Remote sensing and GPS tracking reveal temporal shifts in habitat use in nonbreeding Black-tailed Godwits
Data files
Oct 25, 2024 version files 9.56 GB
-
AKDES_dry.Rdata
2.36 GB
-
AKDES_wet.Rdata
7.20 GB
-
ctmm.Rproj
205 B
-
habitat_use_df.csv
10.41 KB
-
location_data.csv
3.57 MB
-
README.md
4.64 KB
Abstract
Knowledge of the habitat requirements for migratory species throughout their full annual cycle is necessary for comprehensive species protection plans. By describing seasonal shifts of space-use patterns in a key nonbreeding area, the Senegal Delta (Mauritania, Senegal), this study addresses a significant knowledge gap in the annual cycle of the rapidly declining continental Black-tailed Godwit Limosa limosa.
We fitted continuous-time stochastic-process movement models with GPS location data to describe the core areas used by 22 GPS-tagged godwits over the 2022-2023 nonbreeding period. We mapped key habitat types, such as floodplain wetlands and rice fields, via supervised classification of satellite imagery.
Godwits in the Senegal Delta show a distinct shift in habitat use over the nonbreeding period. The core areas of godwits in the early stages of the nonbreeding period (the wet season) were primarily in natural wetlands and fields with newly planted rice. As the rice crop matured and became too dense, godwits moved towards more recently sown rice fields. Later, as floodwaters receded and rice fields dried out, godwits abandoned rice fields and shifted toward natural wetlands with fewer invasive plants, particularly within the marshes and shallow floodplains of nature-protected areas in the lower Delta.
Synthesis and Applications: Our findings illustrate the shifting importance of natural and agricultural wetlands for godwits at different stages of the nonbreeding season. Protected areas in the Senegal Delta, particularly the Djoudj National Bird Sanctuary (Senegal) and Diawling National Park (Mauritania), are crucial habitats during the dry season as godwits prepare for their northward migration, while rice fields take a key role during the wet season. Conservation efforts should prioritize eradicating invasive plants from the Djoudj and Diawling, as well as promote agroecological management in specific rice production complexes indicated in this study.
https://doi.org/10.5061/dryad.4tmpg4fm3
Description of the data and file structure
Authors:
Taylor B. Craft, Theunis Piersma, Jos C.E.W. Hooijmeijer, Bing-Run Zhu, Malaika D’Souza, Eoghan O’Reilly, Rienk W. Fokkema, Marie Stessens, Heinrich Belting, Christopher Marlow, Jürgen Ludwig, Johannes Melter, José A. Alves, Arturo Esteban-Pineda, Jorge S. Gutiérrez, José A. Masero, Afonso D. Rocha, Camilla Dreef, Ruth A. Howison
Data Description:
This dataset contains godwit GPS location data and habitat classification related to the nonbreeding behaviour of Black-tailed Godwits (Limosa limosa). The tracking data spans between June 2022 and March 2023, focusing on the movements and habitat use of godwits within the Senegal Delta. The study aims to investigate habitat use patterns across the nonbreeding period.
Files and variables
File: ctmm.Rproj
Description: R Project containing relative paths to all RStudio scripts and data used in the study.
File: scripts.zip
Description: When run within the ctmm.Rproj, all scripts are self contained by reading/writing data using relative paths. All scripts contain extensive comments describing the workflow and expected outputs.
Scripts should be executed in the following order:
1) movement_models: calculates godwit core areas between dispersal events. First run ctmm_population_HQXS_2022_settlement.R,* followed by ctmm_population_HQXS_2022_dispersal1.R through ctmm_population_HQXS_2022_dispersal8.R, and lastly ctmm_population_HQXS_2022_combined.R
2) AKDE_HR_LC_overlap.R: calculates overlap between godwit core areas and land cover type
3) AKDE_HR_LC_overlap_wet_dry.R: calculates overlap between godwit core areas and land cover type between seasons
4) stats.R: ANOVA for comparing differences in habitat use by godwits across seasons
5) habitat_use_protected_areas_percentage.R: calculates percentage of godwit core areas within protected areas
latitudinal_plot.R: create plot of godwit latitudinal gps positions over time. This script can be run independently of other scripts.
File: location_data.csv
Description: Bird GPS location data.
Variables
- timestamp: date/time of gps location fix
- location.long: longitude
- location.lat: latitude
- individual.local.identifier: bird name
File: AKDES_dry.Rdata
Description: Autocorrelated kernel density estimate object of the class UD from the ctmm package. Contains AKDE model information used to estimate core areas for 10 birds during the dry season. This data is required in order to run script AKDE_HR_LC_overlap_wet_dry.R.
See R documentation for detailed explanation of model variables at https://ctmm-initiative.github.io/ctmm/reference/akde.html
File: AKDES_wet.Rdata
Description: Autocorrelated kernel density estimate object of the class UD from the ctmm package. Contains AKDE model information used to estimate core areas for 18 birds during the wet season. This data is required in order to run script AKDE_HR_LC_overlap_wet_dry.R.
