Biodiversity losses have increased in tropical forests due to fire-related disturbances. As landscape fragmentation and climate change increase, fires will become more frequent and widespread across tropical rain forests worldwide, with important implications for forest dynamics by altering plant-animal interactions. Here we tested the hypothesis that recurrent fires in tropical rain forests change bottom-up and top-down forces controlling the abundance of insect herbivores, which in turn increases herbivory. To quantify herbivory, we collected 50 leaves per tree of five species in burned and unburned experimental plots (N = 75) in southeastern Amazonian forests. We measured leaf nitrogen content and leaf thickness of tree leaves as bottom-up factors that could explain differences in herbivory; we measured predation pressure on model caterpillars and estimated the abundance of predatory ants as top-down factors. We found higher herbivory in burned than in unburned forests, as well as lower predator attacks in caterpillar models and lower abundance of predatory ants. Leaf nitrogen content did not vary across treatments. Birds attacked model caterpillars more frequently in burned than in unburned forests, and leaf thickness was higher in burned forests, but these factors together were not enough to offset the higher herbivory in burned plots. Fire degrades tropical forests not only by killing trees and altering their structure and community dynamics, but also by reducing predatory arthropods and disrupting predator-prey interactions, which triggers increased herbivory. These indirect impacts of recurrent fires probably contribute to further alter forest structure, functioning, and to decrease forest regeneration in Amazonian forests.

Information from our manuscript:

Study site

The study was established at “Fazenda Tanguro”, in the state of Mato Grosso, Brazil, in the southern portion of the Amazon basin (13º 04’ S, 52º 23’ W). The climate is tropical humid, and the transitional forest between the Cerrado (Brazilian savannas) and the southeast Amazon forest composes the vegetation of our experimental plots. Expanding agriculture threatens the region (Brando et al. 2013; Marques et al. 2020), ultimately altering regional climate and increasing fire occurrence (Brando et al. 2013, 2020).

Experimental fires were set in three large and adjacent 50-ha (500 × 1000 m) forest plots edging a soybean field. One plot was burned annually from 2004 to 2010, except in 2008 (Burn 1yr); a second one was burned triennially, in 2004, 2007 and 2010 (Burn 3yr); and a third plot was not burned (Control). Burnings were conducted at the end of the dry season between August and September, with kerosene drip torches. Fires were more intense and severe in the triennially burned plot because of higher litter accumulation (Balch et al. 2015), especially after the extreme drought events of 2007 and 2010 (Brando et al. 2014). Even eight years after fires ceased, these burned forests continued to degrade, with an increased mortality of large trees, and decreased above-ground biomass and canopy cover (Brando et al. 2019b). Our plots were not replicated due to logistical, ethical and legal limitations. However, the plots did not differ regarding several vegetation and microclimatic variables before fire (Balch et al., 2008), suggesting that observed differences can be attributed to the treatment rather than to site effects. For more details of the fire experiments, including pictures of burned and control sites, see Balch et al. (2015).

Sampling design

We selected five individuals from five tree species in each plot treatment (N = 75). These individuals were located across two linear transects, distant 500 and 750 m from forest plot edges, and were spaced at least 10 m from each other. The selection of tree species was based upon the following criteria: absences of extrafloral nectaries and commonness across the study area, according to previous inventories (e.g., Brando et al. 2019b). The selected species were Myrcia multiflora (Lam.) DC. (Myrtaceae), Micropholis egensis (A.DC.) Pierre (Sapotaceae), Sacoglottis guianensis Benth. (Humiriaceae), Sloanea sinemariensis Aubl. (Elaeocarpaceae) and Tapirira guianensis Aubl. (Anacardiaceae).

Herbivory assessment

We quantified leaf area loss by chewing insects (hereafter herbivory) between March and April 2018 by sampling 50 leaves from each tree – and a total of 1,250 leaves per plot. Sampled trees were at least 4m high, so sampled branches were chosen from a distance to avoid selection bias. We sampled branches from the outer part of the crown of all species; for that, we used a slingshot-like apparatus to throw a rope in the selected branch and then bring it down for sampling.  We collected the first 50 leaves from the base to the apex of each branch (according to the standardized protocol of Mendes et al. 2021). Leaves were taken to the laboratory, scanned and had their margins or limb completed to account for leaf area loss to herbivores using the software GIMP. Measurements of leaf area and leaf area lost to herbivory were performed with the R package ‘EBImage’ (Pau et al. 2010). We calculated herbivory in each tree by averaging the percentage of leaf area loss across all 50 leaves sampled.

Plant palatability

Leaf nitrogen content indicates the plant nutritional value to insects (Mattson 1980) and it was estimated through chlorophyll content using a SPAD-502 (Spectrum Technologies, Inc., Plainfield, IL, USA), as leaf chlorophyll concentration and leaf nitrogen content are linearly correlated (Loh et al. 2002). One intact leaf was randomly sampled from each tree (different from those which we measured herbivory), and measurements were taken at the lamina midpoint close to the midrib of the leaves.

