Floodplain land cover affects biomass distribution of fish functional diversity in the Amazon River

Land-cover change often shifts the distribution of biomass in animal communities. However, the effects of land-cover changes on functional diversity remain poorly understood for many organisms and ecosystems, particularly, for floodplains. We hypothesize that the biomass distribution of fish functional diversity in floodplains is associated with land cover, which would imply that fish traits affect behavioral and/or demographic responses to gradients of land cover. Using data from surveys of 462 habitats covering a range of land-cover conditions in the Amazon River floodplain, we fitted statistical models to explain landscape-scale variation in functional diversity and biomass of all fish species as well as subsets of species possessing different functional traits. Forest cover was positively associated with fish biomass and the strength of this relationship varied according to functional groups defined by life history, trophic, migration, and swimming-performance/microhabitat-use traits. Forty-two percent of the functional groups, including those inferred to have enhanced feeding opportunities, growth, and/or reproductive success within forested habitats, had greater biomass where forest cover was greater. Conversely, the biomass of other functional groups, including habitat generalists and those that directly exploit autochthonous food resources, did not vary significantly in relation to forest cover. The niche space occupied by local assemblages (functional richness) and dispersion in trait abundances (functional dispersion) tended to increase with forest cover. Our study supports the expectation that deforestation in the Amazon River floodplain affects not only fish biomass but also functional diversity, with some functional groups being particularly vulnerable.

USGS Earth Explorer site (http://earthexplorer.usgs.gov/). The two contiguous Landsat Thematic Mapper images used to map floodplain land cover correspond to river stages of 2.14 m and 1.75 m at the Óbidos gauge. Aquatic macrophyte coverage (Table 1) was mapped using seven ALOS PALSAR swaths (fine-beam, HH-polarized, L-band synthetic aperture radar; resampled from 12.5 to 30 m) acquired during the early rising-water period in each of five years from 2006 to 2010. ALOS PALSAR imagery was obtained from the Alaska Satellite Facility's Vertex Data Portal (https://vertex.daac.asf.alaska.edu/).
Land-cover data obtained from remotely sensed imagery were assembled according to spatial units defined as local catchments (or "lake systems"). Each local catchment contains lakes, interconnecting channels, forest, and areas with herbaceous vegetation and aquatic macrophytes that are hydrologically connected for about six to nine months per year (Arantes et al. 2018). Local catchments are separated from each other by major secondary channels (areas of low elevation) and/or natural levees (areas of high elevation). We mapped 20 local catchments (Fig. 1, median area: 23.4 km 2 ), which encompassed a gradient of forest cover, ranging from 3 to 70%. The macrophyte metric used here was the percent of the catchment having macrophyte cover during the early-rising water period during three or more of the five years imaged (Table  1).

Field surveys
Field surveys were conducted during five expeditions covering four different stages of the annual hydrological cycle at 462 habitat areas (lakes and secondary channels (open water), and flooded herbaceous and forest) within the 20 local catchments (Fig. 1). For each habitat type within each local catchment, and during two dry periods and one rising-, high-, and falling-water period, we collected fish using a standard set of nets with different mesh sizes (11 gillnets measuring 25 x 2 meters, with mesh sizes 20,30,40,50,60,70,80,90,100,120, and 130 mm, and one gillnet measuring 100 x 3 meters, with 180 mm mesh) to catch multiple fish size classes and species. For each season and all habitats sampled within each local catchment, average gillnet sampling effort was approximately 25 hours (SD ~4 hours). For the same seasons and habitats within each local catchment where we collected fish, we measured water temperature, dissolved oxygen concentration, depth and transparency (Table 1). We also estimated the area covered by aquatic macrophytes via visual observations (Table 1). Whereas macrophyte indices obtained via remotely sensed imagery provided large-scale estimates of coverage in local catchments, visual estimates were useful for characterizing an important feature of fish habitat at a local scale matching that of our fish assemblage surveys.

