Enhancing Animal Movement Analyses: Spatiotemporal Matching of Animal Positions with Remotely Sensed Data Using Google Earth Engine and R

Round 1

Reviewer 1 Report

The ms presents an algorithm for efficiently matching movement data with satellite raster data.  This is useful because it may enable more people to conduct spatial analysis of animal tracking data over wide areas.  The ms presents two case studies of the algorithm being used, but there is no experiment as such.

My main concern is that the ms confuses introduction and method.  Para 60-73 appears to be method, while 128-141 and 155-164 are introduction.  The introduction should include a general paragraph on the appliction of the algorithm to identify more specifically how this algorithm is an improvement (for collecting raster data from a large area? for adding the temporal dimension to spatial data?).  And then, identify a range of datasets that could be used (dem? vegetation height by VOD? regional vegetation height by lidar?)

The analysis of NDVI appears hasty.  For example, the two graphs associated with the elephant show a seasonal pattern in NDVI (more green in Spring) but not that there was a causal relationship.  In other words, there will be more growth everywhere in Spring regardless of where the elephant is.  They also don't account for the seasonal variation in the relationship between NDVI and biomass (in grasslands, if there's lots of tall dry grass over the new green shoots then NDVI will be low even if there's lots of feed)

In conclusion, I'd say that the algorithm is good and deserves publication, but the ecological implications presented are rather poor.  If this is strengthened, particularly in the introduction and abstract, then the ms may have more impact after publication

A few minor things:

19          ecology cannot observe

19, 38   Exponential is a mathematical term often used colloquially to mean ‘rapid’.  In scientific publications it should only be used for the former

184        ran

239        need for

Author Response

Reviewer comment:

The ms presents an algorithm for efficiently matching movement data with satellite raster data.  This is useful because it may enable more people to conduct spatial analysis of animal tracking data over wide areas.  The ms presents two case studies of the algorithm being used, but there is no experiment as such. 

My main concern is that the ms confuses introduction and method.  Para 60-73 appears to be method, while 128-141 and 155-164 are introduction.  The introduction should include a general paragraph on the appliction of the algorithm to identify more specifically how this algorithm is an improvement (for collecting raster data from a large area? for adding the temporal dimension to spatial data?).  And then, identify a range of datasets that could be used (dem? vegetation height by VOD? regional vegetation height by lidar?)

Response: We appreciate the reviewer comment on the usefulness of the coding workflow.  We have moved the code workflow section to the top of materials and methods to highlight the importance of this section. We then expand on the case studies that implement the code workflow. We hope the edits help clarify this issue.  Lines 60-73 introduces Google Earth Engine to the reader, as it is an unfamiliar platform for many ecologists. Lines 128-141 and 155-164 are specific to each case study. It is also important to stress that the goal of the manuscript is to communicate the novel workflow that allows users to extract spatiotemporal information from remotely-sensed data rapidly and efficiently. It is not the goal of the manuscript to investigate any relationships between animal movement and covariates. We only use the case studies to show that the code workflow works and that it has the potential to enhance movement analysis. 

Reviewer comment:

The analysis of NDVI appears hasty.  For example, the two graphs associated with the elephant show a seasonal pattern in NDVI (more green in Spring) but not that there was a causal relationship.  In other words, there will be more growth everywhere in Spring regardless of where the elephant is.  They also don't account for the seasonal variation in the relationship between NDVI and biomass (in grasslands, if there's lots of tall dry grass over the new green shoots then NDVI will be low even if there's lots of feed)

Response: Thank you for the comment. As mentioned in the previous response, the case study was specifically conducted to showcase the usability of the code to extract NDVI information by matching the closest MODIS image to the time at which the GPS animal position was recorded. It was not the goal of this study to investigate any causal relationship. How animal movement relates to NDVI values is beyond the goal of our work.

Reviewer comment:

In conclusion, I'd say that the algorithm is good and deserves publication, but the ecological implications presented are rather poor.  If this is strengthened, particularly in the introduction and abstract, then the ms may have more impact after publication

Response: We hope we were able to clarify the issues raised and that the edits we have made on the manuscript, based on the feedback from the three reviewers, helped improve the study.

