TY - JOUR
AU - Kragt, Marit, E
AB - Abstract Marine recreational fisheries (MRFs) are often highly spatially heterogenous, with effort concentrated into small areas, and fisheries spanning large environmental gradients. However, spatially resolved catch data is rarely collected in MRFs, preventing the study of spatial heterogeneity in catch. This study uses recreational catch reported in 10 × 10 nm blocks across eight degrees of latitude in Western Australia to map spatial predictions of the probability of a recreational catch on an average trip for two key species: West Australian dhufish (Glaucosoma hebraicum) and snapper (Chrysophrys auratus). Two spatial modelling techniques are compared for the analysis, generalized additive mixed models (GAMMs) and boosted regression trees (BRTs). We find that BRTs outperform GAMMs, but performance gains are small. We also find marked spatial variations in recreational catch probabilities: high catches of dhufish are found in the north of the study area, and low catches in the Perth Metropolitan area and in the south; snapper catches are highest in the north and low in the south. These patterns are used to identify important spatial processes in the fishery. The analysis also suggests that modelling approach (GAMMs or BRTs) has only a minor effect on outcomes of spatial catch analysis in MRFs. Introduction Recreational fishing is a popular activity across the world, with an estimated 10% of populations in developed countries participating (Arlinghaus et al., 2015). In light of this popularity, the social and economic importance of recreational fishing is increasingly being recognized (Henry and Lyle, 2003; Hunt et al., 2013; Griffiths et al., 2017), as is the potential for recreational fisheries to impact ecological systems (Post et al., 2002; Coleman et al., 2004). Recent research on recreational fisheries has emphasized the need to integrate human-dimension and biological research, and reframe recreational fisheries as coupled social–ecological systems (Fenichel et al., 2013; Hunt et al., 2013; Elmer et al., 2017). Marine recreational fisheries (MRFs) typically operate over large geographic areas and are highly spatially heterogenous. In many MRFs fishing effort is concentrated near population centres and access points (Ryan et al., 2017; Mitchell et al., 2018). Similarly, environmental gradients across fisheries affect species distributions, reproduction, growth rates, and feeding behaviour (Austin, 2002; Stoner, 2004; Wakefield et al., 2015). These spatial processes can affect stock sustainability, fisher behaviour, and fisher satisfaction, and are therefore critical to managing MRFs as social–ecological systems (Hunt et al., 2013). However, spatially resolved catch and effort data that could be used to study spatial trends in recreational fisheries are rarely collected. In commercial fisheries where spatially resolved catch data are available, spatial trends in catch have been studied (Bigelow et al., 1999; Yu et al., 2013; Lan et al., 2015; Hahlbeck et al., 2017). For example, Hahlbeck et al. (2017) map expected by-catch rates of sunfish and bluefin tuna in a Californian gillnet fishery, using generalized additive models (GAMs) and spatial environmental predictors including sea-surface temperature, rugosity, and longitude. Whilst a range of spatial modelling approaches are available, GAMs and boosted regression trees (BRTs) are often used to model spatial catch trends in commercial fisheries (Bigelow et al., 1999; Yu et al., 2013; Lan et al., 2015; Hahlbeck et al., 2017). GAMs and BRTs are particularly well suited to spatial analysis of patchy fishery-dependent data as environmental variables aid predictions in areas with few observations (Yu et al., 2013). The boat-based recreational fishery in Western Australia provides a good case study to model spatial trends in catch for an MRF as suitable data exists in 10 × 10 nm blocks across eight latitudinal degrees from three 12-month survey years. In this study, we used regression based generalized additive mixed models (GAMMs) and machine learning based BRT models to predict the probability of recreational catch on an average trip across the fishery for two key species: West Australian dhufish (Glaucosoma hebraicum; Richardson, 1845) and snapper (Chrysophrys auratus; Forster, 1801). As few studies have modelled spatial catch in MRFs, we compare GAMMs and BRTs, examining differences in predictive performance and predictor effects. Martínez-Rincón et al. (2012) compare these two approaches in a commercial purse-seine fishery, finding BRTs to perform marginally better. However, GAMMs and BRTs treat noise and variable interactions differently, meaning their performance is likely context dependent. This study provides insights about important spatial processes for each species and is the first to compare GAMMs and BRTs for spatial catch analysis in a recreational fishing context. Methods Study area and species The study area spans ∼1500 km of coastline across two broad fisheries management areas in Western Australia: the West Coast and South Coast Bioregions (Figure 1). This study focuses on two key recreational species from the study region: West Australian dhufish (hereon in referred to as dhufish) and snapper (Fairclough et al., 2014; Ryan et al., 2017). These species are contrasting in terms of biogeographic range: dhufish is a range-restricted endemic species which occurs from Shark Bay (113°E 26°S) to the Recherche Archipelago (122°E 34°S), whilst snapper is cosmopolitan and broad ranging, including across the West Coast and South Coast Bioregions (Gomon et al., 2008). The two species are targeted by recreational fishers as well as commercial and charter fishing vessels. In 2007, stock assessment revealed dhufish and snapper were being overfished in the West Coast Bioregion leading to management reform in 2009 (Wise et al., 2007; Crowe et al., 2013). Currently, recreational catch is managed through a combination of bag and size limits which are set at the bioregional level, and an annual 2-month temporal closure in the West Coast Bioregion. Spawning aggregations of snapper adjacent to the Perth Metropolitan area are also protected by a spatio-temporal closure. Figure 1. View largeDownload slide Study area showing the West Coast and South Coast Bioregions in Western Australia, separated by the dashed line, and 10 × 10 nm blocks that defined fishing locations reported by boat-based recreational fishers. Figure 1. View largeDownload slide Study area showing the West Coast and South Coast Bioregions in Western Australia, separated by the dashed line, and 10 × 10 nm blocks that defined fishing locations reported by boat-based recreational fishers. Catch data Recreational catch of dhufish and snapper was collected in phone diary surveys conducted in 2011/2012, 2013/2014, and 2015/2016 (Ryan et al., 2013,, 2015,, 2017). Separate panels of respondents were randomly selected for each survey from a database of fishers with a Recreational Boat Fishing Licence (RBFL) using a probability-based survey design with proportional sampling from residential strata and similar selection probabilities amongst strata. The sampling frame included all licence holders above 5 years of age that had a RBFL in the 12 months prior to the