3.1. Model comparison
The majority of the candidate models were able to correctly characterize the effects of the covariates based on our prior knowledge. Depth and substrate were, as expected, important environmental drivers of Dungeness crab detection and on average the probability of Dungeness crab detection was found to decrease with increasing depth, salinity range, slope, and fine-scale BPI and increase in muddier and sandier areas, as well as elevated areas (high BPI) at a broad 1–10 km scale (
Fig. 4). However, the ecological plausibility of the candidate models did vary. The trawl, trap-DC, and survey-effect models had weaker effects (i.e., coefficients closer to zero) for almost all fixed effects except depth, where they showed a stronger effect. In contrast, dive- and gear-effect models tended to show stronger effects of most covariates except depth. Given that Dungeness crab are known to inhabit sandy and muddy areas, trawl, trap-DC, and survey-effect models appeared to be under-estimating the effect of substrate as both their sandiness and muddiness index parameter estimates were close to or overlapping zero (
Fig. 4). The effect of the Normal(0,1) priors, that operated as penalties on parameter coefficients, is shown in the supplemental materials (Fig. S7).
Salinity was also an important environmental driver of Dungeness crab detection; however, in this case, only the trap-GC model characterized the expected effect. Most models predicted high probability of Dungeness crab detection in low salinity areas. This mischaracterization could be driven by data gaps. There were few observations in low salinity areas (below 26 PSU) across all data sources (
Fig. 2) and no observations below 20 PSU, thus the models are extrapolating in those areas. Additionally, the tight correlation between bottom salinity and bottom temperature that occurs within the sampled area could be obscuring the true salinity effect, instead showing a combination of the confounded salinity and temperature effects.
The survey-effect and gear-effect integrated models differed in their detectability coefficient estimates. The survey-effect model estimated trap-DC surveys to be 24 times more likely to detect a crab compared to trap-GC surveys and six times more likely to detect a crab compared to trawl surveys. In contrast, the gear-effect model assumes both trap surveys have the same detectability and estimated trap surveys to be 1.4 times more likely to detect a crab compared to trawl surveys (
Fig. 4).
Parameter uncertainty, estimated from the mean standard error of fixed effect coefficients, was larger for the single-survey models on average, and confidence intervals around model coefficients were more likely to overlap zero in single-survey models (
Fig. 4). The trawl model had the highest mean parameter uncertainty (mean SE = 0.84), followed by the trap-DC model (SE = 0.50). The survey-effect and gear-effect integrated models showed a 62% and 48% decrease in parameter uncertainty (SE = 0.19 and SE = 0.26) compared to the trap-DC model, respectively.
The marginal effects of model covariates were largely consistent between models (Fig. S8). The greatest variability among the marginal effect curves was seen between Trawl and trap-GC single-survey models and the integrated models predicting to those surveys. Less variability among marginal effect curves was observed between trap-DC and survey-effect models. However, there were still differences in magnitude: the survey-effect model predicted lower probability of Dungeness crab detection at deeper depths, greater salinity ranges, lower broad-scale BPI, and higher fine-scale BPI compared to the trap-DC model.
The relative influence of the fixed and random effects varied between models (Fig. S9). Latent spatial effects, year, salinity, and survey- or gear-type covariates were the most influential overall. Spatial random effects accounted for more than 37% of the relative influence on average. The trap-GC model had the highest relative influence of latent spatial effects at 86% and the survey-effect had the lowest at 11%. Year was highly influential in both trawl and trap-DC models, but accounted for very little relative influence in all other models. Survey type was the most influential predictor for the survey-effect model, accounting for 71% of the relative influence, followed by depth with 12%. The only other model where depth was a relatively influential predictor was the trap-DC model. Gear type was much less influential (13%) in the gear-effect model than was survey type in the survey-effect model (71%).
Predictive performance, evaluated using spatially buffered LOOCV, varied considerably between candidate models with AUC ranging from 0.30 to 0.90 (
Fig. 5). Dive, trawl, and trap-GC single-survey models showed low predictive performance when evaluated on their respective datasets (mean AUC < 0.61 and mean Tjur’s
R2 < 0.10). Conversely, the trap-DC single-survey model had high predictive performance (mean AUC =0.9 and mean
R2 = 0.35) when evaluated with the trap-DC data.