A few minor things:

Reviewer comment:

19      ecology cannot observe

Response: We modified it to “Movement ecologists have observed”.

Reviewer comment:

19, 38   Exponential is a mathematical term often used colloquially to mean ‘rapid’.  In scientific publications it should only be used for the former

Response: We changed exponential for rapid. 

Reviewer comment:

184    ran

Response: Done.

Reviewer comment:

239    need for

Response: Done.

Reviewer 2 Report

see attached file

Comments for author File: Comments.pdf

Author Response

Reviewer comment:

​​The paper presents a demonstration of the practical use of remote sensing spatial data with GPS locations of tracked animals : this capability opens up new research opportunities on the behavior of animals with different environmental constraints.

The introduction could mention more references about recent advances in the application of remote­sensing data to studies of terrestrial animal movement [cf Neumann et al. (2015) and Remelgado et al. (2018)]. While satellite and animal movement data are used in combination extensively, the conceptual differences in scale (temporal and spatial) between these data sources have not often been addressed objectively; Neumann et al. (2015) concluded that «it is important to critically weigh which remotely-sensed product is best suited for a given scale of analyses (e.g.? fine-but- infrequent versus coarse-but-freguent or minimum mapping unit) and how its limitations may affect inferences.

Response: Thank you very much for your comment and the references provided. We have incorporated both references in the introduction and discussion. We agree that discussing the limitations of remote sensed information in animal movement research is important. Therefore, we included that aspect in the discussion.

Line 277: Our workflow is efficient in terms of time and its simplicity with no need for downloading imagery or relying on powerful computer processors. Likewise, this workflow can be integrated into the most common analyses of movement data applicable in conservation and management of animals such as resource selection [42], step selection functions [43,44] to investigate questions related to how animals move and select habitat across space and time. Selecting the appropriate remotely-sensed product, however, with temporal and spatial resolutions that are relevant to the specific research question and analysis methodology remain a central component of any movement analysis [3].

Reviewer comment:

The 2 choosen examples illustrate this compromise between spatial and temporal resolutions.

In the first example 16-day MODIS NDVI time series with 250m spatial resolution data are used and compared with mean annual NDVI. With the aim of reducing the time difference between each GPS point recorded and the recorded NDVI image composite date, it would be interesting to discuss the possible use of higher temporal resolution data like daily MODIS or VIIRS NDVI products available in ESPA (USGS EROS) with 500m resolution, or even daily MODIS NDVI with 250m resolution (cf Zeng et al.9 2021). Optimal spatial resolution must be also discussed with reference to landscape patterns of studies sites.

Response: Again, thank you for the reference. The limitation of higher temporal resolution data is the amount of pixels with no data due to cloud cover or other pixel quality issues that are reduced in 16 day composite products. However, new work like Zeng et al. 2021 may make better quality daily data available soon. We have incorporated this topic into the discussion.

Line 262: In this study we present examples incorporating two environmental data layers that are recognized for their importance in explaining animal movement and behavior [5,18,28]. The code can be adapted to incorporate other data layers of interest. For instance, extracting snow cover or precipitation from the same ERA5-Land data product can be accomplished by simply selecting different bands from the image collection. Alternatively, other image collections available via GEE such as surface water availability [39] and other vegetation indices (e.g., VIIRS NASA Vegetation Index Product) can be substituted in the code. The step-by-step tutorial presented in the supplemental materials explains how to use the code and adapt it to other datasets. While clouds limit the possibility to use vegetation indexes with higher temporal resolution as they can result in images with large areas lacking information, novel advances for reconstructing daily MODIS NDVI products are opening new possibilities for incorporating these data sets into movement analysis in the near future [40]. Moreover, the flexibility of GEE makes it possible to incorporate information derived from multispectral imagery such as Landsat or Sentinel as potential covariates in models for animal movement data [41].

Reviewer comment:

The second example presents high temporal air temperature data (hourly data) with coarse spatial resolution (9km). Spatial pattern of air temperature should be more precisely described, with reference to spatial patterns of land cover and topography. Homogeneity of 9 km x 9 km pixels could be assessed and discussed using higher resolution data, like daily-1km surface air temperature (cf. Wen et al.. 2020).