Species classification
We tested relations between forest coverage and fish biomass (CPUE) based on 25 fish groups distributed within 6 categories (Table S1). Measures of total fish biomass in local habitats involved subsets of the 186 species that were surveyed in the region. Species were grouped according to their degree of importance in local fisheries and this classification was based on their relative contribution to total yields landed in the main cities in the lower Amazon (Isaac et al. 2016, Castello et al. 2017. Fish of high-importance (28 species, 11 common names) contributed > 85% of the total landing. Fish of medium-importance (83 species, 31 common names) contributed with 15% of the total landing. Both groups (high-importance and mediumimportance) contain important species for subsistence fisheries in the Amazon (e.g., Colossoma macropomum, Cichla monoculus, Prochilodus nigricans, Mylossoma spp. Myleus spp.) (Batista et al. 2008). Seventy-four species (39 common names) were classified as having low-importance and were rarely landed for sale as food, although some of them are used as bait or sold as ornamental fish.
The other categories comprised functional groups based on trophic, migratory, life history, and swimming/microhabitat-use strategies. We classified species according to eight trophic strategies based on dietary information from published reports (Table S1). Herbivores (18 species) feed predominantly on C3 or C4 plant material (seeds, fruits or leaves) and on filamentous algae. Omnivores (47 species) ingest combinations of plant material, detritus, and invertebrates. Detritivores (28 species) predominantly ingest fine particulate organic matter and non-living macrophyte tissues, but also on filamentous algae. Invertivores (23 species) ingest variable fractions of aquatic and terrestrial insects, microcrustaceans from the benthos or water column, spiders, shrimps, and mollusks. Planktivores (10 species) ingest phytoplankton, zooplankton, and occasionally small amounts of plant material and detritus. Piscivores (45 species) ingest adult, juvenile, or larval fish, either whole or in pieces, including scales and fins. Piscivores-macroinvertivores (14 species) feed on the same sources as piscivores but also ingest significant fractions of diverse terrestrial or aquatic macroinvertebrates (e.g., Ephemeroptera, Chironomidae, Coleoptera, Crustacea, etc.).
We classified species according to four migratory strategies based on information on dispersal behavior from published reports. Migratory strategies of Amazon fish often are related to reproduction and/or feeding ecology and influenced by seasonal hydrology and physicalchemical conditions of habitats in the riverscape. Sedentary (55 species) were resident species that spend their entire life-cycles within floodplain habitats eventually performing shortdistances movements. Sedentary species were small-bodied species, or had territorial behavior, or are known to be strongly associated with substrates or complex structured habitat (e.g., tree branches and aquatic vegetation). Species performing local migration (120) comprised a diverse group of fishes that migrate laterally from floodplain lakes or river channels onto flooded floodplain habitats following closely the dynamic 'pulsing' of water levels (Junk et al. 1989;Fernandes 1997;Carolsfeld 2003;Castello 2008). Species performing regional migration (8 species) migrate onto flooded floodplains habitats during high waters, but also conduct longitudinal migrations (often hundreds of kilometers) along river channels to spawn, particularly during falling waters (Goulding 1980;Ribeiro de Brito and Petrere 1990;Benedito-Cecilio and Araujo-Lima 2002;Barthem and Fabré 2004). Long-distance migrators (3 species) were species that migrate thousands of kilometers along river channels, though their juveniles often inhabit floodplain lakes (Barthem et al. 1991(Barthem et al. , 2017. We classified species according to six life history strategies based on maximum body size, size at maturation, batch fecundity, and parental investment per individual offspring (Table S1) (following Winemiller & Rose (1992) and Röpke et al. (2017)). Life history strategies identify suites of intercorrelated functional traits and their associations with patterns of environmental variation involving abiotic factors, disturbance regimes, resource availability and quality, population density, risk of predation or parasitism, and challenges for dispersal (Winemiller & Rose 1992, Winemiller 2005. Sixteen species were classified as equilibrium strategists with maturation at small size (<120 mm standard length, SL), having low batch fecundity, large oocytes, well-developed parental care, and maximum body size between 97 -269 mm SL. Sixteen species were equilibrium strategists with maturation at large size (>170 mm SL), with low batch fecundity, large oocytes, well-developed parental care and maximum size >400 mm SL. Seventy-three species were periodic strategists with maturation at small size (between 63 -148 mm SL), having varied batch fecundity size (average ~ 4,000), small oocytes, maximum size between 137 -410 mm SL and no parental care. Forty-three species were periodic strategists with maturation at large size (>164 mm SL), with batch fecundity highly variable, small oocytes, no parental care and maximum size > 253 mm SL. Thirty-two species classified as intermediate strategists had batch fecundity between 1,000 and 9,000, relatively large oocytes, and intermediate development of parental care. Five species classified as opportunistic had small size (between 26-113 mm SL), early maturation (<60 mm SL), high and sustained reproductive effort but low batch fecundity and no parental care (Röpke et al. 2017).
