Dry season habitat use of fishes in an Australian tropical river

The modification of river flow regimes poses a significant threat to the world’s freshwater ecosystems. Northern Australia’s freshwater resources, particularly dry season river flows, are being increasingly modified to support human development, potentially threatening aquatic ecosystems and biodiversity, including fish. More information is urgently needed on the ecology of fishes in this region, including their habitat requirements, to support water policy and management to ensure future sustainable development. This study used electrofishing and habitat survey methods to quantify the dry season habitat use of 20 common freshwater fish taxa in the Daly River in Australia’s wet-dry tropics. Of twenty measured habitat variables, water depth and velocity were the two most important factors discriminating fish habitat use for the majority of taxa. Four distinct fish habitat guilds were identified, largely classified according to depth, velocity and structural complexity. Ontogenetic shifts in habitat use were also observed in three species. This study highlights the need to maintain dry season river flows that support a diversity of riverine mesohabitats for freshwater fishes. In particular, shallow fast-flowing areas provided critical nursery and refuge habitats for some species, but are vulnerable to water level reductions due to water extraction. By highlighting the importance of a diversity of habitats for fishes, this study assists water managers in future decision making on the ecological risks of water extractions from tropical rivers, and especially the need to maintain dry season low flows to protect the habitats of native fish.

. Location of fish sampling locations in the Daly River catchment. Fish sampling sites are indicated with site identification numbers (1)(2)(3)(4)(5)(6). Refer to Table 1 for site details. The inset shows the location of the study area in northern Australia.  www.nature.com/scientificreports www.nature.com/scientificreports/ Methods study area and focal reach. The Daly River catchment (mean annual streamflow 8653 GL 37 ) is located in the wet-dry tropical climate region of northern Australia (Fig. 1), covers ~53,000 km 2 , and has a largely intact savannah ecosystem throughout the catchment 38 . The catchment's population is 10,000 people, of whom 28% are indigenous 39 . The Katherine (largest tributary) and the Daly Rivers are considered to be in relatively good ecological condition 38,40 . All rivers in the catchment are unregulated (no dams or weirs), but some groundwater extraction occurs for agriculture and water supply and this is likely to increase in the future 18 . The rivers have a characteristic wet-dry tropical flow regime, with predictable and large wet season flows with high interannual variability 41 . Mean annual rainfall averages 1070 mm 42 across the catchment and is highly seasonal, with the majority falling during the wet season months (November-April), and negligible rainfall during the dry season (May-October). Flow in the Daly River and some tributaries is perennial, with significant groundwater inputs coming from two underlying aquifers, resulting in a lengthy period of continuous and stable dry season base flows (May-November). The Daly and Katherine rivers are sand-bed rivers that contain dispersed bedrock outcrops and gravel bars; river banks are typically steep, rising 15-20 m above the river bed 42,43 . Riparian vegetation in the catchment is largely intact and natural, consisting of Eucalyptus woodlands, Melaleuca forests and closed monsoon rainforests 44 .
This study occurred along a 120 km long reach of river from near Katherine township downstream to Oolloo crossing (Fig. 1). Six sites that were accessible by boat were selected; these sites represent the available habitat types for fish within the reach (Table 1).
