Incorporating management action suitability in conservation plans

The conservation of biodiversity is fraught with complexity and uncertainty (Hemming et al., 2022). To deliver on objectives, decision makers must navigate challenging social technical contexts while contending escalating threats (Maxwell 2016). Quantitative models are frequently used to help guide the making process, providing insight into complex scenarios by way attempting remove unstructured subjective judgement (García-Díaz 2019). As scope objectives continues increase, expert opinion alone will be insufficient select between range actions proposed meet so quantitative increasingly relied upon provide support (Farley 2018). A foundational component determining action(s) needed objectives. often in this context predict management outcomes that alternative courses action can compared (Addison 2013). These assume equally likely succeed once implemented or quantify probabilistic specific occurring (Canessa 2015; Regan 2005). However, modern occurs a social-ecological landscape (Yletyinen 2021) which affects whether more less suitable for implementation given context. probability is, therefore, not same as success. Actions appear during planning phases objective may prove inappropriate implementation, ultimately resulting delayed compromised 2020). Thus, informing decisions would benefit immensely from information successfully landscape. Machine-learning suite statistical techniques identify structure data generate predictive (Olden 2008). particularly useful identifying patterns drivers without assuming priori structural relationships. has been widely ecology, including species distribution modelling (Elith & Leathwick, 2009), image recognition (Fairbrass 2019) other instances where performance paramount (Van Doren Horton, This utility appears well suited disentangling relationships associated implementing order determine their suitability Here, we use novel application machine-learning programme across large spatial extent. We frame our study using New Zealand's Predator Free (PF 2050) initiative, seeks eradicate biologically economically harmful invasive vertebrate mammals Zealand 2050 (i.e. rats Rattus rattus, R. norvegicus, exulans; stoats Mustela erminea, M. furo, nivalis; common brushtail possum Trichosurus vulpecula; Russell, Innes, 2015). assist achieving goal national coverage actions, (a) factors played past role different at sites (b) could suitably future sites. then demonstrate applicability approach scenario analysis (c) explore how these vary throughout country individually maximising cost-savings, humaneness probability. Our methodological provides an objective, transparent reproducible framework competing extents under uncertain constraints. Invasive enacted tools suppress population area set areas (hereafter ‘management tools’ ‘tools’). PF government-sponsored initiative well-supported both socially financially since its adoption 2016. Indeed, government championed activities 5,000 community groups iwi M?ori tribes) annual contribution approximately $75 million NZD (totalling over $350 2016–2020) (Department Conservation, 2020, 2021). Currently, patchwork top-down nationally agencies bottom-up individual groups. resulted incomplete coverage, some locations have others do not. Since eradication cannot achieved currently available (Murphy 2019; Peltzer 2019), focuses complete ongoing suppression existing tools, anticipating availability transformative technologies future. conducted main islands ‘the mainland’; Figure 1a), region spanning North Island Te Ika-a-M?ui (106,097 km2), South Waipounamu (165,693 Great Barrier Aotea (242 km2) Stewart Rakiura (2,055 km2). 1 × km grid (267,704 cells) define scale analysis. resolution because it sufficient capture known investigate selected predictor variables. Although many offshore/outlying invaded terrestrial mammals, excluded they managed independently mainland and, most, achievable (Carter Additionally, uninhabited lakes high-elevation (?1,500 m, represents approximate average target persist alpine environments; O'Donnell 2017). All processing R environment (version 4.0.2; Core Team, 2017) ArcGIS 10.7.1 (ESRI, 2011). Four primarily manage mammalian predators – aerially broadcast poisons, ground-based kill-traps exclusion fences (Figure 1b–e). Briefly, pesticides (as palatable cereal baits) sown helicopters underslung bucket spreaders, ground hand, bait-bags bait-stations; attract individuals striking mechanism kill them; designed prevent passage animals location. sometimes perceived unpalatable uncompassionate, lethal control crucial preventing extinctions island biota around world most part accepted (Russell, Jones, investigated correlating applied past, collated records S1) 26 years (1995–2021) provided Trust (PFNZ Trust; Bell pers. comm. 2020), TrapNZ (P. Handford D. Bar-Even comms. 