survey and intended to fish in the survey period. Each month for the 12-month survey period participants were called by trained interviewers and details of each boat-based fishing trip in the month recorded. Fishers reported the spatial location of their catch using a 10 × 10 nm grid (Figure 1). This study considers a subset of the survey data that included fishing trips in the study area (Figure 1) that were launched at a public boat ramp, fished in marine or estuarine waters, and used line fishing gear (including bait and lures). This data subset included 10 919 fishing trips by 3062 fishers over the three survey years. Almost half (48%) of all fishing trips were adjacent to the state capital and major population centre of Perth (Figure 1). Catches of dhufish and snapper were modelled as binary (presence/absence) response variables describing the probability of a catch, based on whether either species was caught and retained on a given fishing trip (hereon in referred to as catch). We chose to model retained catch as both species have high release rates (dhufish: 68%, snapper: 76%) predominantly due to fish being below their legal minimum lengths (dhufish: 78%, snapper: 76%) (Ryan et al., 2017). Therefore, retained catch was thought to be a better reflection of success in the fishery. Using a binary variable removed the need to account for bag limits which truncate the catch distribution and reflects that catch of more than one dhufish and snapper was uncommon due to the low bag and boat limits for the two species. Fishing operation and environmental data Fishing operation variables were recorded in the surveys of recreational catch and included: fisher hours, gear-type, targeting, fisher age, and fisher avidity (Table 1). Fisher hours was calculated as the product of the number of hours spent fishing and the number of licenced fishers in the party. Whilst individual target species were recorded (i.e. dhufish and snapper), we defined targeting using the demersal suite of species as fishers often report their most frequently caught species as their target (Maunder and Punt, 2004). Fishers were considered to target dhufish and snapper if they reported targeting any demersal species, which includes dhufish and snapper, and also other species commonly associated with inshore demersal habitats. Fisher avidity was self-reported as the number of fishing trips in the last 12 months. Table 1. Response and predictor variables and sample averages (n = 10 919) used to model trip-level catch of dhufish and snapper from a boat-based fishery in south-west Australia. Variable Description Units Mean (range) Response variables  Dhufish catch Presence (1) or absence (0): 1 (17%); 0 (83%)  Snapper catch Presence (1) or absence (0): 1 (11%); 0 (89%) Predictor variables  Fishing operation variables   Targeting Targeting (1) or not-targeting (0): 1 (42%); 0 (58%)   Fisher hours Number of licenced fishers on the fishing trip multiplied by number of hours fished Hours 7.9 (0.0–65)   Gear type Fishing method used: bait (80%); lure (8%); both (12%)   Fisher age Age of fisher Years 47.9 (5–91)   Avidity Stated number of fishing trips reported by fisher for the 12 months prior to the survey: 0–9 (23%); 10–19 (27%); 20+ (50%)  Environmental variables   Temperature Sea surface temperature °C 21.0 (15.3–28.2)   Swell Interpolated swell height (m) Metres 1.7 (0.2–6.4)   Depth Average depth across the 10 × 10 nm block the fisher reported to have fished Metres 15.8 (0.6–706.0)   Rugosity Average rugosity index measured across the 10 × 10 nm block 1.4 (0.0–7.1)  Temporal variables   Time of day Time of day fishing started rounded to nearest hour   Month Month of fishing trip   Year Survey year: 2011/2012 (38%); 2013/2014 (33%); 2015/2016 (29%)  Spatial variables   Longitude, latitude Latitude and longitude of the centroid of the 10 × 10 nm block the fisher reported to have fished Decimal degrees Variable Description Units Mean (range) Response variables  Dhufish catch Presence (1) or absence (0): 1 (17%); 0 (83%)  Snapper catch Presence (1) or absence (0): 1 (11%); 0 (89%) Predictor variables  Fishing operation variables   Targeting Targeting (1) or not-targeting (0): 1 (42%); 0 (58%)   Fisher hours Number of licenced fishers on the fishing trip multiplied by number of hours fished Hours 7.9 (0.0–65)   Gear type Fishing method used: bait (80%); lure (8%); both (12%)   Fisher age Age of fisher Years 47.9 (5–91)   Avidity Stated number of fishing trips reported by fisher for the 12 months prior to the survey: 0–9 (23%); 10–19 (27%); 20+ (50%)  Environmental variables   Temperature Sea surface temperature °C 21.0 (15.3–28.2)   Swell Interpolated swell height (m) Metres 1.7 (0.2–6.4)   Depth Average depth across the 10 × 10 nm block the fisher reported to have fished Metres 15.8 (0.6–706.0)   Rugosity Average rugosity index measured across the 10 × 10 nm block 1.4 (0.0–7.1)  Temporal variables   Time of day Time of day fishing started rounded to nearest hour   Month Month of fishing trip   Year Survey year: 2011/2012 (38%); 2013/2014 (33%); 2015/2016 (29%)  Spatial variables   Longitude, latitude Latitude and longitude of the centroid of the 10 × 10 nm block the fisher reported to have fished Decimal degrees View Large Table 1. Response and predictor variables and sample averages (n = 10 919) used to model trip-level catch of dhufish and snapper from a boat-based fishery in south-west Australia. Variable Description Units Mean (range) Response variables  Dhufish catch Presence (1) or absence (0): 1 (17%); 0 (83%)  Snapper catch Presence (1) or absence (0): 1 (11%); 0 (89%) Predictor variables  Fishing operation variables   Targeting Targeting (1) or not-targeting (0): 1 (42%); 0 (58%)   Fisher hours Number of licenced fishers on the fishing trip multiplied by number of hours fished Hours 7.9 (0.0–65)   Gear type Fishing method used: bait (80%); lure (8%); both (12%)   Fisher age Age of fisher Years 47.9 (5–91)   Avidity Stated number of fishing trips reported by fisher for the 12 months prior to the survey: 0–9 (23%); 10–19 (27%); 20+ (50%)  Environmental variables   Temperature Sea surface temperature °C 21.0 (15.3–28.2)   Swell Interpolated swell height (m) Metres 1.7 (0.2–6.4)   Depth Average depth across the 10 × 10 nm block the fisher reported to have fished Metres 15.8 (0.6–706.0)   Rugosity Average rugosity index measured across the 10 × 10 nm block 1.4 (0.0–7.1)  Temporal variables   Time of day Time of day fishing started rounded to nearest hour   Month Month of fishing trip   Year Survey year: 2011/2012 (38%); 2013/2014 (33%); 2015/2016 (29%)  Spatial variables   Longitude, latitude Latitude and longitude of the centroid of the 10 × 10 nm block the fisher reported to have fished Decimal degrees Variable Description Units Mean (range) Response variables  Dhufish catch Presence (1) or absence (0): 1 (17%); 0 (83%)  Snapper catch Presence (1) or absence (0): 1 (11%); 0 (89%) Predictor variables  Fishing operation variables   Targeting Targeting (1) or not-targeting (0): 1 (42%); 0 (58%)   Fisher hours Number of licenced fishers on the fishing trip multiplied by number of hours fished Hours 7.9 (0.0–65)   Gear type Fishing method used: bait (80%); lure (8%); both (12%)   Fisher age Age of fisher Years 47.9 (5–91)   Avidity Stated