Fish and habitat sampling. Fish were surveyed biannually in both the early and late dry season, over a 10-year period from 2006-2015 (n = 89 sampling events in total). A detailed description of fish sampling methods can be found in Stewart-Koster et al. 19 and Chan et al. 18 , but is briefly summarised here. Within each sampling site (500-1000 m reach length, 5-100 m wetted width), fish were sampled at discrete multiple locations (shots or replicates) using a boat or backpack electrofishing (pulsed DC), depending on water depth. Electrofisher settings were adjusted to maximise efficiency of collecting fish with minimum power. At least 15 electrofishing shots of five minute elapsed duration were undertaken for each sampling event, and shots were stratified to ensure each available habitat type within each site was sampled at least once. Prior examination of sampling efficiency has revealed that 15 shots yields an accurate estimate of species composition and assemblage structure within each site (Kennard et al. unpubl. data). Electrofishing shots were conducted in as homogenous area as possible and averaged 77 m (±48 SD) in length (range = 5-263 m). At the completion of each electrofishing shot, fish captured were counted and identified to species, measured (standard length (SL) in mm) and returned alive to the point of capture. The total number of fish from each species observed and reliably identified but not caught during each electrofishing shot was also recorded. Catch data were converted to catch per unit effort (CPUE-total number of individuals caught and observed per electrofishing shot). Some species captured during sampling were separated into juvenile or adult age classes (Table 2), where the minimum length used for juvenile determination for these species was <150 mm SL, except Lates calcarifer, where the minimum length was <300 mm SL.
A total of 20 variables describing fish habitat characteristics was measured within the area of each discrete fishing location (sampling shot) by a second field team. A full description of the methods employed in describing habitat structure is available in Kennard et al. 45 . Briefly, five replicate measures within each electrofishing shot were measured for depth (initially using a staff, then a Garmin 150 Fish Finder) and mean water velocity (Swoffer ™ 2000, and Hach ™ FH950 flow meter, measured as 20% total depth). The total percentage contribution of each mesohabitat type (i.e. riffles, runs, pools) and substrate composition (modified Wentworth scheme-see Pusey et al. 46 ) was visually estimated within each shot (summed to 100%). The proportional contributions of a range of submerged structures (including wood, macrophytes, root masses etc. Table 3) were also visually estimated for each shot (contributions of these variables were not required to add up to 100%). The fish collected within each electrofishing shot were assigned the habitat characteristics of that shot and these data were used to examine the local habitat use of each taxon. Data analysis. The data were filtered to remove 26 rare taxa (i.e. those taxa present in <5% of samples and/or with <50 individuals collected in total). Neoarius graeffei, N. bernei and N. midgleyi were combined to genus level (i.e. Neoarius spp.), due to either low counts or unreliable identification to species. This resulted in a total of 22 taxa suitable for analysis with six taxa separated into juvenile and adult age classes (Table 2). For the entire study period, 54,739 taxa-habitat data points were available for analysis.
Fish habitat variables were assessed for their potential to influence the abundance of each fish taxon across electrofishing shots using boosted regression trees (BRTs). Modelling can improve ecological understanding and allow evaluation of generalizability 47 , and thus has advantages over traditional approaches to quantifying habitat use (e.g. by comparing habitat availability to frequency of use to infer habitat selection or preference) 48,49 . Tree-based modelling approaches such as boosted regression trees (BRTs) are widely used for developing predictive species-environment relationships 50 and have been applied to quantify fine-scale fish habitat use (e.g. Sharma et al. 51 , Greenwood 52 ). BRTs are a decision tree based method which uses a boosting technique to combine large numbers of relatively simple tree models adaptively to optimize predictive performance 53 . Some advantages of BRTs include their ability to model complex and nonlinear relationships and to accommodate missing data and large numbers of potential predictor variables by ignoring non-informative variables 53 . BRTs also perform better than other regression techniques under conditions of high multicollinearity, especially when a large number of variables are analysed.
We assumed a Poisson distribution for the response variable in all models since our data is composed of individual fish counts in each shot. The best combination of parameters required by BRTs (learning rate, tree complexity and bag fraction) was identified for every model using cross-validation 52 . Cross-validation was www.nature.com/scientificreports www.nature.com/scientificreports/ automatically repeated for learning rates from 0.001 to 0.05 (steps of 0.002), tree complexities of 1-3 and bag fractions of 0.5 and 0.75, which span the range of likely optimal values 52 . The combinations that generated the lowest mean cross-validation deviances, calculated from at least 1000 trees, were used for the final models. Following the derivation of full models with all variables, models were investigated to establish whether irrelevant predictors could be removed (procedure as detailed in Elith et al. 52 ).
