Kernel UD considerations

r
biologging
Y2022

Some things to consider before making kernel density estimations.

Author

Miriam Lerma

Published

September 14, 2022

Intro

This post is to exemplify some considerations when calculating kernel density analyses.

Before start calculating kernel density analyses, its useful to consider some sources of error that might change your results.

For the exercises, test data is from masked boobies.
To access the data you have to install the package sula: devtools::install_github(“MiriamLL/sula”)

#devtools::install_github("MiriamLL/sula")
library(sula)
GPS_raw<-(GPS_raw)

To manipulate the data we will use functions from the package tidyverse

library(tidyverse)

For spatial manipulations we will use functions from the packages sp and sf

library(sp)
library(sf)

For creating the polygons of kernel density we will use the package adehabitathr

library(adehabitatHR)

Individuals

Some individuals might drive kernel density calculations in one or other direction as effect of different number of recordings (or days recorded) per individual.

Unpaired

To illustrate this, lets see the calculations using together one individual sampled for 1 day and other for 5 days

Individual one day

ID_1day<-GPS_raw %>%
  filter(IDs=='GPS01') %>%
  filter(DateGMT %in% c('02/11/2017'))

Individual 5 days

ID_5days<-GPS_raw %>%
  filter(IDs=='GPS03')

Unpaired days

Unpaired<-rbind(ID_1day,ID_5days)

Transform to spatial object.

Unpaired<-as.data.frame(Unpaired)
coordinates(Unpaired) <- c("Longitude", "Latitude")
class(Unpaired)

Calculate kernelUD.

UnpairedUD<-kernelUD(Unpaired[,3],h='href')

Obtain polygons.

UnpairedUD95 <- getverticeshr(UnpairedUD, percent = 95, unout = c("m2"))
UnpairedUD50 <- getverticeshr(UnpairedUD, percent = 50, unout = c("m2"))

Here you can check on you polygons visually.

Unpaired95<-st_as_sf(UnpairedUD95)
Unpaired50<-st_as_sf(UnpairedUD50)
ggplot()+
  geom_sf(data = Unpaired95,color='#006d77',fill = "#006d77",alpha=0.3,size=1)+
  geom_sf(data = Unpaired50,color='#5f0f40',fill = "#5f0f40",alpha=0.3,size=1)+
  labs(x = "Longitude", y="Latitude")+
  theme_bw()

Paired

To compare, lets now see the kernel density calculated with these same individuals but recorded at similar number of days (3 days)

ID_01<-GPS_raw %>%
  filter(IDs=='GPS01') %>%
  filter(DateGMT %in% c('02/11/2017','03/11/2017','04/11/2017','05/11/2017'))
ID_03<-GPS_raw %>%
  filter(IDs=='GPS03') %>%
  filter(DateGMT %in% c('02/11/2017','03/11/2017','04/11/2017','05/11/2017'))
Paired<-rbind(ID_01,ID_03)

Transform to spatial object.

Paired<-as.data.frame(Paired)
coordinates(Paired) <- c("Longitude", "Latitude")
class(Paired)

Calculate kernelUD.

PairedUD<-kernelUD(Paired[,3],h='href')

Obtain polygons.

PairedUD95 <- getverticeshr(PairedUD, percent = 95, unout = c("m2"))
PairedUD50 <- getverticeshr(PairedUD, percent = 50, unout = c("m2"))

Here you can check on you polygons visually.

Paired95<-st_as_sf(PairedUD95)
Paired50<-st_as_sf(PairedUD50)
ggplot()+
  geom_sf(data = Paired95,color='#006d77',fill = "#006d77",alpha=0.3,size=1)+
  geom_sf(data = Paired50,color='#5f0f40',fill = "#5f0f40",alpha=0.3,size=1)+
  labs(x = "Longitude", y="Latitude")+
  theme_bw()

As you can see the resulting areas would be different.

To solve this problem, you might want to make sure to have tracking data of similar number of days or recordings.

Intervals

Do you have similar recordings in time?

If some devices have gaps, or record at different intervals, you might underestimate or overestimate specific areas.

For this example, lets see one individuals

GPS01<-GPS_raw %>%
  filter(IDs=='GPS01')

Gaps.

Using the column of hours, lets extract all the recordings after 5 pm.

GPS01$Hour <- as.numeric(substr(GPS01$TimeGMT, 1, 2))
Gaps<-GPS01 %>%
  filter(Hour <= 17)

Transform to spatial object.

Gaps<-as.data.frame(Gaps)
coordinates(Gaps) <- c("Longitude", "Latitude")

Calculate kernelUD.

GapsUD<-kernelUD(Gaps[,3],h='href')

Obtain polygons.

GapsUD95 <- getverticeshr(GapsUD, percent = 95, unout = c("m2"))
GapsUD50 <- getverticeshr(GapsUD, percent = 50, unout = c("m2"))

Here you can check on you polygons visually.

Gaps95<-st_as_sf(GapsUD95)
Gaps50<-st_as_sf(GapsUD50)
ggplot()+
  geom_sf(data = Gaps95,color='#006d77',fill = "#006d77",alpha=0.3,size=1)+
  geom_sf(data = Gaps50,color='#5f0f40',fill = "#5f0f40",alpha=0.3,size=1)+
  labs(x = "Longitude", y="Latitude")+
  theme_bw()

Complete

In contrast, the kernel density calculations without gaps would give different results.

Complete<-GPS_raw %>%
  filter(IDs=='GPS01')

Transform to spatial object.

