Participatory mapping reveals socioeconomic drivers of forest fires in protected areas of the post- conflict Colombian Amazon

1. Wildfires have increased in protected areas (PAs) of the Colombian Amazon follow ing the 2016 peace agreement between the Government and the Revolutionary Armed Forced of Colombia (FARC— Spanish acronym). Recent study efforts to understand this issue suffer from data scarcity and limited consultation of local stakeholder perspectives on factors affecting wildfires. 2. This study uses a social– ecological systems framework to investigate local perceptions of factors driving and/or preventing wildfires in the


| INTRODUC TI ON
Wildfires increasingly threaten the Amazon region and the ecosystem services that it provides (Barlow et al. 2019). In addition to their exceptionally high biodiversity (Jenkins et al., 2013), the Amazon's tropical forests contribute to global climate regulation (Aragão et al., 2014) and function as a valuable carbon sink that is decreasing due to the cumulative effects of fires, drought, deforestation and degradation (Yang et al., 2018). Wildfires not only affect biodiversity by altering species composition  but also stimulate regional forest dieback . This further weakens the Amazon's capacity as a carbon sink by releasing carbon at a faster rate than deforestation alone, threatening increased climate instability (Aragão et al., 2018;Silva Junior et al., 2019). The resultant warming has been projected to put local livelihoods at risk by negatively affecting ecosystem services and water availability, as well as economic growth in the tropics (IPCC, 2018).
Current global and regional economic models exacerbate forest degradation by encouraging land clearance through burning (Aragão et al., 2008;Barlow et al. 2019;Betts et al., 2008), and ignition rates greatly exceed the 500-1,000-year natural fire return interval determined from charcoal studies in the Brazilian Amazon (Cochrane, 2009). Brazil's recent conflagrations bring into stark relief the need for further research into sustainable alternatives to forest clearance, as well as greater attention to escalating deforestation in countries such as Colombia (Nobre, 2019), which contains 7% of the Amazon basin (Armenteras & Retana, 2012).
Colombian Amazonia has experienced a proliferation of wildfires in the highly biodiverse Tinigua-Picachos-Macarena protected areas (PAs) and surrounding regions, following the 2016 peace agreement between the group formerly known as the Revolutionary Armed Forces of Colombia (FARC-Spanish acronym) and the Government . These anthropogenic wildfires clear the PAs' primary forest and largely correspond to the 'deforestation' fires recently defined for the Amazon (Barlow et al., 2019) although they can also become uncontrolled in the dry season (ANLA, 2017;Caracol, 2019;Minambiente, 2018). As well as being exacerbated by climate change, regional fire occurrence has been attributed to the effects of roads, pasture conversion, fragmentation and deforestation (Aragão et al., 2008;Armenteras et al., 2017;Carmenta et al., 2016;Simmons et al., 2016). However, the acquisition of local data on factors driving wildfire increase has been limited in recent years by the armed groups that continue to occupy parts of the region (Armenteras, Negret, et al., 2019;Morales, 2017;Vélez, 2019).
Social considerations such as the local realities facing stakeholders are central, yet often overlooked aspects of natural resource management that provide crucial insights into conservation challenges and solutions (Biggs et al., 2011;Carmenta et al., 2011Carmenta et al., , 2019. Although departments including Guaviare still grapple with dissident guerrilla fronts, the demobilisation process offers a critical opportunity to widen the scale of investigation and incorporate local perceptions of wildfire dynamics in and around Amazonian PAs as they unfold (The Guardian, 2019).
A social-ecological systems framework can effectively utilise social science methodology to include multiple regional perspectives (Alves, 2008;Berkes et al., 1998;. Many of the Amazonian fire studies that incorporate social field research often focus on individual actors such as Indigenous groups (Mistry et al., 2016) or specific countries such as Brazil (Carmenta et al., 2013Eloy et al., 2019;Viana et al., 2016). However, the urgent need for multi-stakeholder participation in PA policy formulation for the Colombian Amazon is clear from the civil unrest surrounding evictions from PAs such as La Macarena, and the military efforts to halt deforestation through 'Operation Artemisa' (El Tiempo, 2018b;Paz Cardona, 2019). Such exclusionary approaches to PA conservation in the tropics have historically exacerbated social conflict and even biodiversity loss through marginalisation of the rural poor (Brockington & Igoe, 2006;Lele et al., 2010;Martin et al., 2014).
Plural valuation of multiple stakeholder perspectives has been advocated as a more equitable and sustainable approach to natural resource governance (Jacobs et al., 2020). Participatory mental mapping approaches offer intuitive methods for identifying common ground and disagreement between local and marginalised social actors, whose perspectives are key to the successful evaluation of conservation and resource management initiatives (Bennett, 2018;Biggs et al., 2011). Research into targeted solutions to wildfire occurrence in Colombian PAs should therefore contest power imbalances by enabling and strengthening participatory dialogue between diverse regional actors (Jacobs et al., 2020;Schreckenberg et al., 2017).