The percentage relative contributions of each habitat variable, including the effect of season (early or late dry season) were calculated from the BRTs for each taxon. The relative contribution of a variable describes the proportion of variation in the data explained by that variable relative to all other variables in the model, scaled to 100 53 . For each taxon, partial-dependence functions were fitted for the top two most influential variables to visualise relationships between fish abundances and each variable. The fitted function for a given variable incorporates the average effect of all other variables (Elith et al. 2008), with positive fitted function values suggesting an increase in abundance, while negative values suggest a decrease. Following initial BRT analyses, upon examination of percent deviance explained for each model, two taxa (Ambassis spp. and Neosilurus hyrtlii) with low calculated percentage deviance (<20%) were removed from further consideration ( Table 4). Hence BRTs for a total of 20 taxa with five taxa separated into adult and juvenile categories are reported, noting that each model has different percent deviance and hence explanatory power ( Table 2).
Hierarchical cluster analysis with complete linkage distance was used to group fish taxa abundance and habitat variables into habitat guilds. The number of classes (guilds) in the clustering dendrogram was determined visually in a way that allowed for a biological interpretation. Pairwise comparisons were conducted between each taxon and all habitat variables using Pearson's correlation coefficient. All analyses were performed in R version 3.3.2 54 using the 'gbm' package version 2.1.3 55 and the 'dismo' package version 1.1-4 56 for the BRT models plus customised code written and described by Elith et al. 52 , and the 'hclust' function of 'vegan' package version 2.4-4 57 for the cluster analysis. ethical statement. All methods were carried out in accordance with ethical guidelines and regulations and approved by Charles Darwin University's Committees for Animal Ethics and Human Research Ethics. Fish sampling was also conducted under NT Fisheries S17 Fisheries Act Permit.

Results
Twelve habitat variables were identified as being the top two most important contributors to the BRT models for the 20 taxa (Table 5). Seasonal variation in habitat use was minimal, with this factor having a low relative influence in BRT models for all taxa (Table 5). Water depth was identified as the top contributor for 16 taxa (Fig. 2  Velocity was the second most important contributor to the BRT models, occurring as one of the top two most important contributor for 12 taxa overall, and was the top habitat contributor for three taxa (Fig. 3, Table 5). Taxa that were most abundant in low velocity habitats (0-30 cm/s) included adult O. lineolata C. stramineus, G. aprion, M. australis, M. mogurnda, T. chatareus and adult S. butleri. Both juvenile H. fuliginosus and P. ordensis were most abundant in fast flowing habitats (>90 cm/s). The remaining taxa, juvenile S. butleri, N. ater and Neoarius spp., were most abundant in habitats with moderate to high velocities (30-90 cm/s).
Submerged wood, root mass, leaf litter, rock, sand, fine gravel, coarse gravel, bedrock, filamentous algae and overhanging vegetation were the next most important contributors to the BRT models (Fig. 4, Table 5). The abundance of both juvenile and adult L. calcarifer, adult O. selheimi, and N. erebi, were all higher in habitats with over 50% submerged wood. Similarly, the abundance of both juvenile and adult O. lineolata, as well as juvenile O. selheimi was higher in habitats with over 40% root mass. Likewise, the abundance of L. unicolor and juvenile S. butleri was also higher in habitats with over 15% leaf litter and 40% rock, respectively. The abundance of L. triramus was higher in habitats with over 80% sand and fine gravel, respectively. Similarly, the abundance of Glossogobius sp. and adult H. fuliginosus was higher in habitats with 50-75% and over 40% coarse gravel, respectively. In contrast, the abundance of A. percoides, C. stercusmuscarum, S. kreffti and adult O. selheimi was lower in habitats with over 15% leaf litter, 60% bedrock, 30% aquatic vegetation and 10% overhanging vegetation, respectively.