2020) Manaaki Whenua Landcare Research (Landcare Research; J. Innes N. Fitzgerald Department Conservation (DOC; P. McMurtrie Eschenmoser Data PFNZ DOC delineated boundaries (e.g. enclosed trapping treated poison operation), detailed location geographic coordinates each trap received own unique coordinate location). Exclusion ?0.25 km2 were identified centroids (Innes 2019); <0.25 Burns al. (2012) personal contacts experts. Boundaries obtained manually digitising aerial images eco-sanctuary managers owners. overlaid tool cells previously implemented. generated map (hereafter, ‘bias map’; Elith 2008) reflect bias presence–absence account sampling presence-only records, model fitting S2). maps ensured informed but solely driven historical practices (see Appendix S1 details pre-processing occurrence generation). successful depends biogeographical 2021; Courchamp 2003; Holmes Oppel Biophysical factors, such geographical isolation geomorphic features, logistically constrain For example, isolated far people preclude regular servicing requirements poisons; Brown located near water bodies better leverage natural dispersal barriers reduce costs fences; 2012). Social human density number landowners), influence viability certain based real impacts (Niemiec 2016; multiple datasets source variables models. Land Institute Information Zealand, 2018), [e.g. Landcover Database (LCDB); Research, 2020], Use Carbon Analysis System (LUCAS; Newsome 2017; Kemp Gorman, comms.), Court (Maori Court, International Union Nature (IUCN; BirdLife International, 23 biogeographic land-tenure (Table S1; S3) Table S2 details). interpretation, classified predictors geomorphic, related land cover related. Specifically, (geomorphic:) mean slope, elevation, distance water, cost distance; (landscape:) core index (mean CAI), circumscribing circle (RCC), shape index; (land cover:) cropland, grassland, shrub/scrub-land forest. related, tenure non-target species. included urban, building density; (tenure:) public land, private (iwi) (non-target species:) sensitive endemic ‘sensitive endemics’). accounted autocorrelation CAI, RCC, tenure-related endemics calculating focal statistics 20 diameter S4). choice was validated measuring effective Block CV package (v2.2.1; Valavi Finally, categorical differences two additional variable indicating cell Island. combine relatively simple single more-complex increased (Polikar, There ways bagging, boosting stacking; Breiman, 1996; Schapire, 1990); however, demonstrated strong similar studies, robust against multicollinearity, reliably important (Friedman, 2001). specifically variant gradient algorithm due recent advances scalability (Chen Guestrin, 2016), decreases time relative algorithms (c.f. adaptive boosting, ‘adaboost’; Freund 1997). fitted boosted trees separately (via XGBoost r package; v1.2.0.1; Chen tool, denoted presences had pseudo-absences randomly recorded use. biases reported implemented, pseudo-absence sampled weights tool. trained described 24 logit link function accommodate presence/pseudo-absence data. increase performance, also calibration involved combinations tuning parameters evaluated them fivefold cross validation parameters). After models, curve (AUC) statistic best mainland. continuous grid, S5). thresholds sum sensitivity specificity create binary indicate S6, wherein 0 = suitable, suitable; Liu AUC scores 0.87 ± 0.01 SD, 0.88 0.00 0.89 SD 0.93 0.05 fences, respectively aims study, confident adequate all >0.75 AUC; Pearce Ferrier, 2000). ability cluster residuals variables' ‘gain’) visualised centred conditional expectation (c-ICE) plots (Goldstein predictions mapping environmental similarity surface 2010). (reviewed Peterson 2003) enacting considered three influencing selection (Brown Each allocated criteria. In instances, four being unsuitable (identified via maps; S6), case fifth ‘future’ contiguous network. lack means required achieve environments yet undergo (but current tools). term conditions. aimed maximise created allocating (based maps, see S5) possible creating network cost, secondary. government, supporting committed reflects desire successful. 1). conducting requires acceptable (Oppel 2011), preference humane tools. single-most among those cell. quantified Sharp Saunders (2011) two-part assessment framework, integrates welfare intensity/duration suffering, alphanumerical score S3). converted tool's ordered factor (1–5, least humane; S7) scenarios. cost-savings minimising expenditure maintaining Given perennially limiting (Wilson Law, 2016) major long-term undertaking, strategies conduct sustainable, prolonged management. cheapest entirety 29 years: 2021–2050). Cost values plausible studies (from primary literature documents) summed per-unit operational entire operation inclusive labour planning, costs). estimate considering longevity poisons every 3 maintain low densities) fair comparisons made. aim estimates