| Participatory cognitive mapping as a boundary object
of interaction can exacerbate urban-rural power imbalances and the policy-practice gap, undermining local faith in fire management initiatives . Participative mapping can help to address this by coordinating multiple stakeholder perspectives in order to better to understand interactions within a complex system (Reddy et al., 2019).
This study adapts participatory maps for use as 'boundary objects' (Star & Griesemer, 1989) to overcome logistical and disciplinary barriers to dialogue among disparate stakeholders. As a conceptually simple and transferable graphical method (Mendoza & Prabhu, 2003), maps can be combined with focus groups, in which individuals' concepts are abstracted into variables by the collective, allowing for the higher levels of confidentiality recommended for research in post-conflict scenarios (Ford et al., 2017). This is especially important in Colombia, where public defenders of the environment are particularly at risk of criminalisation and acts of violence (Forst, 2018). By integrating disparate forms of knowledge through collective, grounded inquiry (Christen et al., 2015) participative mapping has the potential to democratise environmental decision-making in post-conflict regions of the Global South.
The perspectives of four stakeholder groups were aggregated to address two key questions: (a) What do stakeholders perceive as the most important key variables influencing wildfire in PAs?; (b) How can these variables be effectively addressed by policies based on their interactions within the system?

| ME THODS
This study uses a mixed methods approach that combines participative mapping with semi-structured key informant interviews to provide an explorative, baseline understanding of stakeholder perceptions around the social-ecological system driving wildfire in PAs of the post-conflict Colombian Amazon.

| Study area
The study focused on the Macarena Special Management Area (AMEM-Spanish acronym), which is shared by the Caquetá, The AMEM was designated in 1989 to protect this Andean biodiversity corridor from development pressures including timber harvesting and forest clearance for cattle ranching and illicit coca cultivation (Autoridad Nacional de Licencias Ambientales -ANLA, 2017). Beginning with the 19th century rubber boom, the territorial 'colonisation' process was continued by colonos/campesinos: rural landholders displaced by violence and economic pressure in the 20th century (SINCHI, 1999(SINCHI, , 2000.
After declaring Guaviare and most of Caquetá as Forest Reserves (areas promoting low-impact forest use but without strict protection) in 1959, the State gradually rescinded this status in some areas and incentivised settlement from the 1950s (SINCHI, 2011(SINCHI, , 2016. The majority of land titles in Caquetá and Guaviare belong to smallholder (≤50 ha) campesinos (SINCHI, 2011(SINCHI, , 2016, and land inequality remains high (Gini coefficients of 0.70 and 0.81 respectively, UPRA, 2015). The combined effects of these territorial restructurings and inequalities have contributed to uncertainty and conflict surrounding land use and ownership (Acuña & Rincón, 2007). These range from protests over the forced eviction of ranchers from national PAs by the army (La Nación, 2018a) to the expulsion of forest rangers by FARC dissidents that continue to operate in the area (El Espectador, 2020).
Regional studies indicate that local fires are related to anthropogenic activities such as agricultural clearing for pasture and crops, as well as logging (Armenteras et al., 2013;Dávalos et al., 2016;Tasker & Arima, 2016;Xaud et al., 2013). Open fires are prohibited in Colombia's rural areas, except for strictly controlled burns for soil preparation, harvesting and stubble or frost removal, as well as small open pit mining (Meza et al., 2019). These must take place more than 100 metres from conservation areas, natural forest and pro-

| Participatory mapping
We adapted a participatory mapping method from fuzzy cognitive mapping (FCM), following .
This approach allows experts on a social-ecological system to illustrate complex causal connections between its component variables in a graphical causal model (Figure 2), which can be repeated across different individuals and aggregated by the researcher (Kosko, 1988). After defining a single reference point in time (Özesmi & Özesmi, 2004), in this case, the present day, individuals enumerate and discuss variables they perceive as important or influential within a social-ecological system (e.g. wildfires in PAs). where participants lack the time and experience to effectively engage with finer weighting scales (Nyaki et al., 2014). Participants could add, remove or edit variables and edges throughout the mapping exercise.

