Supplementary Material 1

advertisement
Supplementary Material 1. Species selection, environmental variables and modeling procedures
Article title: Land use and land cover effects on regional biodiversity distribution in a subtropical dry forest: a
hierarchical integrative multi-taxa study
Authors: Ricardo Torres, N. Ignacio Gasparri2, Pedro G. Blendinger2 and H. Ricardo Grau2
1
Cátedra de Diversidad Animal II, y Museo de Zoología, Facultad de Ciencias Exactas, Físicas y Naturales,
Universidad Nacional de Córdoba (UNC), Av. Vélez Sarsfield 299, 5000 Córdoba, Argentina. E-mail:
rtorres44@gmail.com
2
Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET) - Instituto de Ecología Regional -
Laboratorio de Investigaciones Ecológicas de Las Yungas (IER-LIEY). Facultad de Ciencias Naturales e
Instituto Miguel Lillo, Universidad Nacional de Tucumán (UNT), CC 34, 4107 Yerba Buena, Tucumán,
Argentina
Corresponding author: Ricardo Torres, Cátedra de Diversidad Animal II, and Museo de Zoología, Facultad de
Ciencias Exactas, Físicas y Naturales, Universidad Nacional de Córdoba (UNC), Av. Vélez Sarsfield 299 (5000)
Córdoba, Argentina. E-mail: rtorres44@gmail.com
Species selected
Given that distribution model performances can differ among species according to their different autoecological
characteristics (Hernandez et al. 2006), species selection was aimed to include the widest possible range of
ecological traits, on the basis of four criteria. First, range of distribution, thus selecting a balanced number of
species both widely distributed (generalist species with continent-wide distributions) and regionally restricted
(species whose distributional area is limited mainly to the Chaco, including several endemisms). Second, based
on the literature (e.g. Short 1975; Cei 1980; Redford and Eisenberg 1992; Prado 1993; and references therein),
specialists opinion and our own knowledge, we selected species representative of the major habitats in the study
area (forests, grasslands and wetlands). As forests currently dominate in the study area, species number resulted
unbalanced in favor to forest-dweller species. Third, for birds and mammals we took into account their diets
(including herbivores, granivores, insectivores, carnivores and scavengers) to ensure good representation of
different functional groups. Finally, we considered both terrestrial and arboreal species of the forests-dweller
fauna. Occurrence records for all species were obtained from museum collections and localities cited in the
bibliography (see Supplementary Material 2 for the sources). Additionally, field surveys of trees, birds and
mammals were carried out during winter of 2009 in 40 sites in the NADC, randomly selecting locations within
areas with scarce or absent previous records of the species considered. Given that models in the ACE were fitted
only with biophysical variables, we used all occurrence records from 1950 to the present, according to the
period covered by the bioclimatic layers in the WorldClim database that we used (Hijmans et al. 2005).
However, in the NADC we modeled only with records between 1970 to present, as the main LULC changes in
NADC occurred after that date. Therefore, an initial list of 241 species was filtered down by eliminating species
with less than five records [the minimal number of occurrence localities that can provide reliable models with
MAXENT (Pearson et al. 2007)] after 1970 in the NADC, to end up with 138 species that were modeled for the
ACE (Supplementary Material 3). For each species, we excluded occurrence records from museum databases
without information of exact geographic location. Therefore, we assumed that the associated error to the varying
accuracy of the occurrence localities is less than the resolution used to model in the NADC (0.5 arc-minutes; see
below). The minimum and maximum numbers of occurrence points used for modeling in the ACE for any
species were 11-107 for trees, 12-84 for amphibians, 13-97 for birds and 14-69 for mammals, while for
modeling in the NADC were 5-82 for trees, 5-39 for amphibians, 5-38 for birds and 6-36 for mammals
(Supplementary Material 3).
Species distribution modeling
The modeling of species distribution was performed using MAXENT v3.3, a software package that implements a
maximum entropy algorithm to generate a probability distribution over pixels in a grid of the modeling area
(Phillips et al. 2006; Elith et al. 2011). The maximum entropy algorithm has been shown to be robust for
modeling presence-only occurrence data, even with very low numbers of occurrence records, outperforming
many other techniques (Elith et al. 2006). MAXENT is also able to manage both continuous and categorical
variables (Phillips and Dudík 2008). As recommended by these authors, we adjusted the MAXENT parameters to
default values. Following Pearson et al. (2004) and Anadón et al. (2007), a hierarchical framework was adopted
for model building, conducting first models on the Argentine Chaco Ecoregion (ACE) and then on the Northern
Argentine Dry Chaco (NADC).
