Data from: Reindeer carcasses modulate vegetation composition and greenness in High-Arctic tundra
Abstract
Milder winters over High Arctic regions have dramatic impacts on local biodiversity on Svalbard, with rain-on-snow events directly correlated with reindeer mortality. Vertebrate carrion can have disproportionately larger impacts on vegetation in nutrient-limited systems, compared with warmer biomes. We conducted a ground survey on the cover of five plant functional groups at paired carcass and control sites and analysed the relationship between cover and carcass presence with generalised linear mixed effect models. Vegetation indices from RGB imagery captured by drones complemented this, assessing plant productivity in terms of ‘spectral greening’. We modelled the relationship between vegetation index values and carcass distance with generalised additive models. We show that graminoids capitalised most from carcass presence, whereas bryophytes and lichen showed decreases in cover. Woody plants and herb covers were not significantly impacted by carcass presence. The Red Green Blue Vegetation Index (RGBVI, our proxy for vegetation productivity) decreased locally at fresh carcasses (i.e. <1 year old) but showed an increase at more established carcass sites (i.e. >1 year). We show that carcasses have differential impacts on the plant functional groups of Svalbard’s tundra and induce a local ‘green-up’ with secondary succession within 2 metres. Given their non-random distribution, carcasses may contribute to vegetation heterogeneity at landscape scales. This is relevant for understanding how climate change-induced reindeer mortalities will impact Arctic tundra composition in the future.
In August 2021, 33 reindeer carcasses were visited in Adventdalen (and surrounding valleys) Svalbard. These carcasses were of varying ages, ranging from approximately half a year to four years. A each carcass site, two primary forms of data collection were conducted, i.e.:
- Vegetation surveys, where a 5m × 5m grid was laid over the carcass site and a nearby control (paired surveys per site)
- Drone imaging surveys over the carcass and its surroundings
Vegetation surveys: The 5 × 5 m grid was subdivided into 1 × 1 m grid cells (subplots), with the central cell over the carcass ‘centre’, i.e. the rumen content/abdomen. A paired control plot was placed 30 – 50 m away from the carcass plot. Total cover (to the nearest 5%) was estimated for five plant functional groups, namely; forbs, graminoids, woody plants, bryophytes, and lichens. This was estimated within each subplot, and at the total plot level.
Drone imaging surveys: The drone images were captured in a survey conducted 70m altitude above the carcass, with 80% image overlap. Images were stitched per survey in Agisoft Metashape Professional. Non-terrain/carcass features such as people/anthropogenic objects were manually masked out of each resulting orthophoto, and then cropped to a 50m radius from the carcass. This was used for further analysis, and these are the images that have been made available here.
Description of the data and file structure
The structure of this dataset are as follows:
- Data - a folder containing the intermediate and processed data files, called on in the analysis
- Scripts - a folder containing the supporting R scripts used for data preparation, analysis, and visualisation
- A Quarto-Markdown file (and associated rendered HTML) file called
Analysis_ProportionalCoverGLM_VegetationIndexGAM.qmd
Within the “Data” folder, there are five primary datatypes. These are:
- Study sites:
33 reindeer carcass sites were used in this study. Their information is stored in the filecarcasscentersV2.gpkg. The centres are used for the drone orthomosaics, where distance from these centres are computed. The only column of relevance here is the column “SITEID” which matches the unique carcass ID to the orthomosaic. The other attributes are autogenerated (fid, IMGID, Altitude) and can be ignored. - Functional groups - proportional covers:
At the 33 sites, paired vegetation surveys at the carcas and a control location were carried out. In these surveys, proportional covers of 5 plant functional groups were estimated in a 5m $\times$ 5m grid, subdivided into 1m $\times$ 1m ‘subplots’ (rounded to the nearest 5%). This resulted in 34 measurements per grid (33 at the sub-plot level, and 1 at the total plot level), for each the carcass and the paired control. This data is stored in the fileproportionalcovers_functionalgroups.xlsx.
There are two tabs within this sheet. The first tab is called “cover” and is comprised of a sheet with each row corresponding to a subplot within the grid, and the proportional cover of a plant functional group. The columns are as follows:- PlotID - a unique ID given to each carcass location, with 3 digits as a numeric ID, followed by the year in which the carcass was recorded
- Type - whether the grid was placed over the carcass or control
- Quadrate - the subplot ID within the overall grid (25 total, numbered from left to right, top to bottom)
- Cadaver - proportion (%) of cadaver material covering the subplot
- Soil - proportion (%) of bare soil seen in the subplot
- Woody - proportion (%) of woody plant cover in the subplot
- Graminoids - proportion (%) of graminoid cover in the subplot
- Bryophytes - proportion (%) of bryophyte cover in the subplot
- Lichen - proportion (%) of lichen in the subplot
- Herbs - proportion (%) of herbaceous plant material in the subplot
- Rocks - proportion (%) of rocky material in the subplot
The second tab is called “tot.cover” and is a sheet with each row corresponding to the same information summarised at the total grid level. Again, the column names refer to the following information:
- PlotID - a unique ID given to each carcass location, with 3 digits as a numeric ID, followed by the year in which the carcass was recorded
- Type - whether the grid covers a carcass plot or a control site
- Tot.Cadaver - proportion (%) of cadaver material at the grid level
- Tot.Bare soil - proportion (%) of bare soil at the grid level
- Tot.Woody plants - proportion (%) of woody plant cover at the grid level
- Tot.Graminoids - proportion (%) of graminoid cover at the grid level