number of fishing trips reported by fisher for the 12 months prior to the survey: 0–9 (23%); 10–19 (27%); 20+ (50%)  Environmental variables   Temperature Sea surface temperature °C 21.0 (15.3–28.2)   Swell Interpolated swell height (m) Metres 1.7 (0.2–6.4)   Depth Average depth across the 10 × 10 nm block the fisher reported to have fished Metres 15.8 (0.6–706.0)   Rugosity Average rugosity index measured across the 10 × 10 nm block 1.4 (0.0–7.1)  Temporal variables   Time of day Time of day fishing started rounded to nearest hour   Month Month of fishing trip   Year Survey year: 2011/2012 (38%); 2013/2014 (33%); 2015/2016 (29%)  Spatial variables   Longitude, latitude Latitude and longitude of the centroid of the 10 × 10 nm block the fisher reported to have fished Decimal degrees View Large Environmental variables were determined from the 10 × 10 nm grid location reported for each fishing trip (Figure 1). Environmental variables included: temperature approximated from grid centroid according to daily mean sea surface temperature satellite records in 0.25 × 0.25° grids from the US National Oceanic and Atmospheric Administration (Reynolds et al., 2007); inverse distance weighted average swell (within a 3-h window from the trip start time) from four insitu wave-rider buoys (Department of Transport, Western Australia); average depth according to standard Australian bathymetry (Whiteway, 2009); and rugosity from standard Australian topography (Whiteway, 2009) using the Benthic Terrain Modeler in the ArcGIS software applying a 900 ×900 m analysis window (Walbridge et al., 2018) (Table 1). Statistical methods Catches of dhufish and snapper were modelled as a function of environmental (temperature, swell, depth, and rugosity); fishing operation (targeting, fisher hours, avidity, fisher age, and gear type); temporal (time of day, month, and year); and spatial (longitude and latitude) variables using two approaches (GAMMs and BRTs). GAMMs are an extension of generalized linear models (GLMs) that replace the linear parameter effects with an additive smooth function. The main advantage over GLMs is that GAMMs can accommodate complex non-linearities in predictor effects (Hastie and Tibshirani, 1986). Our GAMMs also include random variables to account for multiple trips made by the same fisher. The second modelling approach, BRTs differ from GAMMs and GLMs in that rather than fitting a single model, BRTs fit many models in the form of regression trees, that are combined to describe the predictors-response variable relationship (Friedman et al., 2000; Elith et al., 2008). Each regression tree divides observations of the response variable into groups with similar values and describes group membership using a set of binary rules based on the predictors. Regression trees are fit successively to the residuals of previous trees, explaining the variance that previous trees failed to capture. Additional trees are added until a priori stopping criteria is achieved to avoid overfitting. In a practical sense BRTs are like GAMMs in that complex non-linear predictor effects are captured in model fitting. However, unlike GAMMs, predictor interactions in BRTs are also captured in model fitting, rather than having to be specified in the model. Comparisons of the GAMMs and BRTs followed methods used by Leathwick et al. (2006). For each species, two GAMMs were fit: one with no predictor interactions, and one allowing interactions. Interactions in BRTs are controlled by a tree-complexity parameter that specifies the order of interactions (e.g. one implies no-interactions, two implies pairwise-interactions, and so on). To match the GAMMs, we fit a no-interaction and a pairwise-interaction BRT by setting the tree-complexity to one and two, respectively. Higher order interactions were also tested by fitting a BRT with tree-complexity of five. GAMMs were fit using the “gamm4” package in the R statistical software (Wood and Scheipl, 2014). Log and square root transformations were applied to continuous predictors as needed to increase their uniformity. Model selection for each GAMM was based on forward stepwise selection using the lowest Akaike’s information criterion (AIC) score for the converging models. In some instances, convergence problems arose that appeared related to random effects, due to using data from many fishers but with few replicates per fisher. Despite this, the dhufish and snapper GAMMs included six and seven predictors, respectively, which is at the upper end of the number of predictors typically included in GAMMs in the literature (Fisher et al., 2018). The BRTs were fit using the “dismo” package in the R statistical software (Hijmans et al., 2017). Optimal settings for the learning-rate (controlling weighting of previous vs. subsequent regression trees) and bag-fraction (controlling stochasticity) were explored, with a learning-rate of 0.01 and a bag-fraction of 0.7 providing the best all round performance for all BRTs. We also tested whether removing correlated predictors improved predictive performance using the “caret” package in the R statistical software (Kuhn, 2019); no performance gains were found and so all predictors are retained in the BRT models. A key output of BRTs is the relative importance of each variable, which is scaled to sum to 100% across variables, and reflects the average over all trees of the number of times the variable is used in a regression tree, weighted by the resulting squared improvement in model fit. Tenfold cross-validation was used to determine the optimal number of trees in each BRT model and subsequently to compare the predictive performance of the BRTs and GAMMs. For each fold, a model was fit to 90% of the data and predictive performance assessed against the hold-out 10% by calculating fit metrics against the model predicted values. This process was repeated ten times without replacement of the hold-out data set (i.e. rotating through the data set), and overall model performance was determined by averaging fit metrics across the tenfold (Leathwick et al., 2006). The optimal number of trees in each BRT was determined by adding trees until cross-validated residual deviance no longer improved. Having identified optimized GAMMs (by lowest AIC) and BRTs (by cross-validation), the predictive performance of the models were compared using cross-validated residual deviance, deviance explained relative to a null model, area under the receiver curve (AUC), and per cent correctly classified (PCC; the PCC threshold was set at each species prevalence) (Leathwick et al., 2006). AUC is a measure of the ability of the model to discriminate between catch and no-catch events, with higher values indicating better discriminatory ability. To explore spatial variability in recreational catch of dhufish and snapper we made predictions of the probability of a catch for each species on an average trip to each grid. The spatial predictions reflect the effects of latitude, longitude, temperature, depth, and rugosity on catch. Mean annual temperatures for each grid-cell were used. Predictions were based on a trip targeting demersal species. All other non-spatial predictors were kept at their means for normally distributed continuous variables (fisher age, swell), median for skewed continuous variables (time of day, fisher hours), and modes for discrete variables (year, month, method, avidity). Predictions were limited to grids with a mean depth of <200 m. For each species, the model with the highest cross-validated performance was used to generate the catch predictions. Confidence intervals of the predictions were estimated by bootstrapping and are asymmetric as they apply to probabilities which are bounded by one and zero. Colour ramps are from the “viridis” package in the R statistical software (Garnier, 2018). Results Comparison of model predictive performance Predictor variables associated with catch of dhufish and snapper varied between species and to a lesser extent modelling approach (Table 2). The BRTs consistently performed marginally better in the cross-validation tests than the GAMMs (Table 3). For dhufish, when no interactions were included (complexity = 1), the BRT explained 2.9% more deviance than the GAMM, and had a higher AUC score indicating better discriminatory ability. For models with pairwise interactions (complexity = 2), the BRT had a higher AUC score and explained 1.8% more deviance than the GAMM. Overall, for dhufish the best performing GAMM included pairwise interactions [Equation (1)] but was outperformed by the best performing BRT which allowed up to fifth-order interactions [Equation (2)] and explained 2.1% more deviance, had a slightly higher AUC score, and correctly classified 0.2% more observations. GAMM:  Dhufish catch= f(Targeting, Fisher hours, Depth, Latitude,  YearTime of day, Fisher hours×Targeting), (1) BRT:  Dhufish catch=f(All predictors with up to fifth-order interactions). (2) Table 2. Comparison of outcomes for predictor variables used to explain variability in recreational catch using GAMMs and BRTs. Variable GAMM complexity BRT complexity 1 2 1 2 5 Dhufish Targeting X X 46.5 44.3 38.7 Longitude 20.6 19.9 18.4 Latitude X X 14.4 11.0 8.6 Fisher hours X X 7.5 7.4 7.9 Month 1.4 3.3 5.5 Year X X 2.6 2.6 2.8 Time of day X X 2.1 2.6 3.0 Temperature 1.1 2.1 3.4 Swell 1.0 1.8 3.4 Depth X X 1.0 1.7 3.0 Fisher age 0.6 1.4 2.7 Rugosity 0.4 0.8 1.3 Gear type 0.6 0.6 0.6 Avidity 0.3 0.3 0.7 Targeting × fisher hours X Snapper Targeting X X 47.0 38.9 32.9 Fisher hours X X 16.8 14.4 13.3 Temperature X X 8.1 9.3 9.0 Month X X 4.5 8.8 11.0 Longitude X X 8.0 8.0 7.1 Time of day X X 5.8 6.3 7.1 Latitude 2.7 3.4 4.0 Swell X X 1.7 3.2 4.2 Depth 2.5 2.6 3.2 Fisher age 1.7 2.4 4.0 Rugosity 1.1 1.5 2.2 Year 0.0 0.6 0.9 Avidity 0.0 0.6 0.8 Gear type 0.1 0.2 0.2 Targeting × fisher hours X Variable GAMM complexity BRT complexity 1 2 1 2 5 Dhufish Targeting X X 46.5 44.3 38.7 Longitude 20.6 19.9 18.4 Latitude X X 14.4 11.0 8.6 Fisher hours X X 7.5 7.4 7.9 Month 1.4 3.3 5.5 Year X X 2.6 2.6 2.8 Time of day X X 2.1 2.6 3.0 Temperature 1.1 2.1 3.4 Swell 1.0 1.8 3.4 Depth X X 1.0 1.7 3.0 Fisher age 0.6 1.4 2.7 Rugosity 0.4 0.8 1.3 Gear type 0.6 0.6 0.6 Avidity 0.3 0.3 0.7 Targeting × fisher hours X Snapper Targeting X X 47.0 38.9 32.9 Fisher hours X X 16.8 14.4 13.3 Temperature X X 8.1 9.3 9.0 Month X X 4.5 8.8 11.0 Longitude X X 8.0 8.0 7.1 Time of day X X 5.8 6.3 7.1 Latitude 2.7 3.4 4.0 Swell X X 1.7 3.2 4.2 Depth 2.5 2.6 3.2 Fisher age 1.7 2.4 4.0 Rugosity 1.1 1.5 2.2 Year 0.0 0.6 0.9 Avidity 0.0 0.6 0.8 Gear type 0.1 0.2 0.2 Targeting × fisher hours X For the GAMMs and BRTs the model complexity indicates the degree of interactions in each model: 1 indicates no interactions, 2 indicates up-to pairwise interactions, and 5 indicates up-to fifth-order interactions. For the GAMMs, the X’s show variables included in the lowest AIC models. Importance scores of variables are shown for the BRTs. Variables are ordered based on average importance scores across the BRT models. View Large Table 2. Comparison of outcomes for predictor variables used to explain variability in recreational catch using GAMMs and BRTs. Variable GAMM complexity BRT complexity 1 2 1 2 5 Dhufish Targeting X X 46.5 44.3 38.7 Longitude 20.6 19.9 18.4 Latitude X X 14.4 11.0 8.6 Fisher hours X X 7.5 7.4 7.9 Month 1.4 3.3 5.5 Year X X 2.6 2.6 2.8 Time of day X X 2.1 2.6 3.0 Temperature 1.1 2.1 3.4 Swell 1.0 1.8 3.4 Depth X X 1.0 1.7 3.0 Fisher age 0.6 1.4 2.7 Rugosity 0.4 0.8 1.3 Gear type 0.6 0.6 0.6 Avidity 0.3 0.3 0.7 Targeting × fisher hours X Snapper Targeting X X 47.0 38.9 32.9 Fisher hours X X 16.8 14.4 13.3 Temperature X X 8.1 9.3 9.0 Month X X 4.5 8.8 11.0 Longitude X X 8.0 8.0 7.1 Time of day X X 5.8 6.3 7.1 Latitude 2.7 3.4 4.0 Swell X X 1.7 3.2 4.2 Depth 2.5 2.6 3.2 Fisher age 1.7 2.4 4.0 Rugosity 1.1 1.5 2.2 Year 0.0 0.6 0.9 Avidity 0.0 0.6 0.8 Gear type 0.1 0.2 0.2 Targeting × fisher hours X Variable GAMM complexity BRT complexity 1 2 1 2 5 Dhufish Targeting X X 46.5 44.3 38.7 Longitude 20.6 19.9 18.4 Latitude X X 14.4 11.0 8.6 Fisher hours X X 7.5 7.4 7.9 Month 1.4 3.3 5.5 Year X X 2.6 2.6 2.8 Time of day X X 2.1 2.6 3.0 Temperature 1.1 2.1 3.4 Swell 1.0 1.8 3.4 Depth X X 1.0 1.7 3.0 Fisher age 0.6 1.4 2.7 Rugosity 0.4 0.8 1.3 Gear type 0.6 0.6 0.6 Avidity 0.3 0.3 0.7 Targeting × fisher hours X Snapper Targeting X X 47.0 38.9 32.9 Fisher hours X X 16.8 14.4 13.3 Temperature X X 8.1 9.3 9.0 Month X X 4.5 8.8 11.0 Longitude X X 8.0 8.0 7.1 Time of day X X 5.8 6.3 7.1 Latitude 2.7 3.4 4.0 Swell X X 1.7 3.2 4.2 Depth 2.5 2.6 3.2 Fisher age 1.7 2.4 4.0 Rugosity 1.1 1.5 2.2 Year 0.0 0.6 0.9 Avidity 0.0 0.6 0.8 Gear type 0.1 0.2 0.2 Targeting × fisher hours X For the GAMMs and BRTs the model complexity indicates the degree of interactions in each model: 1 indicates no interactions, 2 indicates up-to pairwise interactions, and 5 indicates up-to fifth-order interactions. For the GAMMs, the X’s show variables included in the lowest AIC models. Importance scores of variables are shown for the BRTs. Variables are ordered based on average importance scores across the BRT models. View Large Table 3. Comparison of predictive performance of GAMMs and BRTs of varying complexity (i.e. degree of predictor interactions) for analysis of recreational fishing catch. Species Method Complexity No. of trees Cross-validated AUC (s.e.) Per cent correctly classified (s.e.) Cross-validated residual deviance (s.e.) Deviance explained D2 Dhufish GAMM 1 – 0.854 (0.010) 0.770 (0.004) 0.711 (0.030) 0.262 BRT 1 2 500 0.862 (0.009) 0.772 (0.006) 0.683 (0.027) 0.291 GAMM 2 – 0.855 (0.010) 0.777 (0.004) 0.702 (0.029) 0.271 BRT 2 2 100 0.864 (0.009) 0.777 (0.006) 0.686 (0.027) 0.289 BRT 5 1 450 0.865 (0.009) 0.779 (0.005) 0.683 (0.026) 0.292 Snapper GAMM 1 – 0.801 (0.007) 0.751 (0.005) 0.615 (0.019) 0.153 BRT 1 2 300 0.801 (0.008) 0.751 (0.008) 0.602 (0.019) 0.170 GAMM 2 – 0.802 (0.007) 0.733 (0.005) 0.609 (0.019) 0.161 BRT 2 1 450 0.801 (0.008) 0.692 (0.007) 0.604 (0.019) 0.168 BRT 5 900 0.802 (0.008) 0.688 (0.008) 0.605 (0.019) 0.167 Species Method Complexity No. of trees Cross-validated AUC (s.e.) Per cent correctly classified (s.e.) Cross-validated residual deviance (s.e.) Deviance explained D2 Dhufish GAMM 1 – 0.854 (0.010) 0.770 (0.004) 0.711 (0.030) 0.262 BRT 1 2 500 0.862 (0.009) 0.772 (0.006) 0.683 (0.027) 0.291 GAMM 2 – 0.855 (0.010) 0.777 (0.004) 0.702 (0.029) 0.271 BRT 2 2 100 0.864 (0.009) 0.777 (0.006) 0.686 (0.027) 0.289 BRT 5 1 450 0.865 (0.009) 0.779 (0.005) 0.683 (0.026) 0.292 Snapper GAMM 1 – 0.801 (0.007) 0.751 (0.005) 0.615 (0.019) 0.153 BRT 1 2 300 0.801 (0.008) 0.751 (0.008) 0.602 (0.019) 0.170 GAMM 2 – 0.802 (0.007) 0.733 (0.005) 0.609 (0.019) 0.161 BRT 2 1 450 0.801 (0.008) 0.692 (0.007) 0.604 (0.019) 0.168 BRT 5 900 0.802 (0.008) 0.688 (0.008) 0.605 (0.019) 0.167 Null deviance is 0.964 for dhufish and 0.726 for snapper. The best performing model is shown in bold. View Large Table 3. Comparison of predictive performance of GAMMs and BRTs of varying complexity (i.e. degree of predictor interactions) for analysis of recreational fishing catch. Species Method Complexity No. of trees Cross-validated AUC (s.e.) Per cent correctly classified (s.e.) Cross-validated residual deviance (s.e.) Deviance explained D2 Dhufish GAMM 1 – 0.854 (0.010) 0.770 (0.004) 0.711 (0.030) 0.262 BRT 1 2 500 0.862 (0.009) 0.772 (0.006) 0.683 (0.027) 0.291 GAMM 2 – 0.855 (0.010) 0.777 (0.004) 0.702 (0.029) 0.271 BRT 2 2 100 0.864 (0.009) 0.777 (0.006) 0.686 (0.027) 0.289 BRT 5 1 450 0.865 (0.009) 0.779 (0.005) 0.683 (0.026) 0.292 Snapper GAMM 1 – 0.801 (0.007) 0.751 (0.005) 0.615 (0.019) 0.153 BRT 1 2 300 0.801 (0.008) 0.751 (0.008) 0.602 (0.019) 0.170 GAMM 2 – 0.802 (0.007) 0.733 (0.005) 0.609 (0.019) 0.161 BRT 2 1 450 0.801 (0.008) 0.692 (0.007) 0.604 (0.019) 0.168 BRT 5 900 0.802 (0.008) 0.688 (0.008) 0.605 (0.019) 0.167 Species Method Complexity No. of trees Cross-validated AUC (s.e.) Per cent correctly classified (s.e.) Cross-validated residual deviance (s.e.) Deviance explained D2 Dhufish GAMM 1 – 0.854 (0.010) 0.770 (0.004) 0.711 (0.030) 0.262 BRT 1 2 500 0.862 (0.009) 0.772 (0.006) 0.683 (0.027) 0.291 GAMM 2 – 0.855 (0.010) 0.777 (0.004) 0.702 (0.029) 0.271 BRT 2 2 100 0.864 (0.009) 0.777 (0.006) 0.686 (0.027) 0.289 BRT 5 1 450 0.865 (0.009) 0.779 (0.005) 0.683 (0.026) 0.292 Snapper GAMM 1 – 0.801 (0.007) 0.751 (0.005) 0.615 (0.019) 0.153 BRT 1 2 300 0.801 (0.008) 0.751 (0.008) 0.602 (0.019) 0.170 GAMM 2 – 0.802 (0.007) 0.733 (0.005) 0.609 (0.019) 0.161 BRT 2 1 450 0.801 (0.008) 0.692 (0.007) 0.604 (0.019) 0.168 BRT 5 900 0.802 (0.008) 0.688 (0.008) 0.605 (0.019) 0.167 Null deviance is 0.964 for dhufish and 0.726 for snapper. The best performing model is shown in bold. View Large For snapper, comparing the no interaction models (complexity = 1), the BRT explained 1.7% more deviance, but both models had identical AUC score. For pairwise interaction models (complexity = 2), the BRT marginally outperformed the GAMM explaining 0.7% more deviance and with virtually identical AUC scores. Overall, for snapper the best performing GAMM had no interactions [Equation (3)] but was outperformed by a BRT with no interactions [Equation (4)] explaining 1.7% more deviance but with identical AUC score and PCC to the GAMM. GAMM:  Snapper catch= f(Targeting, Fisher hours, Temperature,  SwellLongitude, Month, Time of day), (3) BRT:  Snapper catch= f(All predictors with no interactions). (4) Predictor effects on catch Comparing the GAMM and BRT with the highest predictive performance for each species shows some discrepancies in predictor effects for dhufish [Equation (1) vs. Equation (2); Figure 2], but broadly similar predictor effects for snapper [Equation (3) vs. Equation (4); Figure 3]. The fishing operation variables targeting, and fisher hours were important in explaining dhufish catch in both models (Figure 2). Depth was the only important environmental predictor of dhufish catch and was only important in the GAMM. Both models show that fishers targeting demersal species were more likely to catch dhufish than those that were not. Dhufish catch also increased with the number of fisher hours, but the effect plateaued in the BRT above 23 fisher hours. The GAMM showed dhufish catch increased with depth, particularly between 0 and 50 m. Figure 2. View largeDownload slide Variations in probability of catching a dhufish using BRT (left) and GAMM (right). Shaded area shows 95% confidence intervals. Confidence intervals were calculated for BRT by bootstrapping (n = 1000). For the BRT, only variables with an importance score of >5% are shown. “n.s.” indicates predictor variable was not significant. Figure 2. View largeDownload slide Variations in probability of catching a dhufish using BRT (left) and GAMM (right). Shaded area shows 95% confidence intervals. Confidence intervals were calculated for BRT by bootstrapping (n = 1000). For the BRT, only variables with an importance score of >5% are shown. “n.s.” indicates predictor variable was not significant. Figure 3. View largeDownload slide Variations in probability of catching a snapper using BRT (left) and GAMM (right). Shaded area shows 95% confidence intervals. Confidence intervals were calculated for the BRT by bootstrapping (n = 1000). For the BRT, only variables with an importance score of >5% are shown. “n.s.” indicates predictor variable was not significant. Figure 3. View largeDownload slide Variations in probability of catching a snapper using BRT (left) and GAMM (right). Shaded area shows 95% confidence intervals. Confidence intervals were calculated for the BRT by bootstrapping (n = 1000). For the BRT, only variables with an importance score of >5% are shown. “n.s.” indicates predictor variable was not significant. Important temporal and spatial predictors of dhufish catch included year and time of day in the GAMM, month in the BRT, latitude in both models, and longitude in the BRT. The GAMM showed that catch of dhufish increased across sample years, and was highest in the early morning hours. Similar year and time-of-day effects were found in the BRT model (Supplementary material) but were not included in Figure 2 due to low importance scores (<5%). Whilst important, month in the BRT showed no clear pattern in dhufish catch. The implied geographic trends of latitude in the GAMMs and latitude and longitude in the BRT were similar showing a (geographically speaking) bi-modal distribution in dhufish catch, with peaks separated by the Perth Metropolitan area (115°E 32°S). In the BRT, a northern peak was observed at low longitudes, dropping east of 115°E, whilst a southern peak was observed at high latitudes, dropping north of 32°S. The same trend is shown in the GAMM using just latitude, with one peak at 30°S and another at 34°S with a low again centreing around the Perth Metropolitan area at 32°S. For snapper catch, important fishing operation and environmental variables included targeting, fisher hours, and temperature in both models, and swell in the GAMMs. Targeting increased snapper catch in both models. Fisher hours also increased snapper catch in both models, though—as for dhufish—the effect plateaued in the BRT, at 36 fisher hours in this case. Both models show high snapper catch in low temperature conditions. A second peak in high temperatures was also suggested—particularly by the BRT—but there was a high degree of error associated with this peak. The GAMM found snapper catch to increase with swell size, but confidence intervals were large at high swells. A similar effect of swell was observed in the BRT (Supplementary material) but was of low importance (<5%). Important temporal and spatial predictors of snapper catch included time of day and longitude in both models and month in the GAMM. Time-of-day in both models showed early morning and late afternoon trips to be the most successful. The GAMM