Models based on biophysical variables
Models for ACE were fitted using the 19 bioclimatic variables from the WorldClim database (Hijmans et al.
2005), at a resolution of 30 arc-seconds. A layer of distances to water bodies was developed from the Shuttle
Radar Topographic Mission (SRTM) Water Body Data (http://dds.cr.usgs.gov/srtm/ version2_1/SWBD/) for
inclusion in the modeling. An elevation layer from 30 arc-seconds SRTM elevation model
(http://srtm.usgs.gov/_) was also used and a slope layer was additionally derived from it. Soil variables extracted
from the Atlas de Suelos de la República Argentina (INTA, 1995) were added to the models of tree species,
burrowing amphibians, burrowing mammals, and mammals that feeds on ground social ants and termites. All
variables were interpolated to a resolution of two arc-minutes (ca 3.5 km). Layers were prepared and formatted
using IDRISI v15 Andes (Eastman 2006).
Occurrence points were unevenly clustered in space. Because this pattern may influence the prediction of
the model, a ‘bias grid’ was created as the inverse of the Euclidian distance to all points; the MAXENT interface
allows the inclusion of this grid in model fit to control the effect of the spatial bias, giving less weight to pixels
closest to each presence point (Elith et al. 2011).
For each one of the 138 species, we obtained an initial set of ten models in the ACE by randomly selecting
75% of the occurrence localities at each run for training, and leaving the remaining 25% for testing. This initial
set of models was used to identify variables with minimal or no contribution to the models using the MAXENT
jackknife test of variable importance to evaluate the relative strength of each predictor variable (Yost et al.
2008). This is done by assessing the behavior of the ‘gain’, an internal measure of goodness of fit defined as the
average log probability of the presence samples, minus a constant that makes that uniform distributions of the
presence samples have zero gain (Phillips 2010). In this process, MAXENT first calculates the average gain
corresponding to training localities, and later the drop in such average training gain when each variable is
omitted from the full model. Thereby, those variables that did not produce a decrease greater than 0.01 in the
average training gain when they were omitted were removed. Covariation between the remaining variables was
tested by the Spearman rs coefficient, considering only the cells with presence data. Only pairs of variables with
an rs value >0.70 were considered as significantly correlated. The average training gain values of correlated
variables was examined once again; those variables showing the lowest decrease in gain values when omitted
from the full model were also removed.
We performed 100 replicates of the maximum entropy model with the reduced set of variables, again
selecting at random 75% of occurrences for training and 25% for testing at each run. One hundred replicates
were appropriate for this procedure, since with 10,000 pseudo-absence points (as we used in MAXENT),
distributions models can reach stability with ten or even less replicates (Barbet-Massin et al. 2012). The values
of the area under the curve (AUC) of the receiver operating characteristic (ROC) plot (Fielding and Bell 1997;
Manel et al. 2001; McPherson et al. 2004) for test points were examined, and the ten models with the greatest
AUC values were averaged to obtain the final model. Any variable identified as not relevant at this point was
removed, and again a run of 100 replicates was performed following the above mentioned procedure. This
process was repeated until no non-relevant variable remained in the model (i.e. the decrease in training gain
when any of the remaining variables was omitted from the full model was greater than 0.01). To generate a
binary prediction of occurrence, a necessary step is to choose a threshold. For this, we selected the ‘maximum
test sensitivity plus specificity logistic threshold’ provided by MAXENT. That threshold, although somewhat
restrictive, generally offered a spatial representation more according with the distribution of presence points in
the study area (Torres and Jayat 2010).
We decided to exclude from further analysis those species with final ACE models with AUC values below
0.65, or those with final ACE models with less than two variables causing a decrease in training gain greater
than 0.01 when omitted from the full model. Thirty out of the 138 species whose models did not meet these
requirements were not modeled in NADC. Although no published compilation exists on the actual number of
species in the NADC, the final number of bird and mammal species here modeled is a small proportion of the
total species pool in the study area. However, the number of species with different ecological traits and
abundances were balanced in terms of abundance and functional characteristics (see Supplementary Material 3),
except of the bias towards forest-dwelling species mentioned above. Therefore, we considered that the sample of
species modeled is generally representative of the total pool of species in the NADC.
