Corridor-based approach with spatial cross-validation reveals scale-dependent effects of geographic distance, human footprint and canopy cover on grizzly bear genetic connectivity.

Understanding how human infrastructure and other landscape attributes affect genetic differentiation in animals is an important step for identifying and maintaining dispersal corridors for these species. We built upon recent advances in the field of landscape genetics by using an individual-based and multiscale approach to predict landscape-level genetic connectivity for grizzly bears (Ursus arctos) across ~100,000 km2 in Canada's southern Rocky Mountains. We used a genetic dataset with 1156 unique individuals genotyped at nine microsatellite loci to identify landscape characteristics that influence grizzly bear gene flow at multiple spatial scales and map predicted genetic connectivity through a matrix of rugged terrain, large protected areas, highways and a growing human footprint. Our corridor-based modelling approach used a machine learning algorithm that objectively parameterized landscape resistance, incorporated spatial cross validation and variable selection and explicitly accounted for isolation by distance. This approach avoided overfitting, discarded variables that did not improve model performance across withheld test datasets and spatial predictive capacity compared to random cross-validation. We found that across all spatial scales, geographic distance explained more variation in genetic differentiation in grizzly bears than landscape variables. Human footprint inhibited connectivity across all spatial scales, while open canopies inhibited connectivity at the broadest spatial scale. Our results highlight the negative effect of human footprint on genetic connectivity, provide strong evidence for using spatial cross-validation in landscape genetics analyses and show that multiscale analyses provide additional information on how landscape variables affect genetic differentiation.

Validation of presence-only models for conservation planning and the application to whales in a multiple-use marine park.

Identification of species' Biologically Important Areas (BIAs) is fundamental to conservation planning and species distribution models (SDMs) are a powerful tool commonly used to do this. Presence-only data are increasingly being used to develop SDMs to aid the conservation decision-making process. The application of presence-only SDMs for marine species' is particularly attractive due to often logistical and economic costs of obtaining systematic species' distribution data. However, robust model validation is important for conservation management applications that require accurate and reliable species' occurrence data (e.g., spatially explicit risk assessments). This is commonly done using a random subset of the data and less commonly with fully independent test data. Here, we apply a spatial block cross-validation (CV) approach to validate a MaxEnt presence-only model using independent presence/absence survey data for a highly mobile, marine species (humpback whale, Megaptera novaengliae) in the Great Barrier Reef (GBR). A MaxEnt model was developed using opportunistic whale sightings (2003-2007) and then used to identify areas differing in habitat suitability (low, medium, high) to conduct a systematic, line-transect, aerial survey (2012) and derive a density surface model. A spatial block CV buffering strategy was used to validate the MaxEnt model, using the opportunistic sightings as training data and independent aerial survey sightings data as test data. Moderate performance measures indicate MaxEnt was reliable in identifying the distribution patterns of a mobile whale species on their breeding ground, indicated by areas of high density aligned to areas of high habitat suitability. Furthermore, we demonstrate that MaxEnt models can be useful and cost-effective for designing a sampling scheme to undertake systematic surveys that significantly reduces sampling effort. In this study, higher quality information on whale reproductive class (calf vs. non-calf groups) was obtained that the presence-only data lacked, while sampling only 18% of the GBR World Heritage Area. The validation approach using fully independent data provides greater confidence in the MaxEnt model, which indicates significant overlap with the main breeding ground of humpback whales and the inner shipping route. This is important when evaluating presence-only models within certain conservation management applications, such as spatial risk assessments.

Advancing methodologies for applying machine learning and evaluating spatiotemporal models of fine particulate matter (PM2.5) using satellite data over large regions.