| Focus groups and selection of participants
In June 2019, we produced four participatory maps through focus groups representing key regional stakeholders, namely: (1) Local authorities; (2) national park service; and (3) cattle ranchers in Guaviare. (4) Cross-sectional stakeholders included participants representing all of the three previous stakeholder groups, but in Caquetá ( Figure 3). These groups were selected on the basis of their widely perceived roles in controlling (groups 1 and 2) and/or stimulating (group 3) wildfire (Minambiente, 2018a). The cattle rancher groups included individuals who had been working in the region for ≥8 years with property sizes ranging from 10 to 140 ha (mean F I G U R E 1 Time series comparing the number of ~1 km 2 active fire pixels detected by the (MODIS MOD14) Fire Information for Resource Management System in the protected areas of (1) Los Picachos; (2) Tinigua; (3) La Macarena; and (4) outside these PAs in the AMEM's Integrated Management Areas property size = 77.5 ha). The cross-sectional stakeholder group included representatives of Caquetá's local government, national park service and cattle ranchers with property sizes ranging from 90 to 120 ha (mean property size = 105 ha).
We used snowball sampling (Atkinson & Flint, 2001)  Stakeholders were asked to identify and discuss the interactions between economic, social, institutional and environmental variables that influenced wildfires directly or indirectly in initial 1-2 hr focus group sessions (Carley & Palmquist, 1992;Özesmi & Özesmi, 2004).
Following discussion, individuals wrote the factors they perceived as most important on an erasable whiteboard using coloured markers, discussing, removing or combining overlapping variables. The F I G U R E 2 Sample map of a hypothetical unrelated system presented to participants

F I G U R E 3 Flow chart detailing focus group locations and exercises
less accessible cattle ranchers carried out the variable selection exercise orally. Separating the variable selection exercises from the map building helped to keep the focus groups appropriately small (~3-4 individuals) for emotionally charged topics (Morgan, 1996;Smithson, 2007).
The cross-sectional focus group carried out both activities (variable selection and assigning causal weights) with representatives of each sector during one session in Caquetá, in order to reduce regional travel in compliance with a departmental risk assessment. It is important to acknowledge that these larger group dynamics can affect the extent to which individuals feel comfortable expressing their views (Smithson, 2007). However, the group's process of taking it in turns to add or remove variables/ edges and then discussing and explaining these as a whole helped to involve all participants while building towards a group consensus (Morgan, 1996).

The Sustainable Development Corporation for Northern and
Eastern Amazonia, who represented local authorities in Guaviare, were presented with the variables defined by the cross-sectional stakeholder group, as this also included local authorities such as municipal councillors. The final, aggregated map involved 25 individuals.

| Key informant interviews
Maps were compared with semi-structured interviews with three Santos, whose administration initiated the peace process and increased Colombia's PAs by 31 million hectares, was interviewed in the United Kingdom, in English. Each informant was asked to discuss interactions between economic, social, institutional and environmental variables that influence wildfire in PAs directly or indirectly.

| Data analyses
Each map was entered into an adjacency matrix in which the numerically weighted connections (edges) of each variable with the others F I G U R E 4 Map of the study area, with five sites in and around the Macarena Special Management Area (AMEM). Interviews took place in (i) Florencia, Caquetá, (ii) El Doncello, Caquetá, and workshops in (1) San Vicente del Caguán, Caquetá, (2) San José del Guaviare and (3) Calamar, Guaviare were recorded in each row, with 0 indicating no edge Kosko, 1988;Mazlack, 2009). The average values of each edge were then combined into a normalised and aggregated matrix , see aggregated matrix in Appendix S2). Equal weightings were given to all participants, so as not to privilege any single group Özesmi & Özesmi, 2004). Those edges that were repeated were reinforced by their appearance in other maps through the average weightings; those that appeared in fewer maps were included in the aggregated map, but were assigned a weight of 0 for maps in which they did not appear.

