Multi-level multi-scale resource selection models using machine learning were compared and contrasted for generating predictive maps of jaguar habitat ( Panthera onca) in the Brazilian Pantanal. Multiple spatial scales and temporal movement levels were run within several analytical modeling frameworks for comparison. Included in the analysis were multi-scale raster grains (30 m, 90 m, 180 m, 360 m, 720 m, 1440 m) and GPS collaring temporal movement levels (point, path, and step). Various analytical methods were used for comparison of models that could accommodate data structural levels (group, individual, case-control). Models compared included conditional logistic regression, generalized additive modeling (GAM), and classification regression trees, such as random forests (RF) and gradient boosted regression tree (GBM). The goals of the study were to discuss the potential and limitations for machine learning methods using GPS collaring data to produce predictive habitat suitability mapping using the various scales and levels available. Results indicated that choosing the appropriate temporal level and raster scale improved model outputs. Overall, larger level analytical modeling frameworks and those that used multi-scale raster grains showed the best model evaluation with the inherent condition that they predict a broader scale and subset of data. The identification of the appropriate spatial scale, temporal scale and statistical model need careful consideration in predictive mapping efforts.
Landscape patterns and processes occur within many spatial and temporal dimensions, and scale is a lens through which to view those dimensions. Landscape ecology examines ecological processes and landscape scales through modeling approaches. Ecological models depend on the entities measured (e.g. which organisms, ecosystems or ecological processes), variables measured (e.g. which environmental or climate covariates), and the processes linking entities and variables. Additionally, they can also respond to changes in spatial extent, spatial grain, temporal duration, and temporal grain used to measure these entities and variables [
Resource selection may also depend on the how the process of interest is represented. For GPS data, resource selection functions may be based on different sampling units used to represent animal movement: points (locations in space), steps (displacements), and paths (sequences of displacements). Binary response variables include (1) animal movement data and (0) background points generated on the landscape within several levels by using a “used vs. available” sampling design broadly referred to as resource selection function (RSF) [
Parametric, semi-parametric, and non-parametric models are important options for analysis on movement ecology. Parametric models generally aim at estimating cause-effect relationships between an organism’s movement features and the environmental covariates. For example, conditional logistic regression is often used in point and step selection approaches to find out which sets of environmental conditions are selected or avoided by organisms. However, these parametric approaches are not recommended for predictive mapping or predicting results on new data. Predictive mapping efforts that utilize non-parametric approaches such as machine learning, for example, are most suitable for predicting on new data [
Emerging studies on machine learning (ML) methods for landscape ecology have used classification and regression trees (CART), showing they outperform other methods for multi-scale modeling [
Previous research has also applied ML methods like RF for movement studies for point selection functions for mule deer [
How predictive modeling, like machine learning methods, may apply to step and path selection functions is unknown. This study extends previous research on movement ecology and machine learning by applying non-parametric and semi-parametric techniques to multi-scale, multi-level methods for GPS collaring data using also path and step selection functions. Step and path selection functions traditionally use conditional logistic regression, and have recently begun exploring alternative methods such as machine learning. We believe that multi-scale predictive modeling and mapping may improve through the use of machine learning and generalized additive modeling approaches. ML methodologies are becoming more useful to ecologists collating big data of high dimensions from data repositories. Both remote sensing data (e.g. Google Earth Engine) and animal movement data (e.g. MOVEBANK) repositories are making data widely available, which also demands for more sophisticated processing and modeling approaches [
Here we aimed at understanding how parametric, semi-parametric, and non-parametric models can contribute to habitat selection estimates and predictive mapping (
The study area is the Brazilian Pantanal region surrounding the Taiama Ecological
Station (17.712061S, 57.415956W.) The station is deep in the Brazilian Pantanal, the world’s largest freshwater wetlands located in the state of Mato Grasso do Sul and Mato Grosso, in Western Brazil. Several major tributaries of the Paraguay River incur a seasonal flooding regime from January to July [
Remote sensing information was particularly useful in this study, and the availability of data layers globally has made compiling such large datasets much easier. A total of 12 raster-based environmental data layers were extracted from Google earth engine and open source data layers available online, namely elevation, slope, aspect, land use and land cover, forest non-forest, total canopy cover, roads, water and hydrology, human density, and cattle density (
| Layer Name | Source |
|---|---|
| Elevation | SRTM Digital Elevation Model |
| Slope | SRTM Digital Elevation Model |