Cluster analysis of the habitat data revealed four distinct habitat-use guilds (Fig. 5a). We interpreted these groups qualitatively, using information on taxon body size (Table 2) and the taxon correlations with each habitat variable (Fig. 5b). The four habitat use guilds comprised: Guild I: fishes occupying deep pools containing root masses and undercut banks, Guild II: large-bodied fishes occupying deep pools containing wood, Guild III: a mixed habitat use guild of small-bodied fishes, and Guild IV: small-bodied fishes occupying shallow riffles with high water velocities and coarse substrates.
Taxa associated with Guild I had a range of body sizes and included both juveniles and adults of O. lineolata, M. australis, G. aprion, adult S. butleri, N. erebi, N. ater, Neoarius spp. and T. chatareus (Fig. 5b). Taxa within Guild II were large-bodied fishes including both juvenile and adult L. calcarifer, adult H. fuliginosus, adult O. selheimi, S. kreffti and P. ordensis. Guild III consisted of six small-bodied taxa with a range of habitat associations. This included L. triramus which was associated with shallow habitats and higher densities of fine gravel and sand, M. mogurnda which was associated with higher densities of coarse gravel, both C. stercusmuscarum and juvenile O. selheimi which were associated with higher bedrock cover, undercut banks and root masses; C. stramineus was associated with increasing densities of aquatic vegetation and filamentous algae, and A. percoides was associated with shallow habitats with increasing bedrock cover. Guild IV consisted of small-bodied fishes occupying shallow riffles with high water velocities and coarse gravel and/or rock substrates (Fig. 5b). This included juvenile H. fuliginosus, juvenile S. butleri, Glossogobius sp. and L. unicolor.

Discussion
This study used a multi-year and data-rich dataset to quantify the dry season habitat use of freshwater fishes in an Australian wet-dry tropical river. Although many taxa displayed distinct associations with the measured habitat variables, depth and velocity were the key habitat variables for most of the taxa examined (16 and 12 taxa, respectively). Other important habitat variables included submerged wood, root masses, vegetation and sediment type. The analysis also identified four distinct fish habitat guilds according to body size and habitat associations. Whilst it is important to also consider taxon-specific differences in habitat use, the use of fish habitat guilds here may have advantages for habitat management in taxa rich systems such as tropical rivers, and also for forecasting habitat use in other systems or similar taxa (e.g. Leonard & Orth 15 ). Depth and velocity are often identified as important factors in determining fish assemblage structure in rivers (e.g. this study; Jackson et al. 8 ; Kennard et al. 58 ). These factors characterise the geomorphological complexity of rivers 59 and hence the availability of habitats for fish; they are also likely to influence fish behavioural characteristics, such as swimming type, reproductive and feeding behaviours. Not surprisingly then, the depth and velocity requirements of fish are common considerations in river habitat and flow management actions, such as environmental flow determinations (e.g. Bunn & Arthington 4 ; Poff et al. 60 ) and habitat restoration activities 61 . In this study, 15 taxa were associated with moderately deep and low velocity habitats or pools. Pool habitats may have higher species diversity relative to other habitat types, because: (i) they are likely to be more effectively buffered against changes in habitat structure, water level and water quality 58,62 , (ii) they contain higher structural diversity (submerged wood, undercut banks, root masses, etc.) for feeding and shelter 63 , and (iii) they usually cover a greater surface area compared to other mesohabitats, such as riffles and runs 9 .