See R documentation for detailed explanation of model variables at https://ctmm-initiative.github.io/ctmm/reference/akde.html
File: habitat_use_df.csv
Description: Dataset used for ANOVA comparing godwit habitat use between different land covers and seasons.
Variables
- LC: land cover type
- Freq: number of 10 m^2 land cover raster pixels contained within a corresponding AKDE
- bird: individual bird ID
- percent: proportion of AKDE contained within corresponding land cover
- area: area (m^2) of AKDE within corresponding land cover
- Season: dry season or wet season
Code/software
R/RStudio
Used for all analyses within study. All R scripts contain extensive comments describing the workflow and expected outputs.
Packages:
- ctmm: continuous time movement models used for godwit core area estimation
- terra: raster analysis
- raster: raster analysis
- sf: importing and analysis of spatial data
- ggplot2: creating figures
- *dplyr: *manipulation and tidying data
Google Earth Engine
Used for acquiring, processing, and analysis of satellite imagery for land cover classification.
QGIS
Used for creating map layouts of land cover with godwit core area overlays.
Land cover classification
To differentiate between major habitat types utilized by godwits during their nonbreeding period in the Senegal Delta, we generated a land cover map of the study area. Although habitat descriptions of the Senegal Delta have been relatively well documented over the years, existing reports often lack specific spatio-temporal information or are outdated.
Using Google Earth Engine (GEE), we processed Landsat 9 images acquired over the Senegal Delta between October 1st and November 30th, 2022, coinciding with the peak nonbreeding period for godwits in the region. This timeframe also corresponds to when most wet season rice fields exhibit high biomass, offering a distinct spectral signature. The eight available images were cloud-masked using bitwise operators on the Quality Assessment (QA) band of each image. Given that a single image may capture rice at different growth stages across various fields, we added a Normalized Difference Vegetation Index (NDVI) band into the image collection to create a greenest pixel composite (GEE guide at https://developers.google.com/earth-engine/tutorials/tutorial_api_06) which was then mosaicked and clipped to the boundary of the Delta du Fleuve Transboundary Biosphere Reserve, producing the final image for classification.
A total of 1,760 ground truthing points were created from inspection of Landsat 9 and Sentinel 2 imagery. For areas which were not readily-discernable through satellite imagery alone, we cross-referenced several data sources, including multiple field expeditions to the region in Nov/Dec 2016 (Hooijmeijer et al., 2016), Nov 2017 (Hooijmeijer et al., 2017), Jul 2019 (Hooijmeijer et al., 2019b), and Oct/Nov 2019 (Hooijmeijer et al., 2019a), existing land cover maps (Zwarts et al., 2009; Bos et al., 2012), and open source products (ESA WorldCover 2021 v200; Google Earth). No permits were required for ground surveys, except in Djoudj National Park, where we were accompanied by park guides.
Ground truthing points were randomly split into a training set comprising 70% of the points (n = 1,232) and a validation set containing the remaining 30% (n = 528). The training set was utilized to train a random forest (RF) classifier with 50 decision trees. The validation set was employed to assess the classification accuracy of the trained model through an error matrix, which tabulates the instances of correct and incorrect classifications. From this matrix, both overall accuracy and Cohen’s kappa coefficient (KC) were computed. Overall accuracy represents the proportion of correctly classified points, providing a general measure of the model's performance. The Serval plugin in QGIS version 3.22.4 (QGIS Development Team, 2009) was used to manually correct misclassified pixels in post-processing. The full GEE script is available at: https://code.earthengine.google.com/8085a86b503dc907fcf90eeeda411103
Land cover was categorized into eight major classes:
(1) Open water: natural or artificial expanses of water, including lakes, rivers, and reservoirs.
(2) Rice fields: areas under active wet season rice cultivation (July to November).
(3) Uncultivated fields: bare fields primarily used for rice production during the dry season (March to June).
(4) Mixed crop fields: agricultural areas for production of various crops, including sugar cane, millet, tomatoes, and other vegetables.
(5) Cattail stands: dense stands of cattail in mostly permanent freshwater bodies.
(6) Floodplain wetlands: broadly characterized by variably flooded floodplains, shallow water, diverse floodplain/marshland vegetation such as Sporobolus, Scirpus, Cyperus, and Phragmites.
(7) Semi-arid grassland: Sahelian Acacia savanna with fixed dunes and plains of low-lying woody vegetation such as Acacia shrubs, grasses, and fallow fields.
(8) Bare land: arid terrain consisting of dry sandy dunes, terrestrial hypersaline plains devoid of vegetative cover.
Our field observations indicate that godwits abandon rice fields when the vegetation becomes excessively dense, similar to cattail-dominated areas, relocating to more recently planted fields or drainage areas between fields. To explore how godwits move between fields in relation to rice growth stages, we examined satellite imagery captured between June 2022 and February 2023. GPS locations of godwits were plotted onto the nearest temporally matched satellite image to assess their presence/absence at different rice growth stages.