| Aspect | SRTM Digital Elevation Model |
| Global Cover | ENVISAT’s Medium Resolution Imaging Spectrometer (MERIS) Level 1B Land Cover |
| Forest Non-Forest | JAXA L-band PALSAR SAR and ALOS mosaics using backscatter coefficient to determine “Forest” and “Non-Forest” |
| Total Canopy Cover | Landsat Vegetation Continuous Fields (VCF) total cover of vegetation of woody plants above 5 m in height. |
| SAR Land Cover | HH and HV L-band ALOS/PALSAR, and HH and HV C-band. RADARSAT-2 using a hierarchical object-based image analysis approach to hydrologically variant subregions. |
| Water | The Global Inland Water dataset shows inland surface water bodies, including fresh and saline lakes, rivers, and reservoirs. |
| Roads | Center for International Earth Science Information Network - CIESIN - Columbia University - Distance Raster generated from Shapefile with Global Roads Open Access Data Set, Version 1 (gROADSv1) 1980-2010. From this distance to roads were calculated. |
| Hydrology | WWF HydroSHEDS Raster - Hydrological data and maps based on Shuttle Elevation Derivatives at multiple scales. |
| Human Density | WorldPop Raster with estimated the number of people per hectare by 2015, with national totals adjusted to match the estimates of the UN population division. |
| Cattle Density | FAO, ILRI, the University of Oxford and the Université Libre de Bruxelles. Raster with Global Distribution of Livestock |
were extracted for time periods suitable for the study, unless unavailable, then previous layers were used, for example the roads layer is current to 2010. Collinearity was checked using pairwise comparison for those models where this may be an issue, any variables that had greater than 0.7 pairwise correlation would be rejected. None of the variables met this criterion, and were uncorrelated. Machine learning models do not consider collinearity to be an issue so all data layers can be included.
The study attempts to understand how various modeling approaches perform, considering spatial information at multiple scales and model levels (Supplementary Information). Point, step (1 hour displacements), and path (24 hour trajectories) approaches were used to represent movement (
| Model | Level |
|---|---|
| Point Selection | Home Range |
| Step Selection | 1 hour steps |
| Path Selection | 24 hour paths |
| Model | Resource Selection Order |
|---|---|
| Gradient Boosting Method | Group |
| Random Forest | Individual ID |
| GAM | Case-Control/Individual ID |
| Conditional Logistic | Case-Control/Individual ID |
| Model | Raster Scales |
|---|---|
| Gradient Boosting Method | All Scales 84 variables |
| Random Forest | All Scales 84 variables |
| GAM | Univariate Scales 12 variables |
| Conditional Logistic | Univariate Scales 12 variables |
To represent the raster layers at multiple scales, a Gaussian kernel smoother was used to average the layers at multiple extents (90 m, 180 m, 360 m, 720 m, and 1440 m), resulting in a total of 84 variables. If the original raster layers were not in a 30 m grain, they were disaggregated to 30 m for analysis. Specifically, the MERIS global land cover (300 m) grain and the SAR land cover (50 m) were disaggregated to 30 m. Assessing the functional grain of analysis, or the grain at which the organism is responding to the landscape, for connectivity studies has been useful to create resistance maps of habitat preferences [
The conditional regression model and GAM models were fit univariately to determine the adequate scale [
Modeling methods rely on niche habitat concepts for environmental covariates to construct accurate models of jaguar distributions on the landscape [
GPS collars (Lotek Globalstar and Iridium Collars) for jaguars (n = 11, five females and six males) are the largest group of animals monitored in the Northern Pantanal. Data were made freely available by Morato et al. (2018), and capture procedures and permits were described in Morato et al. 2016. Monitoring occurred from October 2013 to February 2016 for 909 total days of data collection with individuals ranging from a minimum of 26 days and a maximum of 597 days. Collars were programmed to collect one relocation every hour, summing up 42,741 observed locations for all animals. Data collected followed protocols approved by Instituto Chico Mendes de Conservação da Biodiversidade (Ministério do Meio Ambiente, Brazil (ICMBio-SISBIO)). All procedures followed guidelines approved by the American Society of Mammologists [
Minimum convex polygons have been the traditional method for home range estimation, drawing a polygon around the point locations for the animal, to be used as the area available to animals in their routine movements. This technique provided a crude estimate of the home range, most commonly reported with 95 percent of data points [
The GPS collars captured one point every hour. These hourly point data were subset into “steps”, using the first point as the start of the step, and the next point as the end of the step. Background steps had the same original step distance projected into a different angle around the starting point. Steps from hourly data had a total of 85,388 presences and absences, generating one background step for each step. Step selection was performed in program R, using the package AdeHabitatLT [
Animal trajectories were subset into 24-hour time sequences of hourly steps, generating longer paths for daily intervals, with a total of 4409 present and absent paths. Absences were generated using a correlated random walk (CRW) method. The CRW began at the starting point, simulating a trajectory of a similar length at alternate angles. CRWs are a completely randomized simulation of blind jaguar movement at any chance direction. CRW does not account for any decisions the jaguars normally encounter in time or space.