Four taxa had higher abundance in riffle habitats, including L. unicolor, Glossogobius spp. and juveniles only of S. butleri and H. fuliginosus (Guild IV). Juveniles of the terapontids H. fuliginosus and S. butleri demonstrated ontogenetic habitat shifts, with juveniles using riffle habitats before moving to deeper, slow-flowing pool habitats as adults (deep, structured pools habitat guild, see below). These findings confirm previous research and observations for these two species 18,19,29,64 . Riffle habitats may be providing a suitable refuge habitat for these four small-bodied taxa from piscivorous and avian predators, and/or discrete feeding or reduced competition  65 ). Both Glossogobius spp. and L. unicolor were grouped into the riffle-dwelling habitat guild, due to their association with shallow depths and coarse gravel. This supports in part previous research by Rayner et al. 24 that suggested that Glossogobius sp. 1 (identified in their study) was a fluvial specialist, preferring high velocity, shallow water habitats. Our analysis also provides further evidence that L. unicolor prefers shallow habitats and coarse gravel substrates, but with little velocity preference, thus this species may utilise broader habitat types with various velocities compared to the three riffle-dwelling species grouped in this guild.
In this study, the juveniles and adults of L. calcarifer, adult O. selheimi and adult H. fuliginosus, S. kreffti and P. ordensis, showed a positive association with deep, densely structured habitats composed of submerged wood (Guild II). A number of previous studies have shown that riverine fish assemblages are influenced by the physical structural complexity of river channels, including the presence of large submerged woody debris and instream vegetation 66 . Pettit et al. 63 examined the distribution and movement of large woody debris and its importance as a fish habitat in the Daly River. Using a sub-set of the same fish data analysed in the current study, they demonstrated that fish species richness, diversity and abundance were not correlated with the proportion of wood present at a reach scale. However, juveniles and adults of L. calcarifer and adults of H. fuliginosus were strongly associated with wood cover 63 . Large submerged wood is considered to be an important contributor to habitat heterogeneity, through providing overhead cover which reduces predation risk, and also camouflage for predators to ambush their prey, among other functions 66 .
Fishes associated with pools that contain root masses and undercut banks (Guild I) included a range of taxa and life stages, including O. lineolata, G. aprion, N. ater, N. erebi, M. australis, T. chatareus and adult S. butleri. These taxa were all positively associated with a high density of both root masses and undercut banks, except Neoarius spp., which was more strongly associated with sandy substrates and moderate depths. The presence of both undercut banks and root masses serve as important cover for O. lineolata 29 , which was found to have a strong association with this habitat type in this study. Rainbowfish (Melanotaenia sp.), have previously been recorded as abundant in habitats composed of both root masses and undercut banks, as well as leaf litter 46 . Mouth almighty, G. aprion, a small-bodied predatory fish, may be associated with this habitat type in order to hide and ambush its prey 22,67 . Mouth almighty have previously been classified as a microhabitat generalist 24 , but has also been   www.nature.com/scientificreports www.nature.com/scientificreports/ described as having a close association with root masses, leaf litter and aquatic vegetation 22,67 . Undercut banks have also been suggested as important daytime resting habitat for taxa within the Plotosidae and Terapontidae families 68 , which supports the findings for N. ater and adult S. butleri in this study. In contrast, the presence of www.nature.com/scientificreports www.nature.com/scientificreports/ overhanging vegetation may be more important for T. chatareus, since surface-dwelling and terrestrial invertebrate prey which occur in this vegetation type feature heavily in the diet of this species 67 . The presence of root masses, undercut banks and vegetation are suggested to function as an important refuge and/or foraging habitat for some species 46,68 . Indeed, many of the taxa in this habitat guild are regular prey of L. calcarifer 69 , which supports the hypothesis that the taxa in this guild are using these habitats as a predator refuge. Studies from other systems have also shown that species of small-bodied fishes (e.g. cyprinids) rarely co-exist in the same habitat type as larger predatory species, such as northern pike, Esox lucius 70 , and smallmouth bass, Micropterus dolomieu 71 . This further supports the predator refuge hypothesis.