Finally, we classified species according to five strategies of swimming/microhabitat use based on morphological traits. We based our classification on the classification of Arantes et al.
(2017) that uses traits associated with swimming performance and vertical position within the water column during foraging, phenotypes that influence fitness along gradients of habitat structural complexity and other environmental features (Gatz 1979, Winemiller 1991. Nektonic maneuverable fishes (41 species) had laterally compressed bodies and superior mouth position, whereas nektonic burst swimmers (18 species) had fusiform bodies and terminal mouth position. Both groups had morphological traits associated with efficient swimming performance based on a hydrodynamic body and feeding within the water column. Surface dwellers (2 species) had intermediate lateral body compression, superior mouths and either deep or fusiform bodies. Epibenthic maneuverable fishes (57 species) were a diverse group having relatively deep bodies that are less hydrodynamic than nektonic maneuverable fishes but efficient in making lateral and vertical turns. The two last groups also had more dorsally than laterally positioned eyes. Most Benthic-slow (36 species) and Benthic-fast (23 species) had relatively wide bodies, dorsally located eyes, and inferior mouths, which are characteristic of bottom dwellers. Benthic-fast fishes had higher muscle mass and larger pectoral and caudal fins ratio areas than benthic-slow fishes; morphological traits associated with increasing swimming performance. A few benthicfast (1 species) and benthic-slow (3 species-e.g., Hoplias malabaricus) fishes had terminal or superior mouths. Gymnotiformes (8 species) comprised a diverse group of electric fishes, either substrate or aquatic vegetation dwellers, that are inactive during daylight but actively forage during the night using weak electric organ to locate their prey (Fernandes et al. 2004;Carvalho et al. 2009). As mentioned herein, we did not include gymnotiforms, long-distance migrators and opportunistic strategists in the analyses due to their small sample size and complete absence at some levels of the categorical variables (i.e., seasons and habitat types).

Modelling associations of fish biomass and forest
Fish biomass (CPUE) for each fish group was modeled as a function of linear predictors within a generalized linear model (GLM) framework using a Poisson-Gamma distribution: where, g() is the log link function, i represents the data sampling habitat, y is fish biomass standardized by sampling effort (CPUE), l is a row vector of three measures of land cover (forest, open water, large-scale estimate of aquatic macrophyte cover), e is a row vector of two first PCA axis representing environmental covariates (PC1 represent gradients of transparency and dissolved oxygen and PC2 macrophyte cover, temperature and depth, Fig S1), s is a season factor variable, h is the habitat type factor, m is an indicator for the presence of management, and is the error term. Frequent zero catches, such as observed for our CPUE data, is a common issue in fishery modeling that have been addressed in a straightforward manner within the GLM framework by using a Poisson-Gamma distribution from the family Tweedie, the set of exponential distributions indexed by a power parameter (Jorgensen 1987;Peel et al. 2013). This distribution handles zero values uniformly with positive and continuous values and it was found to outperform other models used for CPUE data containing a point mass at zeros (e.g., delta models, generalized linear models with an additive constant) (Shono 2008;Carvalho et al. 2010;Li et al. 2011). For a random variable Y that is distributed Tweedie, E(Y) = µ and Var(Y) = φµ p where µ is the mean of the distribution, φ is the dispersion parameter, and p is an extra parameter (power parameter) that controls the variance of the distribution. The Tweedy family of distribution include the Normal (when p=0), Poisson (p=1) and Gamma (p=2) distributions. When, p є (1,2), such as in our study, the Tweedie distribution assumes the form of a compound Poisson-Gamma distribution, which allows modelling a variable that has both discrete and continuous components.
