Bottom trawl fishing footprints on the world's continental shelves.

Bottom trawlers land around 19 million tons of fish and invertebrates annually, almost one-quarter of wild marine landings. The extent of bottom trawling footprint (seabed area trawled at least once in a specified region and time period) is often contested but poorly described. We quantify footprints using high-resolution satellite vessel monitoring system (VMS) and logbook data on 24 continental shelves and slopes to 1,000-m depth over at least 2 years. Trawling footprint varied markedly among regions: from <10% of seabed area in Australian and New Zealand waters, the Aleutian Islands, East Bering Sea, South Chile, and Gulf of Alaska to >50% in some European seas. Overall, 14% of the 7.8 million-km2 study area was trawled, and 86% was not trawled. Trawling activity was aggregated; the most intensively trawled areas accounting for 90% of activity comprised 77% of footprint on average. Regional swept area ratio (SAR; ratio of total swept area trawled annually to total area of region, a metric of trawling intensity) and footprint area were related, providing an approach to estimate regional trawling footprints when high-resolution spatial data are unavailable. If SAR was ≤0.1, as in 8 of 24 regions, there was >95% probability that >90% of seabed was not trawled. If SAR was 7.9, equal to the highest SAR recorded, there was >95% probability that >70% of seabed was trawled. Footprints were smaller and SAR was ≤0.25 in regions where fishing rates consistently met international sustainability benchmarks for fish stocks, implying collateral environmental benefits from sustainable fishing.

Using vessel monitoring system data to improve systematic conservation planning of a multiple-use marine protected area, the Kosterhavet National Park (Sweden).

When spatial fishing data is fed into systematic conservation planning processes the cost to a fishery could be ensured to be minimal in the zoning of marine protected areas. We used vessel monitoring system (VMS) data to map the distribution of prawn trawling and calculate fishing intensity for 1-ha grid cells, in the Kosterhavet National Park (Sweden). We then used the software Marxan to generate cost-efficient reserve networks that represented every biotope in the Park. We asked what were the potential gains and losses in terms of fishing effort and species conservation of different planning scenarios. Given a conservation target of 10 % representation of each biotope, the fishery need not lose more than 20 % of its fishing grounds to give way to cost-efficient conservation of benthic diversity. No additional reserved area was needed to achieve conservation targets while minimizing fishing costs. We discuss the benefits of using VMS data for conservation planning.

Towards a scientific community consensus on designating Vulnerable Marine Ecosystems from imagery.

Management of deep-sea fisheries in areas beyond national jurisdiction by Regional Fisheries Management Organizations/Arrangements (RFMO/As) requires identification of areas with Vulnerable Marine Ecosystems (VMEs). Currently, fisheries data, including trawl and longline bycatch data, are used by many RFMO/As to inform the identification of VMEs. However, the collection of such data creates impacts and there is a need to collect non-invasive data for VME identification and monitoring purposes. Imagery data from scientific surveys satisfies this requirement, but there currently is no established framework for identifying VMEs from images. Thus, the goal of this study was to bring together a large international team to determine current VME assessment protocols and establish preliminary global consensus guidelines for identifying VMEs from images. An initial assessment showed a lack of consistency among RFMO/A regions regarding what is considered a VME indicator taxon, and hence variability in how VMEs might be defined. In certain cases, experts agreed that a VME could be identified from a single image, most often in areas of scleractinian reefs, dense octocoral gardens, multiple VME species' co-occurrence, and chemosynthetic ecosystems. A decision flow chart is presented that gives practical interpretation of the FAO criteria for single images. To further evaluate steps of the flow chart related to density, data were compiled to assess whether scientists perceived similar density thresholds across regions. The range of observed densities and the density values considered to be VMEs varied considerably by taxon, but in many cases, there was a statistical difference in what experts considered to be a VME compared to images not considered a VME. Further work is required to develop an areal extent index, to include a measure of confidence, and to increase our understanding of what levels of density and diversity correspond to key ecosystem functions for VME indicator taxa. Based on our results, the following recommendations are made: 1. There is a need to establish a global consensus on which taxa are VME indicators. 2. RFMO/As should consider adopting guidelines that use imagery surveys as an alternative (or complement) to using bycatch and trawl surveys for designating VMEs. 3. Imagery surveys should also be included in Impact Assessments. And 4. All industries that impact the seafloor, not just fisheries, should use imagery surveys to detect and identify VMEs.

Predicting the distribution of out-of-reach biotopes with decision trees in a Swedish marine protected area.

Through spatially explicit predictive models, knowledge of spatial patterns of biota can be generated for out-of-reach environments, where there is a paucity of survey data. This knowledge is invaluable for conservation decisions. We used distribution modeling to predict the occurrence of benthic biotopes, or megafaunal communities of the seabed, to support the spatial planning of a marine national park. Nine biotope classes were obtained prior to modeling from multivariate species data derived from point source, underwater imagery. Five map layers relating to depth and terrain were used as predictor variables. Biotope type was predicted on a pixel-by-pixel basis, where pixel size was 15 x 15 m and total modeled area was 455 km2. To choose a suitable modeling technique we compared the performance of five common models based on recursive partitioning: two types of classification and regression trees ([1] pruned by 10-fold cross-validation and [2] pruned by minimizing complexity), random forests, conditional inference (CI) trees, and CI forests. The selected model was a CI forest (an ensemble of CI trees), a machine-learning technique whose discriminatory power (class-by-class area under the curve [AUC] ranged from 0.75 to 0.86) and classification accuracy (72%) surpassed those of the other methods tested. Conditional inference trees are virtually new to the field of ecology. The final model's overall prediction error was 28%. Model predictions were also checked against a custom-built measure of dubiousness, calculated at the polygon level. Key factors other than the choice of modeling technique include: the use of a multinomial response, accounting for the heterogeneity of observations, and spatial autocorrelation. To illustrate how the model results can be implemented in spatial planning, representation of biodiversity in the national park was described and quantified. Given a goal of maximizing classification accuracy, we conclude that conditional inference trees are a promising tool to map biota. Species distribution modeling is presented as an ecological tool that can handle a wide variety of systems (e.g., the benthic system).