The fourth habitat guild (Guild III) grouped six small-bodied taxa into a mixed habitat guild. Both M. mogurnda and C. stercusmuscarum, which had low associations with all habitat features in this study, have previously been shown to occur across varying substrate types 67 , while C. stercusmuscarum has been described as a habitat generalist 24 . Likewise Oxyeleotris sp. is described as a habitat generalist with a preference for abundant cover 29,46 , and juveniles of O. selheimi in this study were found to have a slight positive association with both root masses and undercut banks. Amniataba percoides was found to have a strong association with bedrock substrate in this study, but has been previously described as associated with a variety of substrates 67 . In comparison, L. triramus in this study was positively associated with both sandy and fine gravel substrates, this supports previous observations of the species being associated with sandy slackwater habitats 29 . Craterocephalus stramineus has previously been described as a riffle-dwelling species 22 , however we found that this species was associated with aquatic vegetation and filamentous algae. www.nature.com/scientificreports www.nature.com/scientificreports/ The habitat use differences observed in this study may be due to differences among rivers and over time, or potentially this study's coarser assessment of habitat use across a 5 minute shot length; i.e. fish may be captured in a rarer, smaller habitat type within the whole habitat sampling shot, and therefore specific associations may be  Table 4 76 ), presumably because the size of an individual influences its ability to respond to velocity, seek shelter from predation and use other resources 77 . In this study, three species (H. fuliginosus, S. butleri, and O. selhemi) demonstrated distinct habitat changes from juvenile to adult life stages. As previously discussed, juvenile H. fuliginosus and S. butleri utilised riffle habitats and then shifted to deeper, slow flowing pools as adults. Similar ontogenetic shifts in habitat use have been observed in other species including brown trout, Salmo trutta, where young-of-the-year trout inhabited fast-flowing riffles, while larger, older trout occupied deeper and slower flowing habitats 78 . In this study, juvenile O. selheimi was associated with habitats dominated by aquatic vegetation, undercut banks and root masses (mixed habitat guild, see above), before switching to habitats with increased levels of submerged wood as adults (deep, structured pools habitat guild, see above). Conversely, our findings showed no change in habitat use between juvenile and adult life stages for L. calcarifer or O. lineolata. Whilst the number of species demonstrating ontogenetic habitat changes could be viewed as low (only three out of five species where life stages were categorised), we did not sample larval or early juvenile life stages in this study, and further research on the habitat use of these important life stages in wet-dry tropical rivers is needed.
The patterns of habitat use revealed in this study improve our understanding of the habitat requirements of a range of fish taxa during the dry season in northern Australian rivers. The study quantified fish-habitat use associations for a large number of common taxa, and identified four distinct habitat use guilds. The fish-habitat guild approach is useful in summarising and simplifying communication of fish habitat use requirements to managers and the broader community. After further exploration of the proposed fish-habitat guilds, the approach may also be useful for predicting habitat requirements in similar systems with little ecological knowledge.
While a range of habitat variables were important habitat descriptors for individual taxa and size classes, unsurprisingly water depth and velocity were key factors influencing dry season habitat use of fish in this system. Many studies have emphasised the maintenance of natural seasonal flow patterns for the persistence of freshwater fish assemblages and the diversity of riverine mesohabitats 4,58,79,80 . In perennial wet-dry river systems, www.nature.com/scientificreports www.nature.com/scientificreports/ the prolonged low water period during the dry season (~May-October), maintains critical riverine longitudinal connectivity and a range of shallow and deeper habitats that sustain aquatic ecosystems (this study 31,63,81 ). Shallow habitats such as riffles are likely to be particularly vulnerable to natural or artificial reductions in water levels 18,31 , and hence changes in flow or dewatering of these habitats during the dry season could reduce connectivity, constrain fish movement and have particularly strong impacts on specific riffle dwellers such as L. unicolor, Glossogobius spp., and juvenile H. fuliginosus and S. butleri (this study 18 ). These results, coupled with data on the extent of riffle habitat area at various water levels, would enable predictions to be made about the extent of riffle habitat at different water levels, and hence predict the impact on riffle-dwelling species. Incorporation of this type of information into water planning is critical for future decision making regarding the management of environmental water in these systems.