Tracking data
During the 2021 and 2022 spring staging and breeding seasons we captured and fitted 87 adult godwits with GPS/GSM transmitters at six sites:
1) southwest Friesland, The Netherlands
2) the Krimpenerwaard in the province of South Holland, The Netherlands
3) Dümmer Nature Reserve, Lower Saxony, Germany
4) Unterelbe Nature Reserve, Lower Saxony, Germany
5) Rice fields in the middle basin of the River Guadiana in Extremadura, Spain
6) Tagus estuary, Portugal
Godwits were captured using automated drop cages and claptraps at breeding sites (1-4), handheld nets and mist nets at Iberian staging sites (5, 6) and fitted with 4.5 g GPS/GSM transmitters (Hunan Global Messenger Technology, China) using leg-loop harnesses (Senner et al., 2015). Location fix rates varied between 1-6 fixes/day. All tracking data were stored in Movebank (Kays et al., 2022). Tracking data were filtered for individual tracks spanning a single nonbreeding period between June 2022 and March 2023, representing the approximate minimum arrival and maximum departure dates of adult godwits in the Senegal Delta (Verhoeven et al., 2021). As godwits use multiple sites along the extensive floodplains of the Senegal River, we narrowed our filtering to include only those that visited the lower Senegal Delta (16.8951°N, 15.7762°E, -15.5571°S, -16.5372°W), and removed in-flight locations with a ground speed filter over 0 km/h.
The GPS tagging for this study was granted by the national Dutch committee for animal experiments following the Dutch Animal Welfare Act Articles 9 and 11.
Home range estimation
Features commonly seen in animal tracking data such as location autocorrelation, irregular sampling intervals, variable telemetry error, and data gaps violate the assumptions of many conventional statistical frameworks (Fleming et al., 2015). However, homogenization and reducing autocorrelation through restrictive subsampling comes at the cost of diminishing biologically meaningful results (De Solla et al., 1999).
We therefore used a continuous-time stochastic-process (CTSP) modeling framework, based on the workflow described in Calabrese et al. (2016). CTSP models were fitted with GPS locations of godwits in the Senegal Delta using the continuous-time movement modeling (ctmm) package (Version 1.1.0) (Calabrese et al., 2016) in R (Version 4.2.2) (RStudio Team, 2020).
Due to the high location accuracy (5-20 m) relative to the fix rate of our tags, where the movements of godwits exceed the location accuracy between fixes, we applied the default prior of 20 m for GPS error into the movement model. The ctmm framework depends on individuals exhibiting range residency, which we verified by visualizing the autocorrelation structure of each bird. Range residency can be determined through the existence of an asymptote in the semi-variogram (Supporting Information, Fig. S1), which indicates the distance and time beyond which there is no longer any correlation between observations. Meanwhile, deviances from the mean location (within a defined observation period) resulting from migration, dispersals, or other non-resident behavior, typically show up in the variograms as peaks, dips, and the absence of an asymptote (Calabrese et al., 2016).
Godwits typically use multiple core areas throughout the nonbreeding period, oftentimes hundreds of kilometers away from their initial settlement area (Verhoeven et al., 2021). We therefore segmented tracks of each individual bird between different core areas.
After track segmentation and range residency verification, we fitted movement models for each bird by first providing an initial movement model parameter estimate. Next, the guesstimated model was iterated and the movement process with the lowest Akaike Information Criterion (AICC) value was selected. All models were fit using perturbative hybrid residual maximum likelihood (pHREML) to account for low sample size (Fleming et al., 2015). The ctmm framework classifies a range of CTSP models which are described in detail in (Calabrese et al., 2016). Once the best fitting model was selected, we used area-corrected autocorrelated kernel density estimates (AKDEs) to calculate 50% utilization distributions (UDs) representing the core area (Calabrese et al., 2016) for each bird with 95% confidence intervals. While testing various UD thresholds, we found that higher UDs included unlikely godwit habitats, while lower UDs underestimated their area requirements. We explored the relationship between the rice cultivation cycle and godwit habitat use by separating core areas into (1) wet season core areas (July-November) and (2) dry season core areas (December-March). AKDEs of individuals with multiple core areas within a season were combined.
Next, we calculated the proportion of habitat within each bird's core areas for the different seasons by extracting land cover values within core areas of each bird using the sf (Pebesma and Bivand, 2023), terra (Hijmans, 2024b), and raster (Hijmans, 2024a) packages in R version 4.2.2 (RStudio Team, 2020). The differences of core area overlap proportion (hereafter referred to as “occupancy”) between floodplain wetlands and rice fields were tested using a 2-way Analysis of Variance (ANOVA) with a seasonal interaction term. Pairwise comparisons were assessed using a post-hoc Tukey HSD test.