We measured leaf thickness as a proxy for sclerophylly. We sampled 10 intact leaves of each individual and measured leaf thickness with a Digital Micrometer (0-25 mm/0.001 mm MDC-Lite 293-821-30). Measurements were taken around the midpoint in both sides of the midrib of the leaves. We calculated mean leaf thickness for each leaf and then for each individual plant, using leaves as replicates per species.

Predation pressure

To estimate predation pressure, we first established a field experiment using artificial model prey – which represent herbivores as caterpillars and other herbivores that do not have anti-predation behavioral forms and strategies (Dáttilo et al. 2016; Roslin et al. 2017). Artificial caterpillars were made with oil-based, odorless, non-toxic soft modeling clay, and were ~ 3 cm long and 5 mm in diameter. All models were green and had a wire passing through their center to attach them to the branch (Figure S1). We established the experiment in May of 2018, at 500 m from the forest plots edges, in different plants than those used for the herbivory experiment, irrespectively of plant species. For each treatment plot we set 15 sampling points across a previously established trail (2 m wide), which were located 10 m apart from each other. Each sampling point consisted of two plants, located at least two meters apart and on opposite sides of the trail. Plants were between 1.5 and 2 m in height, and were chosen arbitrarily. One model prey was fixed distally in an outer branch at ~ 1.5 m height of each plant, totaling 30 artificial prey per plot - each sampling point had a pair of plants with an artificial prey in each. From the total artificial prey, four were colonized by termites and were discarded from our analyses. All model prey remained in the field for four days, and afterwards each model was examined for predation marks, and categorized to major groups of predators (birds, arthropods, mammals) according to Low et al. (2014).

As an additional proxy to our estimation of predation pressure by arthropods we compared the abundances of predatory ants, which are the main arthropod predators in the tropics (Zvereva et al. 2019), across treatment plots. For that, we sampled ant communities in the same study area approximately one year (February 2019) after vegetation and predation experiment samplings. We established six sampling sites in each treatment plot (N = 18); all sampling sites were located at least 250 m from the forest edge, and at least 170 m apart from each other. Each sampling site consisted of a 3 × 3 grid of nine sampling stations located 10 m apart. Each sampling station consisted of unbaited pitfall traps – one in the epigaeic and the other in the arboreal stratum. All pitfall traps were 5 cm in diameter, were partly filled with a salt solution and detergent, and were left open for 48 h. The epigaeic traps were buried with their rims flush to the soil surface, and arboreal traps were tied at a height of 2 m to the trunk of the nearest tree with dbh ≥10 cm. We sorted ants to species, and assigned morphospecies when species identification was not possible. Voucher specimens are held at the Laboratório de Ecologia de Comunidades e Ecossistemas Tropicais (EcoTrop), Universidade Federal de Viçosa (UFV), Brazil.

Data analysis

We adjusted a mixed beta regression model with logit link to test for the effects of fire on herbivory, with tree species as random variable. Results of beta regression modeling are shown in the text and figure as percentage rather than proportion for ease of interpretation. To test whether leaf nitrogen content and leaf thickness varied across treatment plots, we built two generalized linear mixed models, one with leaf nitrogen content and the other with leaf thickness as response variable, and fire treatments as explanatory variable. Both models had Gaussian error distribution, and tree species as a random effect.

We compared predation pressure across treatment plots by considering arthropod and bird marks separately, as these two groups of predators were the most abundant in our study area. We built two generalized linear models using the proportion of predation marks in each sampling point (i.e., a pair of plants with a caterpillar model in each) as response variable - i.e., zero, 0.5, or 1 attack – and fire treatment as explanatory variable. We used binomial error distribution in both models, with a ‘quasi’ correction for overdispersion for the arthropods model. Finally, we classified sampled ants (both from epigaeic and arboreal strata) as predatory or non-predatory (according to Baccaro et al., 2015 and with help of a taxonomist; Julio M.C. Chaul), and used only predatory ants for analyses. We built a generalized linear model with Poisson error distribution using predatory ant abundance (epigaeic and arboreal ants combined) as response variable, and fire treatment as explanatory variable. Ant abundance in each transect was estimated as the sum of occurrence of predatory species in all nine traps of the transect combined.

We conducted all analyses using the software R (R Core Team 2019); we used the glmmTMB package v1.0.1 (Brooks et al. 2017) to adjust mixed beta regression models and the package lme4 to adjust glmm models. We analyzed the residuals to check for distribution suitability and fit in all models. To evaluate differences between fire plot treatments, pairwise contrasts analyses were made between control and two burned treatments combined, which were the most similar levels (Crawley 2013).