When is variable importance estimation in species distribution modelling affected by spatial correlation?

Asso­ciate Pro­fes­sor from the Uni­ver­si­ty of Twente, Thomas Groen, and for­mer Master’s stu­dent, Nivedi­ta Var­ma Harise­na, received the FAIR Data Fund in Spring 2021 to pre­pare their bio­di­ver­si­ty research data and soft­ware code for pub­li­ca­tion in 4TU.ResearchData. Here, we learn about their research on species dis­tri­b­u­tion mod­el­ling and the impor­tance of shar­ing their edu­ca­tion­al dataset to ben­e­fit the wider eco­log­i­cal research com­mu­ni­ty. 

Species distribution modelling

Each species occu­pies a ‘niche’ which is the match of a species to a set of par­tic­u­lar envi­ron­men­tal con­di­tions. Species dis­tri­b­u­tion mod­els use com­put­er algo­rithms to pre­dict the dis­tri­b­u­tion of a species across geo­graph­ic loca­tions using envi­ron­men­tal data. 

Thomas explains that species dis­tri­b­u­tion mod­el­ling typ­i­cal­ly requires two sets of data. “The first dataset con­tains obser­va­tions about species loca­tion, i.e. where species are present and where they are absent. The sec­ond con­tains infor­ma­tion about envi­ron­men­tal con­di­tions [vari­ables] with­in those loca­tions, such as tem­per­a­ture, rain­fall, veg­e­ta­tion cov­er and incom­ing radi­a­tion, for exam­ple.”

He con­tin­ues, “Using sta­tis­ti­cal mod­els, it’s pos­si­ble to describe the rela­tion­ship between the two datasets, allow­ing researchers to pre­dict under which con­di­tions they are like­ly or less like­ly to find cer­tain species. The idea is to use these mod­els to map species dis­tri­b­u­tion across large geo­graph­ic regions and esti­mate the impor­tance of cer­tain envi­ron­men­tal vari­ables for species sur­vival.”

Species dis­tri­b­u­tion mod­els typ­i­cal­ly reuse exist­ing data. “Species loca­tion data is typ­i­cal­ly col­lect­ed by field work, a process by which sci­en­tists observe and col­lect data with­in ecosys­tems,” says Thomas. “Anoth­er method involves cit­i­zen sci­ence, where­by nature enthu­si­asts record obser­va­tions about species sight­ed in vis­it­ed loca­tions and share the data in pub­lic data­bas­es.”

Researchers who mod­el species dis­tri­b­u­tion can also access pub­lic repos­i­to­ries, such as Bio­clim, the Nor­malised Dif­fer­ence Veg­e­ta­tion Index (NDVI) Cli­mate Data Record (CDR) or the Glob­al Bio­di­ver­si­ty Infor­ma­tion Facil­i­ty (GBIF), to down­load valu­able data on envi­ron­men­tal vari­ables or pres­ence of species. 

Real-world impact

“Species dis­tri­b­u­tion mod­els are par­tic­u­lar­ly valu­able for con­ser­va­tion since they can be used to iden­ti­fy areas of land suit­able for endan­gered species, and these areas can be sub­se­quent­ly pro­tect­ed. Such mod­els can also inform about the effects of cli­mate change or land use change” says Thomas. 

“Agron­o­mists also use these mod­els to iden­ti­fy suit­able areas for crops, or where con­trol of pests is most need­ed. In health sci­ences, the tech­niques are used to mod­el dis­ease spread by vec­tors, such as mos­qui­tos,” he adds. 

The problem of spatial autocorrelation 

Nivedi­ta states that ‘spa­tial auto­cor­re­la­tion’ is a com­mon prob­lem affect­ing the accu­ra­cy of species dis­tri­b­u­tion mod­els. It aris­es when geo­graph­ic loca­tions that are close togeth­er have sim­i­lar val­ues for envi­ron­men­tal vari­ables.

“Researchers often go to the same geo­graph­ic loca­tions to col­lect data. And, sam­pling at the same loca­tions can lead to a bias in the mod­el out­put,” says Nivedi­ta.