Response: The goal of our study was simply to demonstrate how users can temporally match ecological data with remotely sensed data products available in Google Earth Enginer. Although datasets with finer spatial resolution may also be preferable for many specific applications, particularly for variables such as temperature that can vary greatly within a 9 x 9 km area, we believe a more detailed treatment of spatial scale is beyond the scope of our study.  However, we did include a discussion on the limitations of remote sensed data as suggested in a previous comment. See our previous response.

In this particular case study, we did correlate temperature data from the remote sensing product with temperature recorded by the collar of the animals to evaluate the accuracy of the remote sensing information. We obtained a high correlation (pearson = 0.896, p < 0.0001), suggesting that the remotely sensed data, even at coarse spatial resolution, is highly accurate in capturing the variation in temperature experienced by the animals.

Reviewer 3 Report

Please see the attachment. 

Comments for author File: Comments.pdf

Author Response

Reviewer comment:

Dear Authors:

I very much enjoyed reading the manuscript. The paper presents a technique to obtain key environmental parameter observations for wildlife (e.g., vegetation cover and temperature) from publicly available remotely sensed data products by leveraging Google Earth Engine and R statistical computing language. Although concerns of biodiversity loss and importance of wildlife conservation have been known topics over decades, research has not advanced as much as we wish for. Data scarcity or limited data access due to the high cost and manpower the main causes along with a weak market pull. The presented technique is very innovative and could make a significant change in the number and advancement of wildlife research.

The manuscript is generally well written and has a clear objective. However, it could be further improved and increase the contributions to the community. My comments and suggestions are summarized below:

Response: Thank you very much. We are pleased that you see the value of our work. Following are our responses to all comments, followed when appropriate, by the new edited text in the manuscript.

Reviewer comment:

1. The paper appears to be written for individuals who have fairly good understanding of GEE and R but not for those who would be inspired this paper and would be interested in utilizing the workflow (e.g., ecologists) for their research.

Response: Thank you for the comment. Most analyses in animal movement are conducted using R programming language. R is also the most popular software for data analysis in the field of ecology (see Lai et al. 2017). Thus, we assume readers of this manuscript have a basic understanding of R. We edited the introduction to clarify this issue.

Lai, J.; Lortie, C.J.; Muenchen, R.A.; Yang, J.; Ma, K. Evaluating the popularity of R in ecology. Ecosphere 2019, 10, 1–7, doi:10.1002/ecs2.2567.

Line 54: Similarly, new statistical methods to analyze and visualize movement data have become more accessible. For instance, there are at least 58 different packages developed for use in the R programming language (https://www.r-project.org)[5,6], one of the most popular open-source programs for data analysis among ecologists [8].

Reviewer comment:

1. For the title, the word, “boosting” may be catchy but too general. The phrase “spatiotemporal extraction of remotely sensed data…” does not seem right; maybe “extraction of spatiotemporal information from remotely sensed data… One suggested title would be “Extraction of spatiotemporal information from remotely sensed data using Google Earth Engine and R for animal movement analyses.”

Response: We modified the title. We would like to keep the original title format, but we replaced ‘Boosting’ for ‘Enhancing’. We think that our work contributes to facilitating extraction of better spatiotemporal covariates. Thus, we like the idea of enhancing animal movement analysis.

Enhancing animal movement analyses: Spatiotemporal matching of animal positions with remotely sensed data using Google Earth Engine and R.

Reviewer comment:

1. In Abstract, I would encourage including a brief summary of results report reported in the Results section to provide examples of the code workflow can be used for animal movement Research.

Response: We included some extra results as suggested. We want to emphasise though that the focus of the manuscript is the process of extracting covariates and not the covariate or case studies per-se. Thus, we put emphasis on the processing times to extract covariates.