Nonetheless, the correlated random walk provided a randomized path to understand the simplest baseline from which to compare the actual jaguar movement. The mean of all steps values in the path aggregated to one single value to represent each path extracted covariate. Paths were generated in program R, package SiMRiv [
Since the use-available (presence-absence) data for steps and paths were generated from the original data points, then these steps and paths can be paired in a case-control framework for the analysis. The conditional logistic and GAM models fit in a case-control framework allowing for a direct comparison between these two models.
However, the RF model used a higher order level at the individual ID strata and was not case-control. Therefore, RF performs the analysis for all steps and all background steps of individual animals during the duration of the study, and not matching presence steps with the generated absence steps them directly. The GBM algorithm is an even larger level using the entire group’s paths and steps together without any subsetting into individual strata or case-control (
Traditional resource selection functions for GPS collaring studies use an explanatory modeling approach such as conditional logistic for case-control for the steps and paths. Hooten et al. (2014) developed a point process model. In a very basic interpretation of the model, the probability density function for use [ x ] u is equal to a weighted distribution of availability [ x ] a , then further indexing resource observations by relocation at time t (Equation (1) and Equation (2)) [
x [ ( s t ) ] u = g ( x ( s t ) , β ) [ x ( s t ) ] a ∫ g ( x ( s ) , β ) [ x ( s ) ] a d s = [ x ( s t ) | β ] u (1)
The likelihood is maximized for resource coefficients β.
∏ t = 1 T [ x ( s t ) | β ] u (2)
A vector of resource covariates, β is a set of regression coefficients, x, a normalizing constant, and g ( x , β ) is a resource selection function [
The generalized additive models are a semi-parametric extension of the generalized linear models that allow for non-linear functions of the environmental covariates. This method assumes that functions are additive and components smoothed. It estimates an additive approximation to the multivariate regression function, employing univariate smoothers and using individual estimates to
| Data Structure | |||||
|---|---|---|---|---|---|
| Group Level - GBM | Individual ID Level - Random Forest | Paired Case Control Level - Conditional Logistic Regression - Generalized Additive Model | |||
| Use | Use | Individual ID | Use | Individual ID | Case Control |
| 1 (Present) | 1 (Present) | 1 | 1 (Present) | 1 | 1 |
| 0 (Absent) | 0 (Absent) | 1 | 0 (Absent) | 1 | 1 |
| 1 (Present) | 1 (Present) | 3 | 1 (Present) | 3 | 2 |
| 0 (Absent) | 0 (Absent) | 3 | 0 (Absent) | 3 | 2 |
explain relationships between variables.
