Uncertainty in Species Distribution Mapping

UNCERTAINTY INPREDICTING SPECIES DISTRIBUTIONA case study using four methods to map tigeroccurrence in Central SumatraSunarto & Marcella J. Kelly, Virginia Tech

Pre-assessment How many of you have modeled species distribution? For those who have done it, how many have: accounted uncertainty/evaluated model accuracy?

Outline Background Why species distribution? How species distribution is mapped Uncertainty and other issues Case study on predicting tiger distribution

Why map/model sp distribution? Understanding the ecology Fulfilling the need for (species) management

www.tamrin.proboards.com/Why: Examples of ApplicationUnderstanding ecology:Dispersal & barrierResources requirementsInteractions: wildlife & human

Brooks et al. 2006Why: Examples of ApplicationBiodiversity conservation priorities6

Conservation vs.DevelopmentHigh ConservationValue Areas (HCVA)7

MCP

Kernel Density

Home-range buffer 1

Home-range buffer 2

Grid

Predicting species distribution: typical steps1. Determining scale2. Selecting variables3. Developing experimental design4. Collecting data5. Developing/choosing statistical procedures:range/envelope, distance/similarity, regressions,…6. Building & selecting models7. Translating mathematical model into distribution map

After Morrison et al. 2006Uncertainty & other issuesImprecise data/measurement error: positional& attribute, applicable for bothspecies presence data (response variable) or environmental/habitat (predictorvariables)Error/uncertainty in data transfer/treatments/manipulationUncertainty in selections of predictor variablesUncertainty in species detectionError/uncertainty in model parameter estimationUncertainty in modeling approachUncertain inferencesFallacious/unrealistic assumptionsNatural variability/stochasticity of the system

Other issues : scaleAdapted from: Karanth & Nichols 2002

A case study: predicting tiger distribution• Critically endangered• Elusive: high association with uncertainty• Urgent need for distribution maps• Year of the tiger

Study Area

Study Area17 km17 km

Types of dataSpecies occurrence (response variables):• Presence only• Presence-’absence’• Detection-Nondetection• CountPredictors: Macro-habitat (global/landscape level features)

Modeling approaches used:Presence only data: Maximum Entropy (MaxEnt)Presence-’absence’ data: Logistic regression (R)Count data: Zero-inflated Negative Binomial regression (R)Detection-non detection data: Occupancy (Program PRESENCE)

Relative importance of variablesMaxEntLogistic Regr.CountOccupancy(% Contr.)(Logit link β)(Log link β)(Logit link β)Intercept NA -11.74 -2.38 (1.11) -8.48 (4.10)Forest area in 2007 59.8 1.30 0.73 (0.22) 0.61 (0.50)Altitude 0.7 113.09 20.94 (11.79) 94.08 (45.55)Distance to core forest area 0.1 -0.77 NA -0.85 (0.66)Road density 14.8 0.92 NA 0.05 (0.69)Distance to road 12.5 NA NA NADeforestation from 2006 to 2007 1.7 NA NA NADistance to core of protected areas 8.8 -1.07 NA NAPrecipitation 1.7 NA NA NA

Summary of predictionsMaxent Logistic Count OccupancyMinimum 0* 0 0 0Maximum 0.9 1.0 128.3 1.0Mean 0.237 0.329 2.036 0.423CV 0.871 1.185 4.499 0.830*) For Maxent outputs, no cell can be interpreted as complete absent even if prediction is (near) 0(Phillips et al. 2006)

MaxEnt

Logisticregression

Count

Occupancy

“All models are wrong, butsome are useful”(George Box, quoted in Kennedy 1992: 73).

Concordance (%) between modelsOverallAccuracyREFERENCESMaxent Logistic Count OccupancyK-hatOverallAccuracyK-hatOverallAccuracyK-hatOverallAccuracyK-hatMaxent 28 13 38 22 38 23Logistic 28 13 47 34 52 40Count 38 22 47 34 56 45Occupancy 38 23 52 40 56 45

‘Correct predictions’ by different models( based on presence-only tiger records in 20 grids collected from independent surveys)Percentage of correct prediction100%75%50%25%MaxEntOccupancyCountLogistic regression0%1 2 3 4Treshholds on probability of occurrence34 1 = Very low, 2= low, 3= medium, 4= high

ConclusionsBetter to model than simply to map occurrence: accounting samplingefforts & environmental variablesIgnoring detection probability ~ underestimating the populationparametersVariation in results from different modelsModel robustness for some variables & areasAccount for missing detections whenever possible: use occupancyFor some cases, presence-only model is still an acceptable choice.

AcknowledgementFinancial & programmatic supportsprovided by: WWF Indonesia & Networks,Virginia Tech, Hurvis Family, STF, PHKA, OFWIMField team: Zulfahmi, Harry Kurniawan,Karmila Parakkasi, Eka Septayuda, Kusdianto,Fendy Panjaitan, Agung Suprianto, E. Tugiyo, L.Subali, H. Gebog, Herri Irawan, Roni Faslah,Kokok Yulianto, Sunandar, Riza Sukriana,TarmisonSpecial thanks to: Drs. Sybille Klenzendorf,Steve Prisley, Jim Nichols, Jim Hines, Dean F.Stauffer, Mike R. Vaughan, WHAPA colleagues