Line 20: Abstract: Movement ecologists have witnessed a rapid increase in the amount of animal position data collected over the past few decades, as well as a concomitant increase in the availability of ecologically-relevant remotely sensed data. Many researchers, however, lack the computing resources necessary to incorporate the vast spatiotemporal aspects of datasets available, especially in countries with less economic resources, limiting the scope of ecological inquiry. We developed an R coding workflow that bridges the gap between R and the multi-petabyte catalogue of remotely sensed data available in Google Earth Engine (GEE) to efficiently extract raster pixel values that best match the spatiotemporal aspects (i.e., spatial location and time) of each animal GPS position. We tested our approach using movement data freely available on Movebank (movebank.org). In a first case study, we extracted Normalized Difference Vegetation Index information from the MOD13Q1 data product for 12,344 GPS animal locations by matching the closest MODIS image in the timeseries to each GPS fix. Data extractions were completed in approximately 3 minutes. In a second case study, we extracted hourly air temperature from the ERA5-Land product for 33,074 GPS fixes from 12 different wildebeests (Connochaetes taurinus) in approximately 34 minutes. We then investigated the relationship between step length (i.e., the net distance between sequential GPS locations) and temperature and found that animals move less as temperature increases. These case studies illustrate the potential to explore novel questions in animal movement research using high temporal resolution remotely-sensed data products. The workflow we present is efficient and customizable, with data extractions occurring over relatively short time periods. While computing times to extract remotely sensed data from GEE will vary depending on internet speed, the approach described has the potential to facilitate access to computationally demanding processes for a greater variety of researchers and may lead to an increased use of remotely sensed data in the field of movement ecology. We present a step-by-step tutorial on how to use the code and adapt it to other data products that are available in GEE.

Reviewer comment:

1. A problem statement is clearly presented in Introduction. There seems to be a lack of balance in providing background information about software. While GEE is briefly summarized (please include the URL of GEE; https://earthengine.google.com/), the R project (for statistical computing) is not described. A brief summary of the R project should be provided along with the URL (https://www.r-project.org/).

Response: We have incorporated the URLs as advised. We have also added a sentence and a reference in an attempt to expand on the R programming language. Google Earth Engine is not a popular platform in the ecology field, so we spent more time in describing it. That is the novelty of our work, making GEE products accessible for analysis of movement data in R.

Line 54: Similarly, new statistical methods to analyze and visualize movement data have become more accessible. For instance, there are at least 58 different packages developed for use in the R programming language (https://www.r-project.org)[5,6], one of the most popular open-source programs for data analysis among ecologists [8]. With such an abundance in data and new methods, the limitation to address relevant scientific questions of interest frequently lies in the required computing power and the technological expertise to do so.

The launch of Google Earth Engine (GEE; https//earthengine.google.com) in 2010 marked an important step towards improving accessibility of remotely sensed data and analysis tools for researchers globally.

Reviewer comment:

1. In Methods and Materials, description of data sources, tools, and methods that were used for the study are not provided. In addition of proper citation (which is already provided), each data source, tool, and method should be briefly described along with providing online sources (e.g., MODIS [including MODIS NDVI], ERA5-Land, JAGS program).

Response: We believe we have provided enough information on the data and methods used with the respective citations. For instance, lines 161-162 describe the MODIS dataset and lines 189-191 explain the ERA5-Land dataset. We provide a detailed step-by-step description on how to use the code workflow with the appendix. We think that expanding on any of the mentioned items will add unnecessary text to the manuscript. We will be happy to expand our descriptions if the editor thinks we should do so.

Reviewer comment:

1. Line 144: I would recommend replacing ‘time difference’ with ‘time lag.’

Response: Done

Reviewer comment:

1. A figure showing the workflow described in the body of the manuscript would be helpful for readers of the paper.

Response: Thank you for this comment, we agree it is a good addition. We have included a new figure as suggested.

Reviewer comment:

1. Lines 148-152: What is the purpose of comparing ‘annual mean’ NDVI and NDVI that matches GPS locations? Should they match? It yes, why? Because I do not quite understand this part of method, I have questions for Figure 1, which include how did you obtain ‘annual mean’ NDVI? Annual mean NDVI should be a single value. How can ‘annual mean’ NDVI have multiple values, which fluctuate, throughout a single year?