y i = β 0 + f 1 ( x 1 i ) + f 2 ( x 2 i ) + ⋅ ⋅ ⋅ + f q ( x q i ) + ϵ i (3)
The smoothing splines, thin-plate splines, cubic splines, and splines with variable knots (k = 3, k = 8) applied to applicable covariates (Equation (3)). In this way, the environmental covariates are split into knots, and data in each knot section are fit independently, furthermore adding functions of knots to predict the link function. GAMs are frequently used in species distribution models [
Decision trees, like Random Forest (RF), are used to create partitions or splits between the predictors, forming them into regression trees. These models are ensemble models that allow for multiple models to be fit, combining the results with the rationale that this will produce a better result than a single model. It does this through the model fitting with training data and then using testing data to estimate error and the importance of each variable. RF is one approach to classify data into decision trees by generating B different bootstrapped training sets in a technique called “bagging” (Equation (4)). Bagging is a non-parametric modeling technique that is useful for high-variance predictors. Bagging averages the observations, and can significantly lower the variance compared to traditional classification trees [
f ^ b a g ( x ) = 1 B ∑ b = 1 B f ^ * b ( x ) (4)
Models generated by bagging alone have issues with becoming correlated due to selecting the strong predictor in the topmost split for all of the trees generated. Random forest is an extension of bagging, such that trees generated by bagging by using a subset of randomly chosen p predictors ( p ), to decorrelate the trees from having dominant predictors in any of the models [
Further, model-averaging with classification trees that have low pairwise correlations, due to this variable separation among trees, reduces model bias and improves model accuracy [
Gradient Boosted Method (GBM) works similarly to RF except that it does not use a bootstrapped dataset. In GBM, trees are built using the residuals from previously grown decision trees to improve the function. Boosting works as an optimization algorithm, gradient descent method. Boosting minimizes the loss function at each step, reducing the residuals through shrinkage methods whereby irrelevant predictors are made to have minimal effect on predictions [
f ^ ( x ) = ∑ b = 1 B λ f ^ b ( x ) (5)
The shrinkage parameter λ works out inconsistencies in the residuals even further by forming new arrangements of trees (5) [
In many wildlife studies, AIC (Akaike’s Information Criterion) is normally used to select models and select for the best set of predictors. AIC does not evaluate the ability of models’ predictive functions. Learning methods that subset the data into training and testing observations are validating the efficacy of the model. Model evaluation using k-fold cross-validation subset data into training data subsets, or folds (k), and then fit models using K − 1 folds for the model training. Five folds were used and divided evenly the presence and background data. The step selection data was a subset in 5 folds (n = 8539). The path selection data was a subset in 5-folds (n = 881). Model evaluation metrics were used to evaluate model accuracy. This study used Area Under the Curve (AUC), Cohen’s kappa (Kappa), and the True Skill Statistic (TSS). AUC is a graphic method for specificity and sensitivity, with AUC values greater than 0.5 known to perform better than completely random noise. The Kappa statistic is based on thresholds derived from a confusion matrix, looking for the maximum Kappa value between 0 and 1 to determine model efficacy. The True Skill Statistic uses sensitivity and specificity for a confusion matrix and ranges from −1 to +1, and any values 0 or less indicate random models [
In this study, we assessed predictive modeling approaches for a multi-level multi-scale GPS-collaring study. Our approach allowed us to determine which of the levels and scales might perform adequately for predictive mapping of the landscape. In this case, the multi-scale path selection function GBM model performed the best (AUC = 1, cor = 0.989, TSS = 1, Kappa = 1). There were comparatively good single-scale path selection function RF model results. Additionally, the single and multi-scale point selection function RF also performed similarly well. Predictive maps were generated for the machine learning outputs (RF and GMB) and habitat suitability results were scaled at equal intervals for comparison of the landscape predictions (
It became imperative to identify the best fitting model for the smallest scale and level that could generate predictive maps. In this case, the smallest level of
data, single-scale and multi-scale SSFs, did not perform adequately on any models. The PSF case-control models were the next smallest temporal level. However, the model did not evaluate well, nor fit the data. The multi-scale PSF, with an individual strata level, did fit the data well and can be considered the best fitting smallest scale model. Through examining the results, it became clear that the largest level of models fit the data better, although were also the most broadly applied and less specific to the data themselves, so we thought to search for the most specific model to the data as possible.
For the largest level model, the “group” level, the gradient boosted tree (GBM), gave the highest model accuracy for the PSF. The level of this analysis is comparatively larger than other models. The GBM model only slightly outperformed the RF algorithm; however, the RF performs at a slightly more specific level of the individual id strata. We assume that larger level RSFs will likely evaluate as being better models because of the broader level of analysis. Models that are more specific, i.e. specifying strata (individual id, case-control), are more specific to pairing the data generated for each path or step not traversed. Additionally, when we consider the point selection function as having no case-controls or anything to pair due to the inherent randomness of background point generation within the MCP, pairing data points is non-consequential and thus at a higher level than the steps or paths. Usually, point selection outperforms step and path selection; although step and path are more specific to an animals’ hourly or daily movements. Therefore, taking methods at different temporal and levels of analysis can mean that the researcher should attempt to classify the best fitting and smallest levels of analysis as possible during model evaluation to generate the most accurate predictions.