Models based on biophysical and LULC variables
We proceeded to model within the NADC each one of the 108 species (18 trees, 25 amphibians, 48 birds and 17
mammals) that met the criteria for the analysis of changes in richness patterns, with a resolution of 30 arcseconds (approximately 880 m in the NADC). The climatic and topographic variables used here were those
relevant at the scale of the ACE, plus four layers related to land use and land cover: woody biomass, density of
puestos, vegetation classes and distance to crops. a) Woody biomass was used as a proxy of human intervention
over the forest structure. The woody biomass was obtained using field estimates and the Normalized Difference
Vegetation Index (NDVI) from the Moderate Resolution Imaging Spectroradiometer (MODIS) of the satellite
Terra [see Gasparri and Baldi (2013) for a detailed explanation of the methodology]. In brief, to generate a
regional map of woody biomass, diameter at breast height data of 50 field forest samples were associated to
MODIS-Terra spectral data (NDVI) in the dry season by using the Random Forest algorithm (Breiman 2001).
The resulting model showed an acceptable mean predicted versus mean observed deviation (below 3%) and an
average deviation for a single prediction of 15% (Gasparri and Baldi 2013). b) A previous map of puestos made
by interpretation of Landsat images (Grau et al. 2008) was edited and extended to cover the complete NADC
following the same methods and the additional use of the GOOGLE EARTH v5.0 software for interpretation of
satellite images. A layer of puestos density was further developed, where the values in each cell represent the
number of puestos at five cell (4.4 km) radius from the centre of that cell. c) Vegetation classes consist of four
vegetation categories (crops, woodlands, flooded and riparian vegetation, and grasslands/bare soil) plus one of
permanent water bodies (Fig. 1). Crops were mapped by visual interpretation of Landsat images following
standard methods on the Argentine forest monitoring system (UMSEF 2012) previously used in the region
(Grau et al. 2005; Gasparri and Grau 2009; Gasparri et al. 2013). To identify the other four classes, a digital
classification of the green, red, infrared and near infrared bands (i.e. bands 1 to 4) was applied in a
multitemporal set of MODIS images with 250 m of spatial resolution. Three different dates (03/06/2007;
06/12/2007 and 10/16/2007) were used to capture the different stages of the vegetation phenology (Gasparri and
Baldi 2013). The classification was performed with the Random Forest algorithm; an independent evaluation of
the resulting maps showed a precision >85%. Random Forest analyses were performed with the R software (R
Development Core Team 2008) using the randomForest package (Liaw and Wiener 2002). Land cover and
above ground biomass mapping were performed in R software using the YaImpute (Crookston and Finley 2008)
and SP (R Development Core Team 2008) packages. The vegetation layer was smoothed by applying a moving
window so that each cell expresses the dominant vegetation class at 1 km radius from the centre of that cell.
Except for deforestation in favor of agricultural areas, we assumed that main vegetation disturbances (as
overgrazing and charcoal and firewood harvesting) were fixed at the current levels prior to 1970s in the NADC,
with little or no changes in the vegetation classes (as well as spatial patterns in woody biomass or density of
puestos) since that date. d) Distance from crops, which was defined as the average distance from cultivated areas
isolated from the landscape in Landsat images of the years 1972, 1990, 1998, 2002 and 2007. The distance to
crops was included because crops grew since 1970, and being the vegetation layer made from images from a
single recent year, older species occurrence records may be erroneously assigned to crops when in fact they
were recorded in a different vegetation class. In this way, any presence locality recorded in forests (or any other
natural vegetation) but falling today in a cell representing crops were related to an average value of distance to
crops rather than assigned to a categorical classification of ‘natural vegetation’ vs ‘crops’.
Models in NADC followed all the same procedures as in ACE, with the exception of species with very few
(five to ten) presence records. Such species were modeled with MAXENT and the help of the pVALUECOMPUTE
software following the Jackknife validation (n-1) procedure detailed in Pearson et al. (2007).
References
Anadón JD, Giménez A, Martínez M, Palazón JA, Esteve MA (2007) Assessing changes in habitat quality due
to land use changes in the spur-thighed tortoise Testudo graeca using hierarchical predictive habitat models.
Divers Distrib 13:324–331
Barbet-Massin M, Jiguet F, Albert CH, Thuiller W. 2012. Selecting pseudo-absences for species distribution
models: how, where and how many? Methods Ecol Evol 3:327–338.