Response: Thank you for this comment. We have edited this section to clarify the analysis. In order to compare between extracting an NDVI value for each location based on an NDVI annual mean vs matching the time of the NDVI image, we plotted in the figures both metrics. The NDVI value is always one value per animal location. There are not multiple NDVI values. What we have is multiple locations per animal, each with one NDVI value. The difference is that for each animal location, one value (in red) is the average annual NDVI for that pixel, and the other (in blue) is the NDVI pixel value of the closest image acquired by MODIS sensor in time to when the animal location was recorded. In seasonal environments, there should be more difference between both NDVI metrics, but they will never match.

Line 168: To showcase the potential benefits of pairing tracking data with the time at which images were acquired, we compared NDVI values extracted using the method described above, with NDVI values extracted for all animal locations from an annual mean NDVI image calculated from all the MODIS images available for the year of each tracking dataset. We plotted the NDVI values against time for each animal and for each method (mean annual NDVI and time matched NDVI) for a visual comparison.

Reviewer comment:

1. One suggestion for Figure 1 is that color coding telemetry location by season may tell patterns in NDVI agreement/discrepancy, as stated in lines 199-201.

Response: Thank you for your suggestion. We believe, however, that the temporal variation is well expressed in the NDVI plots. Because seasons are different across the three African ecosystems represented in the data, adding another color palette will make the figure more difficult to understand, without necessarily adding relevant information. We have modified the methods section to better explain how those NDVI values were obtained. We hope that helps clarify the confusion with the NDVI values reported here.

Reviewer comment:

1. Line: 197: What is the purpose of performing “a Gaussian linear mixed-effects model in a Bayesian framework”? Is this performed in R or something else?

Response: We used the Bayesian framework because it is a personal preference for fitting mixed models given its flexibility and philosophy. The analysis was performed in program JAGS through the jagsUI R package in R software, as stated in line 205.

Reviewer comment:

1. Line 187: What does “the Gelman-Rubin diagnostic” do? What did you evaluate using the diagnostic method? Is this performed in R?

Response: The Gelman-Rubic diagnostic is a metric of model convergence that is reported with the model output. We edited the sentence to clarify the use of this metric.

Line 209: We evaluated model convergence by visually inspecting chain outputs and by ensuring that the Gelman-Rubin diagnostic for all regression parameters was <1.01 [37].

Reviewer comment:

1. In Discussion, snow cover and precipitation are stated as examples for additional data layers. It would also be useful to indicate surface water or water bodies as key data for wildlife movement. Although such data may not be readily available, there are published methods for generating those data. If incorporation of custom layers is possible for the code workflow, the workflow would have a great user base.

Response: Thank you for your suggestion. We have included surface water, which can be obtained from products such as the JRC Monthly Water Recurrence product available in GEE (Pekel et al. 2016). Incorporating custom layers is possible and probably relevant when including products derived from multispectral imagery. We added a reference that shows how to create different variables in GEE and that could be integrated into our workflow (Oeser et al. 2020).

Line 262: In this study we present examples incorporating two environmental data layers that are recognized for their importance in explaining animal movement and behavior [5,18,28]. The code can be adapted to incorporate other data layers of interest. For instance, extracting snow cover or precipitation from the same ERA5-Land data product can be accomplished by simply selecting different bands from the image collection. Alternatively, other image collections available via GEE such as surface water availability [39] and other vegetation indices (e.g., VIIRS NASA Vegetation Index Product) can be substituted in the code. The step-by-step tutorial presented in the supplemental materials explains how to use the code and adapt it to other datasets. While clouds limit the possibility to use vegetation indexes with higher temporal resolution as they can result in images with large areas lacking information, novel advances for reconstructing daily MODIS NDVI products are opening new possibilities for incorporating these data sets into movement analysis in the near future [40]. Moreover, the flexibility of GEE makes it possible to incorporate information derived from multispectral imagery such as Landsat or Sentinel as potential covariates in models for animal movement data [41].

Round 2

Reviewer 1 Report

The ms is acceptable in its present form.  I still think that it could be improved to widen the potential audience, but this isn't essential.