Data for: The fish-mangrove link is context dependent: tidal regime and reef proximity determine the ecological role of tropical mangroves

Full description

Data was collected as described in: The fish-mangrove link is context dependent: tidal regime and reef proximity determine the ecological role of tropical mangroves Tropical mangroves are known to support fish production, but natural variability in the link between mangrove habitats and fish populations undermines our ability to manage, conserve and restore this ecological relationship. This is largely due to undefined context-dependence in the use of mangroves by fish. We collected a spatially extensive dataset of 494 mangrove fish assemblages using standardised Remote Underwater Video surveys of mangrove edge habitat from 5 environmentally heterogenous regions in the Indo-Pacific. We used machine learning methods to define contextual limits of the use of mangroves by reportedly mangrove-affiliated fish. Remote Underwater Video Unbaited Remote Underwater Videos (RUVs) provide a low disturbance snapshot of the fauna naturally associated with coastal habitats. All video surveys were conducted in prop-root habitat of mangroves of the genus Rhizophora, and were within the first 2 metres of the seaward edge of the forest. RUVs were collected when the forest was inundated, with a roughly even distribution across flood, high and ebb tidal states, and rarely during low tide. This was captured in the variable ‘tidal direction’ (see supplementary figure S1) and included in analysis. All video units across all locations were unbaited and deployed for at least 15 minutes, at least 20 m apart. This minimum sampling distance is favoured to achieve high-replication RUV studies (e.g. Bradley et al., 2019), but it does not completely guarantee spatial independence, particularly for large cursorial fishes. Therefore, most samples in this dataset were separated by >50 m. All video units consisted of an underwater camera positioned parallel to the horizon, attached to a weighted landing frame that raised the camera off the substratum. Only samples with a visibility range of greater than 0.5 m were retained. This produced 494 video samples for analysis, from across our 5 regions (for regional breakdown, see Table 1). For optimal and consistent comparison of fish assemblages, 15 minutes of video was watched from each video sample (following Piggott et al., 2020). Species presence-absence, and species richness data was extracted for use in statistical analyses. Abundance data was avoided due to the outsized influence of highly abundant taxa on total fish group abundance. This sampling method is not appropriate for small Gobiiform fishes and cryptic fishes, therefore this group and their relationship with mangroves is not assessed in this study. Seascape life-history strategy categorisation Fish detected in RUVs were identified to the lowest taxonomic grouping possible, and where possible, juvenile stages were differentiated from adult stages, using stage-specific colour patterns and other morphological characteristics (following Bradley et al., 2019, 2021). We examined two distinct sets of fauna – coastal-estuarine fauna which are not associated with reefs, and reef-associated fauna, which are found specifically associated with coral reefs at some point in their life-cycle. For these reef fish, given the known life-history specificity of the role of mangroves, we assessed adults separately from juveniles. We also distinguished between species that are reef associated throughout their life-history (referred to here as ‘juvenile reef specialist fish’), from species that are known to typically use non-reef habitats at some point in their lifecycle (referred to here as ‘multihabitat reef fish’ following Sambrook et al., 2019). Using a simple classification scheme, this produced four distinct groups of fish (see supplementary materials, Table S1). The ‘coastal-estuarine fish’ group were individuals of any life stage of taxa that had a reported association with coastal-estuarine areas and no reported association with coral reefs throughout their life-history. The ‘adult reef fish’ group were fish that were visually identified as adults in RUVs that had a reported association with reef habitat during their adult phase. The ‘juvenile reef specialist fish’ group were fish that were visually identified as juveniles in RUVs that had a reported association with reef habitat, and that were not typically known to use non-reef habitats as either adults or juveniles. The ‘juvenile multihabitat reef fish’ group were fish that were visually identified as juveniles in RUVs that had a reported association with reef habitat during adult life phases, and that were known to use non-reef areas during some part of their non-planktonic life-history. Where consensus could not be reached on taxonomic identity, individuals were assigned to the level of taxonomic grouping (e.g. genus) where consensus was achieved. Where consensus could not be reached on life history stage (i.e. juvenile or adult) a conservative approach was taken to assigning juvenile status. For this reason, the ‘adult reef fish’ group likely contains subadults that are no longer visually distinguishable from adults. Individuals that could not be confidently assigned to a specific group were not included in analysis. Information for ecological and life-history categorisation of each species was gathered via FishBase (Froese & Pauly, 2017), refined using relevant species guides (Allen, 