showed slightly higher snapper catch in the Austral winter months (Figure 3); a similar trend was observed in the BRT (Supplementary material) but was of low importance (<5%). Both models show a similar pattern with respect to longitude with high snapper catch at lower longitudes. Spatial predictions of catch Spatial predictions for dhufish were made using the overall best performing model; the BRT in Equation (2) (Figure 4). Dhufish catch probabilities ranged from 0.11 to 0.56, with estimated confidence intervals widths from 0.06 to 0.41. The highest catch occurred in the north of the study area, between 29 and 30°S, with catch probabilities between 0.43 and 0.56. Similarly, high catch probabilities were found north of 28.5°S, but predictions in this area were very uncertain (confidence interval range up to 0.41). Lowest catch occurred adjacent to the Perth Metropolitan area between 31.5 and 32.5°S and on the south coast (east of 115.5°E), both with catch probabilities <0.2. Figure 4. View largeDownload slide Dhufish recreational catch likelihood predictions and confidence interval range (bootstrapped 1000 times) from a BRT model using fishing operation, environmental, temporal and spatial predictors. Figure 4. View largeDownload slide Dhufish recreational catch likelihood predictions and confidence interval range (bootstrapped 1000 times) from a BRT model using fishing operation, environmental, temporal and spatial predictors. Spatial predictions for snapper were made using the BRT in Equation (4) (Figure 5). Snapper catch probabilities ranged from 0.11 to 0.73 with confidence interval widths ranging from 0.06 to 0.57. The predictions identified that snapper catch was generally higher in the north of the study area than the south. The highest catch probabilities were found north of 29°S with predictions of 0.62–0.73; however, some of this area was characterized by high uncertainty (confidence interval width up to 0.57). Medium catch probabilities (0.3–0.5) were found between 32.5 and 29°S, and in offshore locations between 34 and 32.5°S. Catch in the remainder of the study area was broadly similar between 0.11 and 0.3. Figure 5. View largeDownload slide Snapper recreational catch likelihood predictions and confidence interval range (bootstrapped 1000 times) from a BRT model using fishing operation, environmental, temporal and spatial predictors. Figure 5. View largeDownload slide Snapper recreational catch likelihood predictions and confidence interval range (bootstrapped 1000 times) from a BRT model using fishing operation, environmental, temporal and spatial predictors. Discussion Model comparison We found that BRTs performed slightly better in cross-validation predictive tests than GAMMs for spatial analysis of recreational fishing catch. This matches findings of a previous study that found BRTs to offer superior predictive performance over GAMs for modelling biodiversity (Leathwick et al., 2006). However, in our study the performance margin of BRTs was relatively low, particularly for snapper. This is probably because higher order interactions, whose inclusion is a major advantage of BRTs, were not useful in explaining the likelihood of catching snapper. Where interactions are not relevant or are not explored, other studies have also found only minor performance gains from BRTs (Martínez-Rincón et al., 2012). The predictor effects differed somewhat between BRTs and GAMMs for dhufish but were broadly similar for snapper. Discrepancies for dhufish are likely due to the ability of BRTs to capture high-order interactions important for explaining dhufish catch. For example, month was only important in the BRT with fifth-order interactions, suggesting dhufish catch depends on complex interactions with month that were not captured by the pairwise GAMM. Additional differences between the predictor effects are likely due to the BRTs ability to fit models with correlated predictors allowing inclusion of all 14 predictors in the BRTs, whereas only a subset of predictors could be used in the GAMM. Overall, our results suggest that whilst BRTs perform marginally better than GAMMs at predicting recreational catch, performance gains were small. In addition, predictor effects were similar across the approaches. As such, the analysis suggests that choice of modelling approach (GAMMs or BRTs) has only a minor effect on outcomes of spatial catch analysis in MRFs. A possible exception is when high-order interactions are important, in which case BRTs may perform better. Predictor effects on catch Whilst our focus was on predictive performance and spatial mapping, predictor effects for both species were consistent with other research findings. Time spent fishing (fisher hours) and whether the fisher targeted demersals (targeting) were both positively related to catching dhufish and snapper, as has been found in other recreational catch analysis (Pope et al., 2016). Both species also showed influence from environmental variables. It is important to note that environmental predictors in our analysis were correlated with latitude/longitude and month, and so their effects on catch may not be well represented. This reflects our focus on predictive performance, not modelling predictor effects. Nevertheless, the GAMM showed dhufish catch to increase with fishing depth, matching the species depth distribution (Hesp et al., 2002). High snapper catch was found in low temperatures and in high swell conditions. These trends may relate to increased targeting of snapper in winter, when temperatures are low and swells high (Crisafulli et al., 2019). In addition, it is widely thought amongst fishers that catchability of snapper is higher during or immediately after storm events (Crawford, 2016). The pattern of high early morning and evening catch for both species correlates with observations of crepuscular feeding patterns, which potentially enhance catchability (Helfman, 1986). High evening catch of snapper may also be the result of increased evening targeting of snapper that occurs in the Perth Metropolitan area in August and September leading up to snapper spawning (Crisafulli et al., 2019). The effect of year and month differed between species. Dhufish catch increased with survey year (2011/2012–2015/2016) in the GAMM. The most recent stock assessment for dhufish showed small increases in standardized catch rates for dhufish between 2008 and 2011, immediately prior to the start of our time-series (Fairclough et al., 2014). Thus, the year effect in our analysis may reflect a continued increase in dhufish numbers, possibly resulting from management reforms in 2009 in response to overfishing (Crowe et al., 2013). For snapper, seasonality was observed with a peak in the Austral winter; likely explanations are the same as for temperature and swell discussed above including increased snapper targeting in winter, and increased catchability following storm events (Crawford, 2016; Crisafulli et al., 2019). Spatial predictions of catch Overall a high degree of spatial structure was found in catch for both dhufish and snapper. Higher catches of dhufish were found in the north of the study area (between 29 and 30°S), and lower catches in the inshore waters of the Perth Metropolitan area and the south coast. These findings largely match