, speed and smoothed version of the latter. Covariates dependent on time interval (like speed) are by default calculated with hours, but you can change this with argument units as in the example below.

simple_data <- simulmode[,c("dateTime","x","y")]
full_data   <- add_covariates(simple_data, 
                              coord.names = c("x","y"), 
                              timecol = "dateTime",
                              smoothed = TRUE, 
                              units ="min")
head(full_data)

Advice on data pre-processing

When pre-processing movement data before segmentation/clustering it is common to interpolate missing data points. This may however cause problem if this leads to repetition of values. This can also arise if the individual has a very stable speed (i.e. a boat or a bird deriving on the sea) leading to very similar values.

When the repetition of identical or very similar values are longer than parameter lmin, there are segments with null variance, which cannot be accounted for by the algorithm. Should such cases arise, the algorithm will fail and tell you about it:

df <- data.frame(x = rep(1,500), y = rep(2, 500))
segclust(df, 
         seg.var = c("x","y"),
         lmin = 50, ncluster = 3 )
#> ✖ Data have repetition of nearly-identical values longer than lmin. 
#> The algorithm cannot estimate variance for segment with repeated values. This is potentially caused by interpolation  of missing values or rounding of values.
#> → Please check for repeated or very similar values of x and y

To avoid this problem interpolation should be done rather on the covariates to be segmented rather than the coordinates. Alternatively small and rare gaps of data could be ignored. If the gap is too large there is also the possibility to split the dataset.

Dataset with naturally occurring repetition of similar values (a boat at constant speed) are generally difficult to process with our segmentation/clustering algorithm.