- Tot.Bryophytes - proportion (%) of bryophyte cover at the grid level
- Tot.Lichen - proportion (%) of lichen cover at the grid level
- Tot.Herbs - proportion (%) of herbaceous plant cover at the grid level
- Tot.Rocks - proportion (%) of rocky material at the grid level
- Orthomosaics:
This folder contains (processed) images (.tiffiles, one generated for each site) used in the analysis. This includes a central part of the orthomosaic, covering the first 50m from the carcass centre. The unwanted features (people, rope, research equipment) have been masked out from these images. These are input into further analysis. - Vegetation index selection:
In order to decide which RGB-based index should be applied in this analysis, pixels were sampled from a Sentinel-2 cloud-free composite over the study area. Various vegetation indices were computed for these pixels, and the one with the highest correlation with NDVI was selected. This was done in Google Earth Engine, and the data was then exported to a.csvfile (calledearthengine_vegetationindices.csv). See the Google Earth Engine snapshot for how these pixels were sampled. A copy of this code is also saved in a Javascript file in the “Scripts” folder and is calledArcticScience_CorrelationVIs_NDVI_Sentinel2_2021Adventdalen.js\
The.csvfile contains the pixel values for the following vegetation indices and the columns are named as follows:- NDVI - Normalised Difference Vegetation Index
- RGBVI - Red Green Blue Vegetation Index
- GLI - Green Leaf Index
- MGRVI - Modified Green Red Vegetation Index
- ExR - Excess Red vegetation index
- ExG - Excess Green vegetation index
- ExB - Excess Blue vegetation index
- GRVI - Green Red Vegetation Index
- IKAW - Kawashima Index
- VARI - Visible atmospherically resistence index
- Intermediate data files: \
An intermediate data fileextracttopoints_NAis255.csvis used for analysing the drone imagery and the influence of carcass distance on pixel greenness (as measured by various RGB-derived vegetation indices). This file is created from the script00_VI_data_extraction.RThe attributes of this data table are as follows:- FID - autogenerated row ID
- site ID - Site ID, unique to each carcass (termed “PlotID” in the other files but refers to the same thing, n=33)
- red - pixel red reflectance value
- green - pixel green reflectance value
- blue - pixel blue reflectance value
- mask - pixel mask value (true/NA)
- distrast - distance from carcass center (point defined in
carcasscentersV2.gpkg.) - distclass - categorical variable relatign to which distance bin the pixel falls into (binned in 0.5m increments from carcass centre)
- GLI - Green Leaf Index value
- MGRVI - Modified Green Red Vegetation Index value
- RGBVI - Red Green Blue Vegetation Index value
- ExR - Excess Red vegetation index value
- cell - cell number (from original orthomosaic)
- x - pixel x coordinate, ESPG 32633
- y - pixel y coordinate, ESPG 32633
Code/Software
The analysis was written in in R, with a Quarto extension to generate comprehensible html files of the primary analysis section (.qmd file format).
The “Scripts” folder contains supporting R code for running the primary analysis. These are as follows.
The main analysis script is outlined in the Quarto-Markdown file and rendered in the associated HTML file called Analysis_ProportionalCoverGLM_VegetationIndexGAM, giving a full overview of how the analysis was conducted. Supporting scripts are stored in the “Scripts” folder. These include:
ArcticScience_CorrelationVIs_NDVI_Sentinel2_2021Adventdalen.js: A copy of the Google Earth Engine code used to correlate RGB vegetation indices to NDVI as derived from Sentinel 2.00_ProportionalCover_preparation.R: Script processing the proportional cover data in the .xlsx fileand and making it analysis-ready (for GLMMs).00_VI_data_extraction.R: Script processing the random sampling of the orthomosaics (as stratified by binned distances from the carcass centre) and computing associated RGB (Red-Green-Blue) vegetation indices. This script is what creates an intermediate data fileextracttopoints_NAis255.csvthat is called on in the main analysis script (for the GAM analysis of the drone imaging surveys).zz_functions.R: Custom functions created for GLMM modelling.
33 reindeer carcasses were visited in Adventdalen (and surrounding valleys) Svalbard, in August 2021. Carcasses were of varying ages, ranging from approximately half a year to four years (n = 8 [reindeer dead in 2021] (termed ‘new’); and n = 3 [2020], n = 20 [2019] and n = 2 [2017] (termed ‘old’). There were two forms of data collection.
- Vegetation surveys, where a 5m × 5m grid was laid over the carcass site and a nearby control (paired surveys per site)
- Drone imaging surveys over the carcass and its surroundings
The 5 × 5 m grid was further subdivided into 1 × 1 m grid cells (subplots), with the central cell over the carcass 'centre', i.e. the rumen content/abdomen. The paired control sites were placed 30 – 50 m away from the carcass. Total cover (to the nearest 5%) was estimated for five functional groups, namely; forbs, graminoids, woody plants, bryophytes, and lichens, within each subplot. The subplots were categorised into three ‘bands’ describing their position within the overall grid, i.e. ‘core’ being the centre cell, ‘inner’ being the cells immediately neighbouring the centre cell, and ‘edge’ being the subplots in the outer perimeter. The relationship between these proportional covers ("vegetation composition") and subplot category was modelled with generalised linear mixed effect models.
The drone images were captured in a survey conducted 70m altitude above the carcass, with 80% image overlap. Images were stiched per survey in Agisoft Metashape Professional. Non-terrain/carcass features such as people/rope/research materials were manually masked out of each orthophoto. The orthophoto was then cropped to a 50m radius from the carcass and used for further analysis - these cropped orthomosaics are uploaded here. Stratified random sampling of pixel reflectances from these orthophotos was used to compute vegetation indices, and model the relationship with distance from carcass using generalised additive models.
All data analysis and model creation was carried out in R.