Complete<-as.data.frame(Complete)
coordinates(Complete) <- c("Longitude", "Latitude")

Calculate kernelUD.

CompleteUD<-kernelUD(Complete[,3],h='href')

Obtain polygons.

CompleteUD95 <- getverticeshr(CompleteUD, percent = 95, unout = c("m2"))
CompleteUD50 <- getverticeshr(CompleteUD, percent = 50, unout = c("m2"))

Here you can check on you polygons visually.

Complete95<-st_as_sf(CompleteUD95)
Complete50<-st_as_sf(CompleteUD50)
ggplot()+
  geom_sf(data = Complete95,color='#006d77',fill = "#006d77",alpha=0.3,size=1)+
  geom_sf(data = Complete50,color='#5f0f40',fill = "#5f0f40",alpha=0.3,size=1)+
  labs(x = "Longitude", y="Latitude")+
  theme_bw()

To solve the problem with gaps, you can interpolate the data to fill the gaps and have similar intervals. However, caution should be taken if you have large gaps, it would create a line.

Behaviour

Do you want to know the general areas that the animal used or just where it was feeding?

It depends on your question, but if you are interested in specific behaviours, for example feeding areas, the kernel density analyses might be bring very different results than when using all movement data.

Foraging

Here, we are using only areas where the animal was foraging.

Load data

GPS_raw<-as.data.frame(GPS_raw)
GPS01<-subset(GPS_raw,GPS_raw$IDs=='GPS01')

Use an specific period

GPS_bc<-recortar_periodo(GPS_data=GPS01,
                                inicio='02/11/2017 18:10:00',
                                final='05/11/2017 14:10:00',
                                dia_col='DateGMT',
                                hora_col='TimeGMT',
                                formato="%d/%m/%Y %H:%M:%S")

Convert to the correct format

GPS_bc$tStamp<-paste(GPS_bc$DateGMT,GPS_bc$TimeGMT)
GPS_bc$tStamp <- as.POSIXct(strptime(GPS_bc$tStamp,"%d/%m/%Y %H:%M:%S"),"GMT")
GPS_bc$lon<- as.numeric(GPS_bc$Longitude)
GPS_bc$lat<- as.numeric(GPS_bc$Latitude)
GPS_bc$id <- as.factor(GPS_bc$IDs)

Keep only the important columns

GPS_bc<-GPS_bc %>%
  dplyr::select('id','tStamp','lon','lat')

Load the package

library(EMbC)

Run the function

BC_clustering<-EMbC::stbc(GPS_bc[2:4],info=-1)

Add the behavioral classifications

GPS_bc$Behaviours<-(BC_clustering@A)

Rename it so you can understand what each behaviour means

GPS_bc<-mutate(GPS_bc, BC = ifelse(GPS_bc$Behaviours == "1", "Resting",
                                        ifelse(GPS_bc$Behaviours == "2", "Intense foraging",
                                               ifelse(GPS_bc$Behaviours == "3", 'Travelling',
                                                      ifelse(GPS_bc$Behaviours == "4", "Relocating", 
                                                             "Unknown")))))

Filter to keep only foraging

Foraging<-GPS_bc %>%
  filter(BC=='Intense foraging')

Transform to spatial object

Foraging<-as.data.frame(Foraging)
coordinates(Foraging) <- c("lon", "lat")

Calculate kernelUD.

Note Here the href is of 0.0048 which is giving the error of subscript out of bounds Lets then better calculate using other h value

#ForagingUD<-kernelUD(Foraging[,3],h='href') 
#ForagingUD95 <- getverticeshr(ForagingUD, percent = 95, unout = c("m2"))

The new h value is of 0.01

ForagingUD<-kernelUD(Foraging[,3],h=0.009) 
ForagingUD

Obtain polygons.

ForagingUD95 <- getverticeshr(ForagingUD, percent = 95, unout = c("m2"))
ForagingUD50 <- getverticeshr(ForagingUD, percent = 50, unout = c("m2"))

Here you can check on you polygons visually.

Foraging95<-st_as_sf(ForagingUD95)
Foraging50<-st_as_sf(ForagingUD50)
ggplot()+
  geom_sf(data = Foraging95,color='#006d77',fill = "#006d77",alpha=0.3,size=1)+
  geom_sf(data = Foraging50,color='#5f0f40',fill = "#5f0f40",alpha=0.3,size=1)+
  labs(x = "Longitude", y="Latitude")+
  theme_bw()

All recordings

Here, we are using all the areas.

GPS_bc<-GPS_bc %>%
  dplyr::select('id','tStamp','lon','lat')
Behas<-as.data.frame(GPS_bc)
coordinates(Behas) <- c("lon", "lat")

Calculate kernelUD.

BehasUD<-kernelUD(Behas,h='href') 
BehasUD

Obtain polygons.

BehasUD95 <- getverticeshr(BehasUD, percent = 95, unout = c("m2"))
BehasUD50 <- getverticeshr(BehasUD, percent = 50, unout = c("m2"))

Here you can check on you polygons visually.

Behas95<-st_as_sf(BehasUD95)
Behas50<-st_as_sf(BehasUD50)
ggplot()+
  geom_sf(data = Behas95,color='#006d77',fill = "#006d77",alpha=0.3,size=1)+
  geom_sf(data = Behas50,color='#5f0f40',fill = "#5f0f40",alpha=0.3,size=1)+
  labs(x = "Longitude", y="Latitude")+
  theme_bw()

If you want to classify the behaviour, please check the post on EmBC