| Network structure analysis
Network analyses were applied to map adjacency matrices in RStudio v3.5.2 (Newman, 2010). We used local indices of the edges immediately connected to each node to calculate degree centrality and classify them as (a) transmitters; (b) receivers; and (c) ordinary variables (Table 1; Özesmi & Özesmi, 2004;Solana-Gutiérrez et al., 2017;Văidianu, 2013). Transmitters' high ratio of outgoing to incoming edges suggests higher capacities to act as drivers that influence other variables . Receivers' high ratio of incoming to outgoing edges suggests they are susceptible to the effects of drivers that connect to them. Ordinary variables are those with neither a disproportionately high number of incoming nor outgoing edges.
Degree centrality is limited to measuring the effects of immediate neighbours upon a variable (Sharkey, 2017). However, Katz centrality allows variables to accumulate influence from across the network, with more distantly connected variables contributing less than immediate neighbours (Lavin et al., 2018). We therefore ranked variables according to their Katz centrality to determine their importance in the network, using the default attenuation factor α = 0.1 in Python's Network X library (Hagberg et al., 2008). Maps were visualised using the igraph package in RStudio 3.5.2 (Csardi & Nepusz, 2006).

| Network structure analysis
The aggregated network (see Figure 5 for a simplified diagram and Appendix S3 for the entire network) contained 36 variables affecting wildfire with 112 edges (density = 0.09). Although stakeholders tended to agree on the positivity/negativity of shared edges, the connection between roads and ranching evinced disagreement. The national park service (Appendix S4.3) and cross-sectional stakeholders (Appendix S4.2) attributed a positive relationship (roads drive ranching), and ranchers (Appendix S4.1) attributed a negative one (roads reduce ranching). The resulting 'weak' positive relationship (+0.2) represents the mean of two strong positive connections (+0.8), one strong negative connection (−0.8) and one non-existent connection (0).

| Receiver variables
The aggregated map contained no 'True' receiver variables (zero outdegree centrality); Ordinary receivers are listed in Table 3. These include five of the upper percentile of central variables (Table 4),

Metric Description
Degree centrality Transmitter variables 'True' transmitters had ID value of zero; 'Ordinary' transmitters were within the upper quartile of OD/ID ratio, functioning as system 'drivers'  Receiver variables Variables within the upper quartile of ID/OD ratio, functioning as system 'outcomes'  Central (Katz) variables Variables with highest Katz centrality, which accounts for the cumulative influence of connections with other influential variables (Lavin et al., 2018) Density Measure of connectivity between variables, calculated by dividing the number of edges (E) by the maximum possible number of edges between (N) variables Özesmi & Özesmi, 2004).
suggesting that some of the most important variables may be driven by underlying processes that influence them in the map.

| Central (Katz) variables
As expected, forest fires possessed highest Katz centrality, followed by deforestation and climatic phenomena (Table 4). These were followed by extensive cattle ranching, land grabbing, lack of basic public services, lack of governance, unsustainable ranching methods and untitled land.

| Key informant interviews
All of the nine central variables were also mentioned by every key informant (see Appendix S5 for a table detailing commonality between the maps and key informants). The former President and PNN Director listed cattle ranching as the most important direct forest fire driver, and the SINCHI Researcher selected land grabbing, both central variables. This indicated broad agreement on variables across interviewees and data collection methods.
Variables which were mentioned by ≥2 key informants, but which did not appear in the maps included the forest fire driver of mining, and the following forest fire preventatives, responses and mitigators: military operations, early warning systems, conservation credits and Indigenous practices. Indigenous groups in particular were stressed as 'the best guardians of the environment' 2 who 'take care of the forests, the fauna and all the ecosystem' 3 because 'they know how to harvest', 4 despite not appearing in the locally constructed map. Conservation credits, which were mentioned by two key informants, and oil palm, which was only mentioned by the PNN Director, also appeared in the rancher map. However, these were described as foreseeable (not current) and therefore removed in F I G U R E 5 A simplified visual of the aggregated map of key variables in the wildfire system. Central (Katz) variables and their labels are in bold, transmitter variables and their labels are in turquoise. Edge (connection) widths are determined by their relative strength, and arrowheads by edge direction. Positive edges above the average edge weight are blue, positive edges below this are grey. Negative edges are orange. Weaker edges and variables have been removed for clarity. Variables were selected for visualisation purposes only, and any analyses need to consider the entire network (Appendix S3) order to restrict system perceptions to the present day (Schneider et al., 1998).