Breiman L (2001) Random forest. Mach Learn 45:5–32
Cei JM (1980) Amphibians of Argentina. Monit Zool Ital Monograph 2
Crookston NL, Finley AO (2008) YaInpute: an R package for kNN imputation. J Stat Softw 23:1–16
Eastman JR (2006) IDRISI Andes Tutorial. Clark Labs, Clark University, Worcester, MA
Elith J, Graham CH, Anderson RP, Dudik M, Ferrier S, Guisan A. et al (2006) Novel methods improve
prediction of species’ distribution from occurrence data. Ecography 19:129–151
Elith J, Phillips SJ, Hastie T, Dudík M, Chee YE, Yates CJ (2011) A statistical explanation of MaxEnt for
ecologists. Divers Distrib 17:43–57
Fielding AH, Bell JF (1997) A review of methods for the assessment of prediction errors in conservation
presence/absence models. Environ Conserv 24:38–49
Gasparri NI, Baldi G (2013) Regional patterns and controls of biomass in semiarid woodlands: lessons from the
Northern Argentina Dry Chaco. Reg Environ Change DOI: 10.1007/s10113-013-0422-x
Gasparri NI, Grau HR (2009) Deforestation and fragmentation of Chaco dry forest in NW Argentina. Forest
Ecol Manag 258:913–921
Gasparri NI, Grau HR, Gutiérrez Angonese J (2013) Linkages between soybean and neotropical deforestation:
Coupling and transient decoupling dynamics in a multi-decadal analysis. Global Environ Chang 23: 1605–
1614
Grau HR, Gasparri NI, Aide TM (2005) Agriculture expansion and deforestation in seasonally dry forests of
north-west Argentina. Environ Conserv 32:140–148
Grau HR, Gasparri NI, Aide TM (2008) Balancing food production and nature conservation in the Neotropical
dry forests of northern Argentina. Glob Change Biol 14:985–997
Hernandez PA, Graham CH, Master LL, Albert DL (2006) The effect of sample size and species characteristics
on performance of different species distribution modeling methods. Ecography 29:773–785
Hijmans RJ, Cameron SE, Parra JL, Jones PG, Jarvis A (2005) Very high resolution interpolated climate
surfaces for global land areas. Int J Climatol 25:1965–1978
INTA (1995) Atlas de Suelos de la República Argentina. CDROM. Instituto Nacional de Tecnología
Agropecuaria, Buenos Aires, Argentina
Liaw A, Wiener M (2002) Classification and regression by Random Forest. R-news 2:18–22 http://CRAN.Rproject.org/
Manel S, Williams HC, Ormerod SJ (2001) Evaluating presence-absence models in ecology: the need to account
for prevalence. J Appl Ecol 38:921–931
McPherson JM, Jetz W, Rogers DJ (2004) The effects of species’ range sizes on the accuracy of distribution
models: ecological phenomenon or statistical artefact? J Appl Ecol 41:811–823
Pearson RG, Dawson TP, Liu C (2004) Modeling species distribution in Britain: a hierarchical integration of
climate and land-cover data. Ecography 27:285–289
Pearson RG, Raxworthy CJ, Nakamura M, Peterson AT (2007) Predicting species distributions from small
numbers of occurrence records: a test case using cryptic geckos in Madagascar. J Biogeogr 34:102–117
Phillips SJ (2010) A brief tutorial on Maxent. Lessons in Conservation 3:107–135
Phillips SJ, Dudík M (2008) Modeling of species distributions with Maxent: new extensions and a
comprehensive evaluation. Ecography 31:161–175
Phillips SJ, Anderson RP, Schapire RE (2006) Maximum entropy modeling of species geographic distributions.
Ecol Model 190:231–259
Prado D (1993) What is the Gran Chaco vegetation in South America? I. A review. Contribution to the study of
the flora and vegetation of the Chaco. V. Candollea 48:145–172
R Development core team (2008) A language and environment for statistic computing, R foundation for
statistical computing. Vienna, Austria ISBN 3-900051-07-0. http://CRAN.R-project.org/
Redford KH, Eisenberg JF (1992) Mammals of the Neotropics. Vol. 2. The Southern Cone. Chicago University
Press, Chicago
Short L (1975) A zoogeographic analysis of the South American Chaco avifauna. B Am Mus Nat Hist 154:163–
352
Torres R, Jayat JP (2010) Modelos predictivos de distribución para cuatro especies de mamíferos (Cingulata,
Artiodactyla y Rodentia) típicas del Chaco en Argentina. Mastozool Neotrop 17:335–352
Yost AC, Petersen SL, Gregg M, Miller R (2008) Predictive modeling and mapping Sage Grouse (Centrocercus
urophasianus) nesting habitat using Maximum Entropy and a long-term dataset from Southern Oregon.
Ecol Inform 3:375–386
Download