those reported by Aidoo et al. (2016) who analysed the 2011/2012 subset of the same survey data using soft-indicator kriging, though the south coast was outside their study scope. Low recreational catch in the inshore waters of the Perth Metropolitan area also matches anecdotal reports from recreational and commercial fishers documented a decade prior to the start of our data series (Hesp et al., 2002). A hotspot of snapper catch was found in the north of the study area (north of 29°S), and to a lesser extent between 33 and 29°S, and in offshore locations between 34 and 33°S. Aidoo et al. (2016) also identified the northern hotspot; however, their analysis suggested lower catches off the southern west coast (west of 115°S) which was not confirmed in this study. This discrepancy appears to be the result of year-to-year variation rather than analytical approach, as raw catch in the south was higher in 2013/2014 and 2015/2016, years not included in the analysis of Aidoo et al. (2016). Spatial trends in recreational catch represent the combined effects of fish population size and behaviour, fisher behaviour, and fisheries management. The confounding effect of these factors, particularly in the way they vary over space and time makes it difficult to determine the cause of spatial trends in recreational catch. However, the latitudinal gradient in snapper catch across the study region suggests that latitudinal correlated processes are important drivers of snapper catch in the fishery. Temperature, which decreases with latitude in the study region, has been linked to snapper growth and reproduction in this fishery and may be important for producing the observed latitudinal catch trend (Wakefield et al., 2015). Dhufish catch on the other hand appears to be substantially lower in the inshore Perth Metropolitan area. Possible causes include lower fish abundance in the Perth Metropolitan area, spatial variability in fisher behaviour (e.g. practicing catch-and-release), or reduced catchability due to the localized effect of high fishing intensity on hook-avoidance behaviour in fish (Alós et al., 2015). Regardless of mechanism, spatial catch variability can have important sustainability implications. Anecdotal reports suggest that fishers have responded to low dhufish catch in the Perth Metropolitan area by fishing in deeper waters (Hesp et al., 2002), subjecting released dhufish to higher mortality rates [dhufish release mortality increases with depth from 21% in 0–14 m to 86% in 45–59 m (St John and Syers, 2005)]. As release rates are high (68%) and a large portion of the effort in the fishery occurs near the Perth Metropolitan area (Ryan et al., 2017), the causes of low recreational catch in the waters of the Perth Metropolitan area, and the behavioural response of fishers, warrants further research. Acknowledgements We acknowledge contributions from the survey participants that were interviewed and staff from the Survey Research Centre (Edith Cowan University) for conducting the phone surveys. Brent Wise, Dave Fairclough, and Alex Hesp provided constructive comments in reviewing the manuscript. We also thank two anonymous reviewers, whose comments helped improve the manuscript. References Aidoo E. N. , Mueller U. , Hyndes G. A. , Ryan K. L. 2016 . The effects of measurement uncertainty on spatial characterisation of recreational fishing catch rates . Fisheries Research , 181 : 1 – 13 . Google Scholar Crossref Search ADS WorldCat Alós J. , Palmer M. , Trías P. , Díaz-Gil C. , Arlinghaus R. 2015 . Recreational angling intensity correlates with alteration of vulnerability to fishing in a carnivorous coastal fish species . Canadian Journal of Fisheries and Aquatic Sciences , 72 : 217 – 225 . Google Scholar Crossref Search ADS WorldCat Arlinghaus R. , Tillner R. , Bork M. 2015 . Explaining participation rates in recreational fishing across industrialised countries . Fisheries Management and Ecology , 22 : 45 – 55 . Google Scholar Crossref Search ADS WorldCat Austin M. P. 2002 . Spatial prediction of species distribution: an interface between ecological theory and statistical modelling . Ecological Modelling , 157 : 101 – 118 . Google Scholar Crossref Search ADS WorldCat Bigelow K. A. , Boggs C. H. , He X. I. 1999 . Environmental effects on swordfish and blue shark catch rates in the US North Pacific longline fishery . Fisheries Oceanography , 8 : 178 – 198 . Google Scholar Crossref Search ADS WorldCat Coleman F. C. , Figueira W. F. , Ueland J. S. , Crowder L. B. 2004 . The impact of United States recreational fisheries on marine fish populations . Science , 305 : 1958 – 1960 . Google Scholar Crossref Search ADS PubMed WorldCat Crawford J. 2016 . Snapper Basics Land-Based and Shallow Water. http://www.fishingworld.com.au/how-to/snapper-basics-part-1 (last accessed 8 October 2018). Crisafulli B. M. , Fairclough D. V. , Keay I. S. , Lewis P. , How J. R. , Ryan K. L. , Taylor S. M. et al. 2019 . Does a spatiotemporal closure to fishing Chrysophrys auratus (Sparidae) spawning aggregations also protect individuals during migration? Canadian Journal of Fisheries and Aquatic Sciences , doi:10.1139/cjfas-2017-0449 WorldCat Crowe F. M. , Longson I. G. , Joll L. M. 2013 . Development and implementation of allocation arrangements for recreational and commercial fishing sectors in Western Australia . Fisheries Management and Ecology , 20 : 201 – 210 . Google Scholar Crossref Search ADS WorldCat Elith J. , Leathwick J. R. , Hastie T. 2008 . A working guide to boosted regression trees . Journal of Animal Ecology , 77 : 802 – 813 . Google Scholar Crossref Search ADS PubMed WorldCat Elmer L. K. , Kelly L. A. , Rivest S. , Steell S. C. , Twardek W. M. , Danylchuk A. J. , Arlinghaus R. et al. 2017 . Angling into the future: ten commandments for recreational fisheries science, management, and stewardship in a good anthropocene . Environmental Management , 60 : 165 – 175 . Google Scholar Crossref Search ADS PubMed WorldCat Fairclough D. V. , Molony B. W. , Crisafulli B. M. , Keay I. S. , Hesp S. A. , Marriott R. J. 2014 . Status of demersal finfish stocks on the west coast of Australia. Fisheries Research Report, 253. 96 pp. Department of Fisheries, Western Australia. Fenichel E. P. , Abbott J. K. , Huang B. 2013 . Modelling angler behaviour as a part of the management system: synthesizing a multi-disciplinary literature . Fish and Fisheries , 14 : 137 – 157 . Google Scholar Crossref Search ADS WorldCat Fisher R. , Wilson S. K. , Sin T. M. , Lee A. C. , Langlois T. J. 2018 . A simple function for full-subsets multiple regression in ecology with R . Ecology and Evolution , 8 : 6104 – 6113 . Google Scholar Crossref Search ADS PubMed WorldCat Friedman J. , Hastie T. , Tibshirani R. 2000 . Additive logistic regression: a statistical view of boosting . Annals of Statistics , 28 : 337 – 374 . Google Scholar Crossref Search ADS WorldCat Garnier S. 2018 . viridis: Default Color Maps from “matplotlib”. R Package Version 0.5.1. http://CRAN.R-project.org/package=viridis. Gomon M. F. , Bray D. J. , Kuiter R. H. 2008 . Fishes of Australia’s Southern Coast. New Holland Publishers Pty Ltd, Sydney, Australia. Griffiths S. P. , Bryant J. , Raymond H. F. , Newcombe P. A. 2017 . Quantifying subjective human dimensions of recreational fishing: does good health come to those who bait? Fish and Fisheries , 18 : 171 – 184 . Google Scholar Crossref Search ADS WorldCat Hahlbeck N. , Scales K. L. , Dewar H. , Maxwell S. M. , Bograd S. J. , Hazen E. L. 2017 . Oceanographic determinants of ocean sunfish (Mola mola) and bluefin tuna (Thunnus orientalis) bycatch patterns in the California large mesh drift gillnet fishery . Fisheries Research , 191 : 154 – 163 . Google Scholar Crossref Search ADS WorldCat Hastie T. , Tibshirani R. 1986 . Generalized additive models . Statistical Science , 1 : 297 – 318 . Google Scholar Crossref Search ADS WorldCat Helfman G. S. 1986 . Fish behaviour by day, night and twilight. In The Behaviour of Teleost Fishes , pp. 366 – 387 . Ed. by Pitcher T. J. . Springer US , Boston, MA . Google Preview WorldCat Henry G. W. , Lyle J. M. 2003 . National recreational and indigenous fishing survey. Final Report for FRDC Project, 99/158. 