1985; Allen et al., 2012) and primary research (Newman & Williams, 1996) and further supplemented by the expertise of relevant authors. Contextual variables A range of different variables were used to examine context-dependence, namely salinity, distance to reef, tidal range, tidal state, biophysical typology, substratum and depth. Table 2 provides a detailed description of each variable, the methodology used to measure each variable, and the associated hypothesis as to why each could be important in determining the use of mangrove habitat by fish (Table 2; see supplementary methods for text 1 for details). Correlations between variables and imbalances in the dataset (see supplementary material, Figure S1) are representative of natural variation and do not violate underlying assumptions in the machine-learning analyses employed. Results were interpreted with these imbalances in mind. Software/equipment used to create/collect the data: Unbaited Remote Underwater Videos (RUVs) provide a low disturbance snapshot of the fauna naturally associated with coastal habitats. All video surveys were conducted in prop-root habitat of mangroves of the genus Rhizophora, and were within the first 2 metres of the seaward edge of the forest. RUVs were collected when the forest was inundated, with a roughly even distribution across flood, high and ebb tidal states, and rarely during low tide. This was captured in the variable ‘tidal direction’ (see supplementary figure S1) and included in analysis. All video units across all locations were unbaited and deployed for at least 15 minutes, at least 20 m apart. This minimum sampling distance is favoured to achieve high-replication RUV studies (e.g. Bradley et al., 2019), but it does not completely guarantee spatial independence, particularly for large cursorial fishes. Therefore, most samples in this dataset were separated by >50 m. All video units consisted of an underwater camera positioned parallel to the horizon, attached to a weighted landing frame that raised the camera off the substratum. Only samples with a visibility range of greater than 0.5 m were retained. This produced 494 video samples for analysis, from across our 5 regions Software/equipment used to manipulate/analyse the data: We used Random Forest classification, a high-accuracy machine learning technique, to determine variable importance and model the relationship between contextual variables and fish groups. Random Forest is a non-parametric statistical classifier that employs classification trees to partition data into homogeneous subgroups using predictor variables, until no further reduction in group heterogeneity can be achieved (Breiman, 2001). Random Forest grows many trees, each with a randomised subset of data and predictor variables, and then tests each tree with the observations in the respective excluded data (out-of-bag (OOB)). Aggregating the proportions of OOB predictions across the entire ‘forest’ of trees allows for the estimation of probability of class membership based on predictor variables without the dangers of over-fitting associated with single trees. The contribution of each variable to model accuracy (variable importance) is determined by comparing the misclassification rates when using actual and randomly permuted values for each predictor variable (Cutler et al., 2007). To visualize the relationship between predictor variables and the response variable, we used the feature contribution method (Palczewska et al., 2014), which extracts the influence of the variable of interest on the prediction for each observation from the Random Forest model. To examine context-dependence in the presence of each fish group, we built a Random Forest model for each fish group, calculating variable importance and the feature contributions of each variable. Using species richness data of the fish group as the response factor, Random Forests of 5000 trees were grown, weighted by the prior proportion of presence vs absence of the inshore-user group. For each Random Forest, the OOB error rates were calculated to evaluate model fit, and variable importance was calculated using the permutation process described above. Feature contributions were calculated for each predictive variable, however, only the two most important variables were selected for interpretation to avoid the use of variables that contribute little to model accuracy. In feature contribution plots, the influence of the predictor variable on class prediction (species richness of the inshore-user group) was displayed for each observation, along with an average for each value of the contextual variable to aid visualisation of the relationship, from which goodness of fit was calculated (Welling et al., 2016). This provides a model of the relationship between a fish group and Rhizophora habitat. A strong positive contribution indicates an increased likelihood of encountering species of that fish group, and strong negative contribution indicates a reduced probability of encountering species of that fish group. A contribution close to zero indicates that the variable had little influence on prediction at that value. Collinearity among variables does not reduce prediction accuracy, but must be considered in the interpretation of the resulting model. All analyses were performed using R version 3.3.3 (R Core Team, 2017). Random Forests were built using the ‘randomForest’ package (Liaw & Wiener, 2002), and feature contribution plots were displayed with the ‘forestFloor’ package (Welling et al., 2016).