She pro­vides an exam­ple. “If the tem­per­a­ture is con­sis­tent­ly high in the loca­tions sam­pled, one might assume that high tem­per­a­ture is impor­tant for species resid­ing with­in those loca­tions. How­ev­er, it’s like­ly that the close sim­i­lar­i­ty in tem­per­a­tures record­ed in those loca­tions inflates the impor­tance of tem­per­a­ture as an explana­to­ry vari­able.”

Nivedita’s educational dataset 

To help researchers under­stand the effects of spa­tial auto­cor­re­la­tion, Nivedi­ta cre­at­ed an edu­ca­tion­al dataset con­tain­ing sim­u­lat­ed land­scapes of envi­ron­men­tal vari­ables and vir­tu­al species that respond to these vari­ables in pre­dictable ways. 

By con­trol­ling the response of the vir­tu­al species to the dif­fer­ent sim­u­lat­ed envi­ron­men­tal vari­ables, the impor­tance of each vari­able in explain­ing species pres­ence is known. By sub­se­quent­ly esti­mat­ing the impor­tance of these vari­ables with estab­lished meth­ods, the (mis) match between “true impor­tance” and “esti­mat­ed impor­tance” at dif­fer­ent lev­els of auto­cor­re­la­tion can be demon­strat­ed 

Nivedi­ta used the FAIR Data Fund from 4TU.ResearchData to refine and pub­lish the data and code under­ly­ing her research paper, ‘When is vari­able impor­tance esti­ma­tion in species dis­tri­b­u­tion mod­el­ling affect­ed by spa­tial cor­re­la­tion?’

Aside from sim­u­la­tion data, her pub­lished dataset com­pris­es a README file and an R script to allow future users of species dis­tri­b­u­tion mod­els to be able to change mod­el para­me­ters, visu­alise the results and reuse the mod­els with­in their own con­text. 

“My dataset allows researchers to explore the sim­u­la­tions and learn how to cor­rect for spa­tial auto­cor­re­la­tion in their own datasets,” says Nivedi­ta. “Researchers can change para­me­ters with­in my mod­el, such as land­scape size and res­o­lu­tion; sam­pling den­si­ty; species response to envi­ron­men­tal vari­ables; and spa­tial auto­cor­re­la­tion lev­els, to repli­cate the con­di­tions of their own dataset.”

“What’s more, they can learn how spa­tial auto­cor­re­la­tion could affect their dataset and can choose to adapt their meth­ods in order to avoid bias, such as increas­ing sam­pling effort, sam­ple size or loca­tion,” she adds. 

The FAIR Data Fund

Work­ing now as PhD researcher at ETH Zurich, Nivedi­ta explained that the FAIR Data Fund helped her set aside a num­ber of hours to ret­ro­spec­tive­ly pre­pare her Master’s dataset for open access pub­li­ca­tion online. 

“I had to for­mat, clean and anno­tate the dataset myself. I knew that if I hand­ed that task to some­one else they would not have been able to under­stand my data. I cre­at­ed under­stand­able doc­u­men­ta­tion, func­tions and vari­able names; removed any errors from the dataset; and, struc­tured the data to make it eas­i­er to find and access by oth­ers,” says Nivedi­ta. 

With sup­port from the Uni­ver­si­ty of Twente’s Open Sci­ence Offi­cer, Markus Konkol, Nivedi­ta also organ­ised for her com­pu­ta­tion­al research to be inde­pen­dent­ly exe­cut­ed and ver­i­fied by CODECHECK. A cer­tifi­cate is avail­able con­firm­ing that Nivedita’s com­pu­ta­tions could be inde­pen­dent­ly exe­cut­ed by the review­ers. 

Nivedi­ta con­cludes with some pos­i­tive com­ments about data pub­li­ca­tion. 

“Pub­lish­ing my data was easy and straight­for­ward, and I received a lot of sup­port from my super­vi­sor, Thomas. Our dataset has been down­loaded more than 100 times since it was pub­lished! It’s excit­ing to observe that data shar­ing and reuse is becom­ing more com­mon in ecol­o­gy. Col­lect­ing sam­ples dur­ing field­work costs a lot of time, mon­ey and effort, so it’s impor­tant that researchers share their work and get prop­er­ly cred­it­ed for their con­tri­bu­tion.”

Authored by Con­nie Clare (4TU.ResearchData)