We assessed the quality of models fit via visual inspection of plots of model residuals. We used randomized quantile residuals as recommended by Dunn and Smyth (1996) and Dunn (2009) for model fits using the Tweedie family, and as used in several cases of fishery modeling studies ( e.g., Tascheri et al. 2010;Peel et al. 2013). The randomized quantile residuals were examined for heteroscedasticity and approximate normality. Because the data have a non-Normal nature, Pearson and deviance residuals are intrinsically non-Normal, and difficult to interpret due to a large proportion of exact zeros (Peel et al. 2013). Instead, quantile residuals have an exact Normal distribution provided if correct response distribution is used and indications of non-Normality are interpretable as deficiencies in the model (Dunn 2009). Spatial dependence is a common feature of ecological studies because data collected at sites that are located closer to each other tend to be more similar than data collected from sites that are farther apart (Legendre and Fortin 1989). Therefore, we also tested for spatial autocorrelation of the models' residuals using Moran's I statistics. Moran's I is a correlation coefficient that measures the overall spatial autocorrelation of the data (I=-1 indicates a perfect dispersion, I=0 indicates a perfect randomness, and I=1 a perfect clustering) (Sokal and Oden 1978). We tested whether Moran's statistics values differ from random by comparing the observed Moran's I from each model residuals with bootstrapped generated Moran's I. The Moran's I statistic was bootstrapped by randomly assigning longitude and latitude values to the residuals values and Kernel density estimates of the Moran's I statistics were used to calculate a 95% highest density region (i.e., confidence interval). Moran's I results indicated that data is not strongly dependent upon space across distances (see Figures S2); therefore, incorporating spatial autocorrelation was not a concern for our models.
Analyses were performed in R v. 3.3.3. Models were fitted using the statmod (Giner and Smyth 2016) and Tweedie (Dunn and Smyth 2005) packages and Moran's I were calculated using the ape (Paradis et al. 2004), geoR (Ribeiro Jr et al. 2001) and fields (Nychka et al. 2005) packages.  Figure S1. Principal components analysis with habitats ordination according to the local environmental variables: depth, dissolved oxygen, transparency, temperature and the local-scale estimate of aquatic macrophyte cover (macro.obs) ( Table 1). PC1 is associated with a gradient of transparency (score -2.4) --dissolved oxygen (score 2.2) and PC2 with a gradient of macrophyte (score -1.6) and temperature (score -1.4) --depth (score 2.0). Data were standardized and PCA was performed using stats library and prcomp function in R software (R Core team 2017) Figure S2. Partial effects of management on relative biomass (CPUE) for each fish group. The reference level is absence of management (coefficient = 0), meaning that the coefficient size for presence of management reflect its size being compared to that of absence of management while controlling for the effects of other variables. Figure S3. Partial effects of habitat type on relative biomass (CPUE) for each fish group. Habitat type are: Lake, Flooded herbaceous (F.her), Channel (Chan), and Flooded forest (F.for) (See Table 1). The reference level is lake habitat type (coefficient = 0), meaning that the coefficient size for the other habitat types reflect their size being compared to that of lake while controlling for the effects of other variables. Figure S4. Partial effects of season on relative biomass (CPUE) for each fish group. Season are: High water (Hig), Low water (Low), Rising water (Ris), and Falling water (Fall) (See Table 1). The reference level is high water (coefficient = 0), meaning that the coefficient size for the effect other seasons reflect their size being compared to that of high water season while controlling for the effects of other variables. Figure S5. Randomized quantile residuals versus linear predictor for all models. Fish category is indicated on the top of the graphs (see Table S1). Table S1-Fish species (common and scientific names and families) and their degree of importance for regional fisheries, and trophic, migratory, life history and microhabitat use strategies. Trophic strategies classification was based on Barbarino and Winemiller (2003) Mérona and Mérona (2004), Santos et al. (2008), Shibuya and Zuanon (2013), Rӧpke et al. (2014), Correa and Winemiller (2014), Lopes et al. (2009) and Rӧpke et al. (2017). Migratory behavior classification was based on Goulding (1980), Junk et al. (1989), Ribeiro de Brito and Petrere (1990), Barthem et al. (1991), Fernandes (1997)