| D ISCUSS I ON
Participative mapping and supplementary interviews identified nine central variables that influence forest fires in PAs, according to local stakeholders and key informants. Aside from the three ecological receiver (response) variables of forest fires, deforestation and climatic phenomena, these were (a) extensive cattle ranching; (b) land grabbing; (c) lack of basic public services; (d) lack of governance; (e) unsustainable ranching methods; and (f) untitled land. Each of these will be discussed in turn, in relation to their functioning within the social-ecological system driving wildfire.

| Extensive cattle ranching
As a central receiver variable, extensive cattle ranching was the strongest direct forest fire driver (excluding the ecological receiver variables of climate change and deforestation). It not only drives these variables, but also shares a positive feedback with the latter, while its direct (+0.6) effect on forest fires ranks far above the aggregated map's absolute average edge weight (0.27). This corroborates reports that ranching drives up to 80% of deforestation in Brazil's Northern region (Nepstad et al., 2008). One rancher described how cattle ranching employs burning to clear land for pasture: Here the rancher and the farmer, in order to cultivate effectively, according to them, needs to crop and leave the trees and cuttings to dry, to burn and afterwards to plant. If you did not burn, there will supposedly be no crops.
Ranchers attributed this practice especially to large-scale landowners (>500 ha), 28 of whom own approximately 91% of Guaviare's unrestricted territory (outside forest reserves), with the remaining 9% held in smaller parcels across 5,456 owners (SINCHI, 2016 because there is no marketing and distribution, and people invested in agriculture and they lost their money. The pastural expansion which accompanies frontier development, often replacing coca, is an important deforestation driver in the Colombian Amazon, which must be stabilised through improved infrastructure and basic services, as well as conservation governance (Dávalos et al., 2016). With coca plantations covering 41,382 ha in Putumayo-Caquetá, and 10,500 ha in Meta-Guaviare (UNODC, 2018), unsupported land use transitions could add another 50,000 ha of pasture, at least 8,000 ha of which is located in PAs. 5

| Land grabbing
As a direct driver of deforestation (0.2) and forest fires ( (Lizcano, 2018), the lack of focus on 'mafias' in the maps could illustrate variation between the ways in which national policymakers and local stakeholders can perceive the same issues (Kingdon, 1984). Critical evaluation of other land grabbing forms mentioned in the local focus groups, for instance on behalf of multinational companies, is therefore also important to avoid policy misfits that fail to address reported stressors (Bunce et al., 2010).
For example, the rancher group discussed and eventually removed the oil industry variable from their map, explaining: Thank God, they have not arrived here yet. We hope they never will.
The sentiment is clear, as is that of the Siona group protesting oil exploration on their reserve in neighbouring Putumayo. With every key informant listing Indigenous practices, reserves and their traditional ecological knowledge as key to forest fire mitigation, violation of Indigenous land rights must be treated as seriously as other forms of land grabbing. Any other approach risks extending Colombian land reform's unresolved bias towards local elites to multinational corporations (Bucheli, 2006;Grajales, 2011

| Lack of investment in basic public services
Lack of investment in basic public services appeared as a central ordinary variable, reflecting similar priorities identified by smallholders in Brazil (Carmenta et al., 2013). It directly drove lack of governance, illegal armed groups and extensive cattle ranching, while contributing to poor road conditions and unsustainable ranching methods.
Ranchers described how the lack of investment in road maintenance drove farmers to choose livestock over arable agriculture, which requires better transportation infrastructure: If you need to sell an animal, you can take it by foot; you don't need the road, […] but if we are going to plant, for example, cassava and plantain here, we end up getting stuck with them.
Smallholders have explained the proliferation of cattle ranching in the Brazilian Amazon with similar reasoning (Pereira et al., 2016). In Guaviare, ranchers particularly stressed the damage suffered by roads as climate change intensifies flooding, indicating that the map's feedback between climatic phenomena and road deterioration could stimulate further extensive ranching, if road maintenance is not improved.
However, all such conclusions should take into account the uncertainty implicit in the map's contested opinions over the connection between roads and ranching.
Although the projected expansion of Caquetá's road network (Foro Regional, 2018-2022) could facilitate alternative livelihoods reliant on connectivity, new roads expose forests to destructive exploitation (Laurance et al., 2009), and should be supplementary to rural reform and distributional improvements. Otherwise, the aggregated map indicates that road development should be exercised with extreme caution, prioritising maintenance over expansion to avoid 'uncontrolled colonisation' (Clerici et al., 2019).