When comparing each single-survey model to each integrated model using their respective single-survey datasets, the integrated models performed at least as well as the single-survey models based on both metrics, as indicated by their overlapping confidence intervals (
Fig. 5), with the exception of the gear-effect model that did not perform as well as the trap-DC model when evaluated with trap-DC data. Based on the LOOCV routine, the survey-effect integrated model outperformed the gear-effect integrated model. However, the LOOCV routine was unable to differentiate between the highest performing models: the trap-DC model and the survey-effect model (mean AUC > 0.89 and mean Tjur’s
R2 > 0.34), which had overlapping confidence intervals when evaluated with trap-DC data.
For validation plots comparing theoretical and sample randomized quantile residuals for each model, see the supplemental materials (Fig. S10).
3.2. Coastwide model predictions and fisheries catch
Here, we compare the coastwide predictions of Dungeness crab detection from the best performing single-survey model with the integrated models, using fisheries catch records as means of independent evaluation. Although it is unlikely that the trap-DC model will perform well when predicting the coastwide occurrence of Dungeness crab, given that the model was trained only on trap-DC data, it is included here both to confirm that hypothesis and serve as a control from which to measure any potential improvements in the integrated models.
Independent data evaluation with the fisheries CPUE index yielded different results to the LOOCV evaluation. In this case, gear-effect predictions were the best performing, with the largest mean hotspot overlap (54.7 ± 4.5 % overlap, ± 2SE) calculated across all quantiles from 0.65 to 0.95 (
Fig. 6). Survey-effect predictions had the second highest mean hotspot overlap (49.9 ± 4.9), followed by trap-DC predictions (44.3 ± 4.4).
Qualitatively, large differences between the trap-DC and the integrated models were observed in their coastwide predictions of probability of Dungeness crab detection (
Fig. 6). The trap-DC model predicted high probabilities throughout most of the region, with only deeper areas of the shelf and the inlets showing lower, but still relatively high probability of Dungeness crab detection. In comparison, gear-effect and survey-effect models had much patchier predictions, with more areas of the coast with low probabilities. At a broad scale, the gear-effect and survey-effect coastwide predictions were similar and areas predicted to have high probability of Dungeness crab detection aligned well with the 90th percentile CPUE hotspots (
Fig. 6 light blue polygons). All three models predicted high probability of Dungeness crab detection around CPUE hotspots near Dogfish Bank, Chatham Sound, Clayoquot Sound, the southern Gulf Islands, and the banks off Vancouver (
Fig. 6). However, the survey-effect predictions differed from the gear-effect predictions along the North Central Coast where the survey-effect model predicted higher probability of Dungeness crab detection in rocky coastal areas.
When comparing coastwide predictions of habitat—areas with greater than threshold probability of Dungeness crab detection—the differences between model predictions were even more pronounced (
Fig. 7). The trap-DC model predicted the largest area of Dungeness crab habitat within the study region (82 217 ± 11 008 km
2, ± 2SE), double the habitat area predicted by the survey-effect model (43 226 ± 12 445 km
2), and more than four times the habitat area predicted by the gear-effect model (18 0292 ± 5391 km
2). Standard error values around the mean coastwide habitat area represent the variability across habitat predictions from the three optimized thresholds used to convert probability into binary habitat predictions.
When comparing the distribution of the probability of Dungeness crab detection inside and outside the fishery footprint (the area exploited by the Dungeness crab fishery between years 2007 and 2017), we found similar results to the hotspot overlap analysis (
Fig. 8). Both integrated models had roughly three times the area of correctly than incorrectly classified habitat based on the fishery footprint (gear effect = 2.97 ± 0.27 and survey effect 2.70 ± 0.60, ±2SE). In contrast, the trap-DC model predictions had roughly the same area of correctly and incorrectly classified habitat (0.93 ± 0.27).
To further compare the trap-DC, gear-effect, and survey-effect coastwide predictions, we considered their spatial prediction uncertainty. Prediction confidence intervals differed more between trap-DC and integrated models (
Fig. 9). For the trap-DC model, even the lower bound of the coastwide prediction showed high probability of detection predictions across most of the area, exhibiting a similar spatial pattern to the higher bound prediction from the survey-effect model. Conversely, the lower bound of the gear-effect model showed lower probability of detection in most areas. While all three models had wide prediction confidence intervals in some areas of the coast, they all also exhibited lower uncertainty (i.e., narrow confidence interval range) in the high probability of Dungeness crab detection areas around Dogfish Bank—the area with the largest commercial CPUE hotspot—and hotspot areas near Chatham Sound, Clayoquot Sound, Vancouver, and the southern Gulf Islands (see
Fig. 6 for the locations of these regions).