A research survey led by Marine Scotland Science as part of the SMARTFISH Work Package 8 (WP8): West of Scotland and northern North Sea on-board FRV Scotia (Figure 1) set sail on the evening of Saturday 07th November to commence fishing on grounds East of Fair Isle. The survey was collecting data for the testing and demonstration of high-tech systems for the EU fishing sector developed under the SMARTFISH 2020 international research project. During the first 3 days 12 fishing tows were made fishing with the BT237 demersal otter trawl gear (Figure 2). The mixed species catches were passed along the on-board conveyor belt below several cameras capturing video and images for machine learning (ML) applications. These ML applications are being developed to count fish, identify them to species level and where possible estimate their length and weight. The goal is to improve automated data collection for fish stock assessment and provide evidence of compliance with fishery regulations.
Figure 1. Marine Scotland Fishing Research Vessel (FRV) Scotia in Aberdeen harbour.
Figure 2. WP8 Scotia research cruise fishing locations within the northern North Sea (ICES Division 27.4.a).
Photos of fish along with their weights and lengths were taken using the CatchSnap photography boards (Figure 3). These images will be used to train a mobile phone app style ML system designed for industry self-sampling or at-sea observer programmes. Melbu systems and SINTEF Ocean the partners in this project requested around 300 images per species to train the species classifier. This target number of images was reached for 5 species (haddock, whiting, Norway pout, mackerel and herring) and > 2500 images were collected in total across 20 species.
Figure 3. Individual fish being sampled on the CatchSnap photography boards, the ID number, length (mm) and weight (g) of the individual is recorded on the wipe clean information strip.
Images were taken for three species using the CatchScanner (Figures 4 and 5), this is a laser scanning unit that sits over the conveyor belt and scans single fish as they pass beneath a set of 3 lasers. This unit was calibrated remotely by Melbu partners during the loading of the vessel. While there were some teething issues with the 3 coloured lasers (R,G,B) not firing at the same time resulting in a sub-optimal image and colour quality > 1000 haddock, 8 saithe and 38 cod were passed through the scanner in optimal laser conditions. In addition > 2000 haddock, 13 saithe and 56 cod were scanned while not all of the lasers were operational. Although the issues with the lasers were unfortunate the images with missing laser colours will be used to train the system for robustness.
Figure 4. The CatchScanner unit, with laser set up.
Figure 5. Saithe passing through the laser beam and being detected on the CatchScanner monitor.
All trawl catches were passed along the conveyor 3 times at different fish densities/levels of occlusion, from complete belt coverage to only a few fish overlapping. The catches were then sampled to get the number of fish by species and the length distributions. These will be compared to the classification of fish from the CatchMonitor ML system developed by the University of East Anglia (UEA) to explore the effect of fish density (in camera view) on the performance of the instance segmentation system (that detects and outlines individual fish in images) and the image classifier (that identifies species of individual fish).
The last two days of the survey were dedicated to scallop dredging in the Moray Firth which was well timed as the wind speeds and swell had picked up and we were able to shelter closer to shore. This is the first time scallop dredges have been deployed from Scotia and a bespoke short bar and bridle with two dredges was assembled to enable the dredges to be lifted over the stern using the GAMMA frame (Figure 6). The crew quickly found a consistent deployment process which was aided by the use of guide ropes to prevent the bar from rotating whilst suspended. 10 successful tows were made with 2 additional foul hauls where the dredges were upside down when retrieved (Figure 2).
The dredge catches from multiple tows were combined, including the scallops, bycatch, shell debris and rocks and passed along the conveyor to replicate the conveyors on commercial scallop dredge vessels. Scallops were then measured and aged on board by scientists, and individual photos were taken using the CatchSnap boards (Figure 6). This will be the first footage of scallops to be provided to UEA and it is hoped that over the coming year elements of automated catch sampling will be developed for this fishery. This will include identifying and counting scallops within the catch, estimating their width, and potentially their age (using annual growth rings). A MASTS PhD student visiting from the University of Aberdeen also made additional recordings with a go-pro to test the hypothesis that scallops could be aged from video footage and what the optimal camera distance to belt and configurations of resolution, lighting etc. should be.
Figure 6. Scallop dredge deployment aft, scallop video footage collection and scallop CatchSnap sampling.
Overall the sea conditions and weather were great for a November survey with light winds, little swell and occasional fog. The testing of a variety of new catch sampling technologies within the fish house as well as the first deployment on Scotia of a novel gear was been a great success thanks to the hard work of the scientists and crew.
Scientist In Charge