188 pp. Australian Government Department of Agriculture, Fisheries and Forestry, Canberra, Australia. Hesp S. A. , Potter I. C. , Hall N. G. 2002 . Age and size composition, growth rate, reproductive biology, and habitats of the West Australian dhufish (Glaucosoma hebraicum) and their relevance to the management of this species . Fishery Bulletin , 100 : 214 – 227 . WorldCat Hijmans R. J. , Phillips S. , Leathwick J. , Elith J. 2017 . dismo: Species Distribution Modeling. R Package Version 1.1-4. http://CRAN.R-project.org/package=dismo Hunt L. M. , Sutton S. G. , Arlinghaus R. 2013 . Illustrating the critical role of human dimensions research for understanding and managing recreational fisheries within a social–ecological system framework . Fisheries Management and Ecology , 20 : 111 – 124 . Google Scholar Crossref Search ADS WorldCat Kuhn M. 2019 . caret: Classification and Regression Training. R Package Version 6.0-84. https://CRAN.R-project.org/package=caret. Lan K.-W. , Lee M.-A. , Wang S.-P. , Chen Z.-Y. 2015 . Environmental variations on swordfish (Xiphias gladius) catch rates in the Indian Ocean . Fisheries Research , 166 : 67 – 79 . Google Scholar Crossref Search ADS WorldCat Leathwick J. R. , Elith J. , Francis M. P. , Hastie T. , Taylor P. 2006 . Variation in demersal fish species richness in the oceans surrounding New Zealand: an analysis using boosted regression trees . Marine Ecology Progress Series , 321 : 267 – 281 . Google Scholar Crossref Search ADS WorldCat Martínez-Rincón R. O. , Ortega-García S. , Vaca-Rodríguez J. G. 2012 . Comparative performance of generalized additive models and boosted regression trees for statistical modeling of incidental catch of wahoo (Acanthocybium solandri) in the Mexican tuna purse-seine fishery . Ecological Modelling , 233 : 20 – 25 . Google Scholar Crossref Search ADS WorldCat Maunder M. N. , Punt A. E. 2004 . Standardizing catch and effort data: a review of recent approaches . Fisheries Research , 70 : 141 – 159 . Google Scholar Crossref Search ADS WorldCat Mitchell J. D. , McLean D. L. , Collin S. P. , Taylor S. , Jackson G. , Fisher R. , Langlois T. J. 2018 . Quantifying shark depredation in a recreational fishery in the Ningaloo Marine Park and Exmouth Gulf, Western Australia . Marine Ecology Progress Series , 587 : 141 – 157 . Google Scholar Crossref Search ADS WorldCat Pope K. L. , Chizinski C. J. , Wiley C. L. , Martin D. R. 2016 . Influence of anglers’ specializations on catch, harvest, and bycatch of targeted taxa . Fisheries Research , 183 : 128 – 137 . Google Scholar Crossref Search ADS WorldCat Post J. R. , Sullivan M. , Cox S. , Lester N. P. , Walters C. J. , Parkinson E. A. , Paul A. J. et al. 2002 . Canada’s recreational fisheries: the invisible collapse? Fisheries , 27 : 6 – 17 . Google Scholar Crossref Search ADS WorldCat Reynolds R. W. , Smith T. M. , Liu C. , Chelton D. B. , Casey K. S. , Schlax M. G. 2007 . Daily high-resolution-blended analyses for sea surface temperature . Journal of Climate , 20 : 5473 – 5496 . Google Scholar Crossref Search ADS WorldCat Ryan K. L. , Hall N. G. , Lai E. K. , Smallwood C. B. , Taylor S. M. , Wise B. S. 2015 . State-wide survey of boat-based recreational fishing in Western Australia 2013/14. Fisheries Research Report, 268. 208 pp. Department of Fisheries, Western Australia. Ryan K. L. , Hall N. G. , Lai E. K. , Smallwood C. B. , Taylor S. M. , Wise B. S. 2017 . Statewide survey of boat-based recreational fishing in Western Australia 2015/16. Fisheries Research Report, 287. 205 pp. Department of Primary Industries and Regional Development, Western Australia. Ryan K. L. , Wise B. S. , Hall N. G. , Pollock K. H. , Sulin E. H. , Gaughan D. J. 2013 . An integrated system to survey boat-based recreational fishing in Western Australia 2011/12. Fisheries Research Report, 249. 168 pp. Department of Fisheries, Western Australia. St John J. , Syers C. J. 2005 . Mortality of the demersal West Australian dhufish, Glaucosoma hebraicum (Richardson 1845) following catch and release: the influence of capture depth, venting and hook type . Fisheries Research , 76 : 106 – 116 . Google Scholar Crossref Search ADS WorldCat Stoner A. W. 2004 . Effects of environmental variables on fish feeding ecology: implications for the performance of baited fishing gear and stock assessment . Journal of Fish Biology , 65 : 1445 – 1471 . Google Scholar Crossref Search ADS WorldCat Wakefield C. B. , Potter I. C. , Hall N. G. , Lenanton R. C. J. , Hesp S. A. 2015 . Marked variations in reproductive characteristics of snapper (Chrysophrys auratus, Sparidae) and their relationship with temperature over a wide latitudinal range . ICES Journal of Marine Science , 72 : 2341 – 2349 . Google Scholar Crossref Search ADS WorldCat Walbridge S. , Slocum N. , Pobuda M. , Wright D. J. 2018 . Unified geomorphological analysis workflows with Benthic Terrain Modeler . Geosciences , 8 : 94. Google Scholar Crossref Search ADS WorldCat Whiteway T. G. 2009 . Australian bathymetry and topography grid, June 2009. Geoscience Australia Record, 2009/21. 46 pp. Geoscience Australia. Wise B. S. , St John J. , Lenanton R. C. J. 2007 . Spatial scales of exploitation among populations of demersal scalefish: implications for management. Part 1: stock status of the key indicator species for the demersal scalefish fishery in the West Coast Bioregion. 130 pp. Department of Fisheries, Western Australia. Wood S. , Scheipl F. 2014 . gamm4: Generalized Additive Mixed Models Using mgcv and lme4. R Package Version 0.2-5. http://CRAN.R-project.org/package=gamm4 Yu H. , Jiao Y. , Carstensen L. W. 2013 . Performance comparison between spatial interpolation and GLM/GAM in estimating relative abundance indices through a simulation study . Fisheries Research , 147 : 186 – 195 . Google Scholar Crossref Search ADS WorldCat © International Council for the Exploration of the Sea 2019. All rights reserved. For permissions, please email: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)
TI - Determining spatial patterns in recreational catch data: a comparison of generalized additive mixed models and boosted regression trees
JF - ICES Journal of Marine Science
DO - 10.1093/icesjms/fsz123
DA - 2019-09-01
UR - https://www.deepdyve.com/lp/oxford-university-press/determining-spatial-patterns-in-recreational-catch-data-a-comparison-o1hd5Lr3yq
SP - 1
VL - Advance Article
IS - 
DP - DeepDyve
ER -