For the explanatory models, the conditional logistic regression only performed on the 30 m2 raster data point selection function (AUC = 0.821, cor = 0.54, TSS = 0.50, and Kappa = 0.50), and GAM performed on the multi-scale point selection function (AUC = 0.90, cor = 0.654, TSS = 0.657, Kappa = 0.656). These results can be interpreted as better than random. All other models were unable to fit GAM or conditional logistic regression. In this case, the case-control framework did not improve the analysis, as point selection functions generate background points randomly without any temporal or spatial “pairing” in the data, rendering the case-control functionality completely random although the level of this analysis could be considered “individual strata” within the home range MCP that the background points were generated. GAM operated at the level of individual id, and was able to outperform the conditional logistic regression for single and multi-grain point selection.
Model ResultsSingle Grain (30 m2) Point Selection FunctionUsing the smallest available grain for all layers, the point selection function showed the RF model performed very well, followed by GBM (
| Model | AUC | cor | TSS | Kappa |
|---|---|---|---|---|
| RF | 1 | 0.989337 | 1 | 1 |
| GMB | 0.988558 | 0.903022 | 0.896347 | 0.895994 |
| GAM | 0.879735 | 0.625121 | 0.605795 | 0.605548 |
| Conditional Logistic | 0.821929 | 0.540328 | 0.507053 | 0.50537 |
| Model | AUC | cor | TSS | Kappa |
|---|---|---|---|---|
| RF | 0.739253 | 0.46462 | 0.393011 | 0.393657 |
| GBM | 0.703558 | 0.372691 | 0.305952 | 0.306359 |
| GAM | 0.558275 | 0.102469 | 0.085603 | 0.085621 |
| Conditional Logistic | 0.549032 | 0.081891 | 0.076443 | 0.076415 |
| Model | AUC | cor | TSS | Kappa |
|---|---|---|---|---|
| RF | 1 | 0.979821 | 1 | 1 |
| GBM | 0.978913 | 0.8614446 | 0.8516936 | 0.8516711 |
| GAM | 0.746084 | 0.416801 | 0.383126 | 0.383275 |
| Conditional Logistic | 0.630573 | 0.216279 | 0.209406 | 0.209485 |
| Model | AUC | cor | TSS | Kappa |
|---|---|---|---|---|
| RF | 1 | 0.989594 | 1 | 1 |
| GBM | 0.993498 | 0.927041 | 0.921899 | 0.9216746 |
| GAM | 0.902305 | 0.654069 | 0.657657 | 0.6568519 |
| Conditional Logistic | 0.746150 | 0.248888 | 0.000709 | 0.0006918 |
| Model | AUC | cor | TSS | Kappa |
|---|---|---|---|---|
| RF | 0.739331 | 0.464646 | 0.394652 | 0.395227 |
| GBM | 0.712909 | 0.391378 | 0.319602 | 0.320075 |
| GAM | 0.559197 | 0.104748 | 0.081383 | 0.081383 |
| Conditional Logistic | 0.5028 | 0.000508 | 0 | 0 |
| Model | AUC | cor | TSS | Kappa |
|---|---|---|---|---|
| GBM | 1 | 0.997403 | 1 | 1 |
| RF | 1 | 0.98323 | 1 | 1 |
| Conditional Logistic | 0.630938 | 0.130467 | 0.187526 | 0.187623 |
| GAM | 0.842983 | −0.59011 | −0.1051 | −0.1048 |
successful model out of all of the various levels (step or path) at this 30 m2 raster scale. This scale and level also had the best performing conditional logistic regression models.
