The 179,300 km² Pantanal biome, a geological depression of the Upper Paraguay River basin (Fig. 3A), is the largest inland tropical wetland on Earth. The Pantanal is located at the epicenter of South America, with most of its extent within Brazil and smaller portions located in Bolivia and Paraguay13. The Pantanal vegetation is a macromosaic influenced by the central Brazilian wooded scrubland (Cerrado) savannah, the Amazon, and Chaco biomes (Fig. 3B, 3C). This wetland is shaped by the hydrological dynamics exerted by annual and supra-annual flood pulses16. For the purposes of this study, we considered only the Brazilian Pantanal, because it encompasses 78% of the entire ecoregion and our socioeconomic and biophysical databases did not include Bolivia and Paraguay.
Jaguar distribution model (JDM)
This study used the jaguar presence database published in Brazil’s Jaguar Conservation Action Plan11 and 22 additional localities that we obtained from 2010-2017. We spatially rarified the presence database (SDM Toolbox41) at a distance of 6 km42 to preclude spatial autocorrelation, resulting in 147 unique presence points. We included the following biophysical layers as predictors in our models: bioclimate (gridded climate data, http://worldclim.org/version243); topography (SRTM Digital Elevation Data; https://www2.jpl.nasa.gov/srtm/); and land cover (Global Land Cover Map; http://due.esrin.esa.int/page_globcover.php). We avoided multicollinearity by selecting only uncorrelated or weakly correlated variables (< 0.70) for modelling: elevation; land cover; annual mean temperature (bio1); mean diurnal temperature range (bio2); maximum temperature of the warmest month (bio5); mean temperature of the wettest quarter (bio8); precipitation of the driest month (bio14); and precipitation of the warmest quarter (bio18). All variables were resampled at a spatial resolution of 1 km2.
The JDM was developed using Maxent (v.3.4.144, 45, 46), the most widely used SDM algorithm. Maxent estimates a target probability distribution by finding the probability distribution of maximum entropy, subject to a set of constraints that represent incomplete information about target distributions44. We set the default parameters in Maxent (convergence threshold of 1.0 x 10-5 = 0.000010, with 500 interactions and 10,000 background points, auto features), plus a variable importance analysis based on the jackknife, response curves, and random seed. The JDM was generated via bootstrapping methods with 10 random partitions and replacement, with 70% of the dataset used for training and 30% for testing the models. The result was a probabilistic model with pixel values ranging from 0.0 to 1.0. Higher suitability values represent higher probabilities of finding the species in the field. The output threshold of the JDM (i.e. habitat suitability) was interpreted to represent the probability of encountering one or more jaguars at a given site.
Bovine cattle density (CD) was estimated using data available13, which considered both exotic pasture and natural grassland vegetation maps, the total pasture area per property, and the occupation rate of heads of cattle per hectare (i.e. stocking density) at each of the 22 municipal counties across the Pantanal. For our purposes, 10,000 points were randomly plotted throughout the Brazilian Pantanal while retaining a minimum distance of 3,266 m between neighbouring points. This distance was based on the average size of rural loandholdings in the Pantanal13. We then intersected the points with the CD map to extract density values for each of the 3,631 rural properties. We derived a kernel map using a 25 km-radius and 900-m pixels, weighted in relation to the CD. Finally, we normalized the entire raster data within a range between 0.0 (minimum) and 1.0 (maximum) for each pixel.
We identified and mapped all non-urban lodges and hotels throughout the Pantanal using the compulsory federal registry of the Brazilian Ministry of Tourism (https://cadastur.turismo.gov.br/hotsite/). We conducted internet searches to identify lodges that were not yet registered in the federal system and confirm that all selected lodges indeed operated as ecotourism enterprises. Following identification, information on geographic location and private landholding boundaries were collected directly from the lodges via site visits and telephone calls. A kernel map was then generated based on these coordinates, with a radius of 25 km and pixels of 900 m. After this step, we normalized the raster data within a range between 0.0 (minimum) and 1.0 (maximum) for each pixel.
Wildfire data were derived from a model that identified and assigned dates to all burnt areas. These burnt areas were identified from Chrono sequences of daily multispectral images retrieved from satellite imagery without the preprocessing need of cloud masking and image selection. The model used input data from the 750 m bands of VIIRS that was resampled to a 0.01° spatial resolution grid. The derived burned areas were validated against higher resolution reference maps and compared to the global burned area datasets MCD64A1 Collection 6 and FireCCI5147. These spatial data were made available by the Environmental Satellite Applications Laboratory of the Federal University of Rio de Janeiro48. For our study, we used shapefiles depicting all annual burnt areas of the entire Brazilian Pantanal for four consecutive years (2017, 2018, 2019 and 2020).
After developing each layer (i.e. jaguar habitat suitability, cattle stocking density; density of ecotourism lodges; and 2017-2020 burnt areas), we divided the Brazilian Pantanal into 4,951 hexagonal cells, each of which corresponding to the average size of an operational cattle ranch in the Pantanal13. For each hex-cell, we extracted the average pixel value for each raster (jaguar habitat suitability; cattle density; and density of ecotourism lodges). For the wildfire data, we extracted the proportional area that was burnt in each hex-cell in each of the four years (2017-2020). To facilitate the analysis, all data were log-transformed. A Pearson correlation matrix was then performed to examine the relationships between the seven spatial layers at the scale of hex-cells. Permutation tests were used to further explore pairwise spatial differences at the hex-cell scale (N = 4,951) between JDM, cattle suitability, and ecotourism potential.
Our null model used randomly shuffled observed values without replacement while keeping sample sizes constant for 5,000 iterations. We compiled difference values for each iteration to create 5,000 distributions of potential differences. To obtain a probability value, these differences were compared to their respective probability distributions from the permutations. Spatial layers were processed using QGIS (v. 3.16.5; QGIS Development Team 2021) and ArcGIS (v. 10.03; ESRI 2011) software. Data were analyzed using the “tidyverse” workflow49 and the “infer” 50 package in R (R Core Team 2021). To formally assess spatial autocorrelation, we used the “SpatialPack” R package51, which quantifies the spatial association between any two deﬁned processes on a ﬁnite subset of a spatial plane.