Reconstructing the distribution of fine particulate matter (PM2.5) in space and time, even far from ground monitoring sites, is an important exposure science contribution to epidemiologic analyses of PM2.5 health impacts. Flexible statistical methods for prediction have demonstrated the integration of satellite observations with other predictors, yet these algorithms are susceptible to overfitting the spatiotemporal structure of the training datasets. We present a new approach for predicting PM2.5 using machine-learning methods and evaluating prediction models for the goal of making predictions where they were not previously available. We apply extreme gradient boosting (XGBoost) modeling to predict daily PM2.5 on a 1×1 km2 resolution for a 13 state region in the Northeastern USA for the years 2000-2015 using satellite-derived aerosol optical depth and implement a recursive feature selection to develop a parsimonious model. We demonstrate excellent predictions of withheld observations but also contrast an RMSE of 3.11 µg/m3 in our spatial cross-validation withholding nearby sites versus an overfit RMSE of 2.10 µg/m3 using a more conventional random ten-fold splitting of the dataset. As the field of exposure science moves forward with the use of advanced machine-learning approaches for spatiotemporal modeling of air pollutants, our results show the importance of addressing data leakage in training, overfitting to spatiotemporal structure, and the impact of the predominance of ground monitoring sites in dense urban sub-networks on model evaluation. The strengths of our resultant modeling approach for exposure in epidemiologic studies of PM2.5 include improved efficiency, parsimony, and interpretability with robust validation while still accommodating complex spatiotemporal relationships.

The forestecology R package for fitting and assessing neighborhood models of the effect of interspecific competition on the growth of trees.

Neighborhood competition models are powerful tools to measure the effect of interspecific competition. Statistical methods to ease the application of these models are currently lacking.We present the forestecology package providing methods to (a) specify neighborhood competition models, (b) evaluate the effect of competitor species identity using permutation tests, and (cs) measure model performance using spatial cross-validation. Following Allen and Kim (PLoS One, 15, 2020, e0229930), we implement a Bayesian linear regression neighborhood competition model.We demonstrate the package's functionality using data from the Smithsonian Conservation Biology Institute's large forest dynamics plot, part of the ForestGEO global network of research sites. Given ForestGEO's data collection protocols and data formatting standards, the package was designed with cross-site compatibility in mind. We highlight the importance of spatial cross-validation when interpreting model results.The package features (a) tidyverse-like structure whereby verb-named functions can be modularly "piped" in sequence, (b) functions with standardized inputs/outputs of simple features sf package class, and (c) an S3 object-oriented implementation of the Bayesian linear regression model. These three facts allow for clear articulation of all the steps in the sequence of analysis and easy wrangling and visualization of the geospatial data. Furthermore, while the package only has Bayesian linear regression implemented, the package was designed with extensibility to other methods in mind.

Mapping the geogenic radon potential for Germany by machine learning.

The radioactive gas radon (Rn) is considered as an indoor air pollutant due to its detrimental effects on human health. In fact, exposure to Rn belongs to the most important causes for lung cancer after tobacco smoking. The dominant source of indoor Rn is the ground beneath the house. The geogenic Rn potential (GRP) - a function of soil gas Rn concentration and soil gas permeability - quantifies what "earth delivers in terms of Rn" and represents a hazard indicator for elevated indoor Rn concentration. In this study, we aim at developing an improved spatial continuous GRP map based on 4448 field measurements of GRP distributed across Germany. We fitted three different machine learning algorithms, multivariate adaptive regression splines, random forest and support vector machines utilizing 36 candidate predictors. Predictor selection, hyperparameter tuning and performance assessment were conducted using a spatial cross-validation where the data was iteratively left out by spatial blocks of 40 km*40 km. This procedure counteracts the effect of spatial auto-correlation in predictor and response data and minimizes dependence of training and test data. The spatial cross-validated performance statistics revealed that random forest provided the most accurate predictions. The predictors selected as informative reflect geology, climate (temperature, precipitation and soil moisture), soil hydraulic, soil physical (field capacity, coarse fraction) and soil chemical properties (potassium and nitrogen concentration). Model interpretation techniques such as predictor importance as well as partial and spatial dependence plots confirmed the hypothesized dominant effect of geology on GRP, but also revealed significant contributions of the other predictors. Partial and spatial dependence plots gave further valuable insight into the quantitative predictor-response relationship and its spatial distribution. A comparison with a previous version of the German GRP map using 1359 independent test data indicates a significantly better performance of the random forest based map.