Single Grain (30 m2) Step Selection FunctionThe best model for the 30 m2 grain step selection function was RF, followed by GBM (
The results from the single grain path selection function showed better model evaluation metrics for all models than step selection (
The multi-scale data improved all model results, improving all of the model estimates when compared with the 30 m2 grain (
The multi-scaled step selection function produced similar results to those found in the 30 m2, with negligible increases in the AUC, cor, Kappa, and TSS (
The multi-grain path selection GBM was the best fit model and was fit with 84 multi-scale environmental predictor variables, making only slight improvements over the 30 m2 and multi-scale point selection RF models (
This research explored methods for modeling the predictive habitat suitability and jaguar resource selection in the area surrounding the Taiama NR in the Brazilian Pantanal. The applicability of parametric models like conditional logistic regression, and non-parametric models that use machine learning (ML) methods, were applied GPS collaring data for various movement sampling units, study designs, scales, and grains available for habitat selection mapping. This study was able to apply multi-level multi-scale modeling similar to other studies [
This study also assessed the applicability of ML methods operating at RSF orders such as individual strata or group level, not strictly within a case-control framework like conditional logistic regression or GAM. Similar to other studies that have sought to compare ML methods to conditional logistic regression [
Previous studies have demonstrated the importance of choice of raster grain for producing resistance surfaces that are then used with movement simulations such as least cost paths [
These results are generally consistent with other findings that random forest and other machine learning algorithms perform “better” than logistic or conditional logistic regression. In this case, the conditional logistic regression fit only slightly better than random except for the point selection function at a single 30 m2 grain. This study demonstrates a direct comparison between a semi-parametric GAM and conditional logistic regression using case-control data. The GAM improved model estimates compared to conditional logistic regression for all level models, and on the right dataset for these techniques a large improvement in model fit could be shown to improve model evaluation. Advancements in machine learning may include developing specific tools to accommodate case-controls, where specific random forest algorithms developed with conditional logistic regression could be developed specifically for this purpose. Models would have to all be fit using a case-control framework to be directly comparable at one level. Otherwise, levels for individual strata, and group level can be seen as having this overarching discrepancy.
This multi-level approach allowed us to further understand the effects of level on predictive modeling approaches. From these results, it may be assumed that larger temporal and model levels fit these particular data better, and that looking at smaller levels like case-control for the most specific step level was not able to perform adequate enough to use the results. It was our objective to identify the levels and scale to suit the analysis and the results indicated that larger scale models fit the data best, although are not as specific to movements of the organism.
Due to the computational package functionality of each statistical model to operate at the levels for the entire group, or strata for individual id, or work within a case-control framework, this difference in model level potential within the analysis became a subject of concern. For example, GBM operates at the group level, and random forest at the individual id strata, GAM and conditional logistic at the individual id strata and case-control. Furthermore, this study explores the inherent limitations of current machine learning packages for working within one level of analysis, suggesting a supervised learning approach that can either be specific to one level of RSF for direct comparison, or comparison at higher order levels if necessary.
The machine learning derived predictive maps and models used in this study are not meant for direct interpretation of variable importance such as giving direct estimates of preferred canopy cover or distance to water. In predictive modeling, the goal is to accurately predict and project something new and optimize accuracy of making predictions, in contrast to understanding why these models predicted in the way they do [
This study analyzed non-parametric, semi-parametric, and parametric methods for multiple temporal levels (point, step, and path selection), model levels (group, individual, case-control), and raster grains to compare the applicability of predictive statistical methods in comparison to explanatory methods, such as conditional logistic regression for predicting large areas of the landscape. We compared the parametric statistical approach using conditional logistic regression, with non-parametric and semi-parametric models that accounted for non-linearities such as generalized additive models and classification trees (RF and GBM), comparing the results using model selection methods (AUC, cor, KAPPA, TSS) derived from k-fold cross validation. The results revealed differences in predicting landscape resource selection using non-linear modeling approaches. This case study illustrates inability for a direct comparison of machine learning and explanatory modeling approaches for step and path selection functions due to the inherent data levels (case-control, individual stratum, group) that the models will accept.
E.B. conceived of the idea. E.B. developed the analytical methods and performed the computations, wrote the manuscript. D.H. and T.H. supervised the final draft and finding of the work. D.H. suggested major revisions to conceptual structures and language. B.A. discussed the results and contributed comments to the manuscript.
Special thanks to the field team at Centro Nacional de Pesquisa e Conservação de Mamíferos Carnívoros, Instituto Chico Mendes de Conservação da Biodiversidade, Atibaia, SP, Brazil who were involved in the fieldwork and project discussion for this project. Espcially Dr. Morato and Dr. Santos who made extensive comments throughout manuscript preparation. Special thanks to the GIS preparation team that helped assemble remote sensing data layers for the analysis. Special thanks to Kathy Zeller for making initial comments.
The authors declare no conflicts of interest regarding the publication of this paper.
Bohnett, E., Hulse, D., Ahmad, B. and Hoctor, T. (2020) Multi-Level, Multi-Scale Modeling and Predictive Mapping for Jaguars in the Brazilian Pantanal. Open Journal of Ecology, 10, 243-263. https://doi.org/10.4236/oje.2020.105016