| Lack of governance
The central ordinary variable of lack of governance was driven by weak institutions and the lack of basic public services. The former President also made this connection when describing 'the lack of judicial action against forest fires, (…) which has to do with lack of territorial control and lack of adequate institutions'. By contrast, a rancher described how guerrilla groups limited forest burning and clearance before the peace agreement: While there is no doubt these human rights violations outweigh any potential conservation gains, the power vacuum left by the FARC has allowed habitat destruction to continue largely unchecked (Clerici et al., 2019). Some ex-combatants have even returned to arms from dissatisfaction with the promised amnesty (La Nación, 2018b), and such key conditions of the peace agreement must clearly be fulfilled to prevent dissention and further social and environmental harm (El País, 2019;UN News, 2019). Conflict reignition could exacerbate the lack of governance, as well as illicit crops, another wildfire driver in the map (Davalos, 2003).

| Unsustainable ranching methods
This ordinary central variable referred to the lack of technical education in more sustainable livestock management practices, such as that offered by an agroecological training partnership between the Centro de Pensamiento Anticorrupción, ETH Zurich and two Colombian universities. While these initiatives were praised by ranchers who had access to them, their reliance on international partners combines with the prohibitive expense of techniques such as pasture rotation and trail verge management 6 to limit their current scale.
Although the National Development Plan promises renewed investment in rural sustainability (Mintransporte, 2019), local authorities' use of the conditional tense (e.g. 'more social investments This institutional encouragement towards land clearance for unsustainable ranching is indicative of the rigidity traps that arise from prioritising homogenous production systems over the maintenance of heterogeneous landscapes (Carpenter & Brock, 2008). The cattle ranchers even went so far as to designate their profession 'a necessary evil'. Improved credit access could incentivise landowners to preserve, rather than clear forests, and finance sustainable ranching techniques currently described as prohibitively expensive.

| Untitled land
Extensive literature explores the ongoing land rights disputes influencing Colombia's conflicts, including the unequal distribution of property, the difficulties of formalising land ownership in rural communities and communities' violent displacement from their territories (Botero, 2016;Latorre, 2015;Valencia Agudelo, 2017). Such insecure land rights negatively affect smallholder productivity and resource conservation (Godfray, 2013).
Stakeholders saw abandoned land resulting from previous armed conflicts as directly perpetuating land grabbing, lack of investment in basic public services, forest fires and colonisation culture (Colombian Constitutional Court, 1995;USAID, 2013

| Strengths and limitations
Participative mapping offers many advantages for social-ecological research, including the ability to integrate knowledge from varied stakeholders

| CON CLUS ION
Participative mapping of the system driving wildfire occurrence in three PAs of Colombian Amazonia reveals the perceived influence of extensive and unsustainable cattle ranching, land grabbing, untitled land and the lack of governance and basic public services. Both local stakeholders and key informants agreed on the importance of these variables, as well as the interlinked ecological variables of deforestation and increased climatic phenomena.
Although the peace agreements' aims to strengthen territorial control and promote rural development intersect with the stakeholders' aggregated perspectives, the map revealed the contested nature of variables such as roads and their changing relationship to forest fire occurrence when implemented in conjunction with other forms of development. With cattle ranching strongly affected by access to basic public services and rural connectivity, this study reveals the importance of holistic development that addresses systemically important variables such as territorial control, rural economic alternatives and the titling and distribution of land (Armenteras, Negret, et al., 2019). Systemic change may involve readdressing the legacies of agrarian colonisation that continue to drive Amazonian forest fires.

ACK N OWLED G EM ENTS
The research was supported by the University of Oxford's School of Geography and the Environment and Oriel College. We are grateful to Dr Ariell Ahearn for guidance on focus group and interview methodologies; Dr Meredith Root-Bernstein for insightful comments on an earlier version of the manuscript; Joan Sebastian Barreto Rivera for assistance with mapping the study area; and Magdalena Drożdż for assistance with Python. Our thanks extend to all of the participants who gave up their time to share their valuable perspectives.

CO N FLI C T O F I NTE R E S T
The authors have no conflict of interest to declare.

DATA AVA I L A B I L I T Y S TAT E M E N T
Interview transcripts are not available due to data privacy agree-