programme, ratio conversion enabled us compare magnitude pricing nationally. might starting position estimates, no considers benefits programme. drew 2). words, there effectively predicted locality. Moreover, analysed generally than Landscape metrics, CAI gain 0.130 0.03 0.120 0.02 respectively, scale), expansiveness), contiguity Geomorphic describing walk km), elevation lacustrine waterbody) 0.109 0.116 0.095 0.0 distance, Notably, [iwi] landowners) influential any model, marked 0.007 levels). Many align S8). gaps filled-in, regions Gisborne Hawke's Bay Otago Southland features). Similarly, naturally defendable fiords/sounds Marlborough Sounds), peninsulas Banks Peninsula, Farewell Spit), mountainous along coastlines, S9). note limited overall, novelty technology few established date. insights developing contiguous, national-scale 3). Under scenario, deployed (comprising 57.8% land-area; employed land-area 82.6% c.f. widest general 48.3% 2), lower impact animal parts S6). When comparing humanness (cf. 3b 3a), replaced lesser extent, scenario. balanced instance, 22.2% 25.7% mainland, 41.2% (similar scenario; Allocating criteria overall financial managing 10% 20% expensive, difference finding highlights importance cost-effectively tracts (Elliott Kemp, Overall, 89% eligible (238,518 require Areas requiring tended geographically clustered. contained largest 21,802 74.7% area), majority Fiordland West Coast mammal initiative. By predictions, analyses tool-use configurations what look like prior eradication. suggest others, translate depending first strategic deployment inform complex, landscape-scale envision research informative policy-makers who incorporate alongside considerations address international challenges global change, biosecurity (Faulkner climate change (Kappes well-being (de Wit broad-scale when results highlight interplay consider should (contiguousness expansiveness features within cell; 2a). exhibited highly non-linear relationship S10), suggesting effect contiguousness context-dependent trends scalable engaging practitioners local experts designing plans increasing (Crowley expertise essential integrating larger cohesive units programmes continue expand, regional jurisdictions, (Peltzer little result contrasts previous causal mechanisms underlying inhabited Niemiec One explanation discrepancy metrics too coarse fully (Beever proxies habitation smaller, sparsely populated Carter 2021), densely areas. Another derived biophysical RCC already implicitly aspects land-tenure. measures (in minutes/km) traverse nearest vehicle-accessible trail centre significantly negatively correlated scale, caused underestimate residual clustering occurred land; Disentangling socio-ecological enabling investigation readily scales. reveal 2050. First, swaths full potential. up 5.8- 1.8-times levels, respectively. Second, small scenarios, seem further 3; safeguarding vulnerable predation (Burns 2012) serving vital nationwide high broad usage Third, 10.9% land-area, especially higher elevations, necessitating tool(s), Clusters inaccessible terrain S3c), Tongariro 3), outcome coincide difficulty rather appropriateness per se. Indeed developed overcome accessibility issues disentangle truly if only remote priority. formulating alternatives planners (and hence contributes paradigms, dynamic policy pathways; Haasnoot Alternate pathways cornerstone structured cognitive biases, differing value judgements, and/or interests present overcoming querying commonly held fallacious beliefs remains option status quo bias; Gregory provisioning pragmatic solutions; ideal practical, identifies satisfactory solutions highest Future extend efficacy meeting density-impact functions; Green Grosholz, subsequent steps process S4 limitations). Zachary T. James C. Russell conceived study; Jeffrey O. Hanson collected data; led writing manuscript; authors contributed substantially revising text gave final approval publication. grateful project. thank Rebecca Peter Daniel GroundTruth; John Neil Pete McMurtrie, Patrik Eschenmoser, Josh Nicholas Gorman Conservation; Bruce University Auckland; numerous owners giving permissions include eco-sanctuaries study. Z.T.C. supported Doctoral scholarship, Auckland Faculty Science 2021 PhD Output Award, Ltd; J.O.H. Environmental Climate Change Canada (ECCC) Conservancy (NCC); J.C.R. Royal Society Rutherford Discovery Fellowship (Grant No. RDF-UOA1404), Biological Heritage National Challenge. Thanks reviewers editor feedback earlier drafts manuscript Wareko Angina abstract translation. Open access publishing facilitated Auckland, Wiley - agreement Council Australian Librarians. funded advisor Zero Predators, Ltd. code Zenodo Digital Repository https://doi.org/10.5281/zenodo.6339723 S3 Please note: publisher responsible content functionality supplied authors. Any queries (other missing content) directed corresponding author article.