Deep Vision, 2016 Shale Rosen, Havforskningsinstituttet
Developments, 2016 Live acoustic communication of statistic and metadata (not images) via SIMRAD sensor Design of active sorting / subsampling mechanism (J2 Subsea, Scotland, possible test 11.2016) System integration (integrating DeepVision data and images with LSSS acoustic analysis software, Christian Michelsen Research AS) Automatic object detection, partnership with University of Girona, Spain)
Use as fisheries research tool (outside of CRISP) Study trawl method used for mackerel in North-Atlantic Improved species identification for acoustic survey Masters project, Bård Aarbakke: Suggests changes to trawling technique Showed problems with current trawl technique (too slow to catch saithe effectively)
Use as fisheries research tool (outside of CRISP) EU Horizon 2020 project MINOUW Deep Vision to estimate likely bycatch and discards without actually catching fish
Winner, Innovation Prize Nor-Fishing 2016
Bård Aarbakke MSc (03.08.2016) Used Deep Vision data to evaluate trawling technique used to assess Atlantic mackerel Currently employed by Scantrol Deep Vision, testing analysis software
Deep Vision performance Bård Aarbakke, Scantrol Deep Vision AS
Hired to look into different aspects analysis of Deep Vision data 1. Time to count all fish in a typical trawl haul 2. Time to randomly measure lengths of 100 mackerel within a haul 3. Measurability of fish in different concentrations 4. Consistency of length measurements 5. Comparison of Deep Vision lengths with lengths collected by technicians
New software
Time to count all fish in a typical trawl haul Analysed first 20 seconds in each 1 minute interval over 30 minutes (1/3 of total images). Fish counted as they left the DV channnel (codend side) Special technique when a lot of fish entered in a dense aggregation
Used 1 hour and 12 minutes to analyze the mackerel haul 42 minutes on the Saithe haul Time to count all fish in a 30-minute haul Mackerel: Time intervals Entrance rate (20 seconds intervals per minutes) Entrance rate in total (Estimated per minute) 12:35:00-12:36:00 76 228 12:36:00-12:37:00 72 216 12:37:00-12:38:00 45 135 12:38:00-12:39:00 24 72 12:39:00-12:40:00 32 96 12:40:00-12:41:00 23 69 12:41:00-12:42:00 16 48 12:42:00-12:43:00 11 33 12:43:00-12:44:00 5 15 12:44:00-12:45:00 8 24 12:45:00-12:46:00 8 24 12:46:00-12:47:00 7 21 12:47:00-12:48:00 7 21 12:48:00-12:49:00 7 21 12:49:00-12:50:00 5 15 12:50:00-12:51:00 12 36 12:51:00-12:52:00 6 18 12:52:00-12:53:00 7 21 12:53:00-12:54:00 15 45 12:54:00-12:55:00 3 9 12:55:00-12:56:00 13 39 12:56:00-12:57:00 6 18 12:57:00-12:58:00 6 18 12:58:00-12:59:00 8 24 12:59:00-13:00:00 12 36 13:00:00-13:00:00 4 12 13:01:00-13:01:00 6 18 13:02:00-13:02:00 4 12 13:03:00-13:03:00 8 24 13:04:00-13:04:00 10 30 Total 466 1398 Saithe: Time intervals Entrance rate (20 seconds per minutes) Entrance rate in total (Estimated per minute) 20:40:00-20:41:00 0 0 20:41:00-20:42:00 0 0 20:42:00-20:43:00 0 0 20:43:00-20:44:00 0 0 20:44:00-20:45:00 0 0 20:45:00-20:46:00 0 0 20:46:00-20:47:00 0 0 20:47:00-20:48:00 0 0 20:48:00-20:49:00 0 0 20:49:00-20:50:00 1 3 20:50:00-20:51:00 0 0 20:51:00-20:52:00 1 3 20:52:00-20:53:00 0 0 20:53:00-20:54:00 1 3 20:54:00-20:55:00 0 0 20:55:00-20:56:00 0 0 20:56:00-20:57:00 0 0 20:57:00-20:58:00 0 0 20:58:00-20:59:00 3 9 20:59:00-21:00:00 3 9 21:00:00-21:01:00 0 0 21:01:00-21:02:00 2 6 21:02:00-21:03:00 1 3 21:03:00-21:04:00 5 15 21:04:00-21:05:00 16 48 21:05:00-21:06:00 21 63 21:06:00-21:07:00 48 144 21:07:00-21:08:00 50 150 21:08:00-21:09:00 71 213 21:09:00-21:10:00 112 336 Total 335 1005
Time to randomly measure lengths of 100 mackerel within a haul The length measurements done by: 1. Generate 20 random times within the time period of interest 2. Measure 5 fish on each time period If the picture did not contain 5 fish, the next picture was used Measured every measurable fish in the last image If the time period overlapped, measurements of 5 new fish was started 3. Stop once 100 fish were measured All fish in the last picture were counted Sometimes this method leads to more then 100 measurements
Time consumption of length measurements (100 random mackerel) Measuring 100 random mackerel takes approximately one and a half hours (1:28:17), 53 seconds per fish
Measurability of fish in different concentrations (Percentage measurable fish) Found 1 minute periods with different fish concentrations Counted entrance rates in the 1 minute intervals (entrance rate per minute) Counted fish which could not be measured (visual assessment) Did not try to measure every fish
Measurability of fish in different concentrations % non-measurable 80 60 40 20 0 Average measurability in Mackerel = 95,2% Average measurability in Saithe = 71,8% Average measurability in Cod = 68,3% 0 100 200 300 400 500 # per minute Mackerel Saithe Mackerel: Saithe : Cod: Fish per minute (N) Non-measurable fish (%) 32 2,6 38 3,1 191 1,1 277 8,7 407 6,9 Fish per minute (N) Non-measurable fish (%) 12 25 14 42,9 92 28,4 112 23,21 154 27,3 Fish per minute (N) Non-measurable fish (%) 3 0 6 16,7 52 28,9 75 25,3 215 64,2
Consistency of length measurements Found fish which remained in the DV chamber from 8 31 pictures Did repeated measurements on the same fish 80 155 measurements Calculate standard deviation and percentage of length measurements within +/- 10 mm of the average length Two points on straight fish, more on bent fish
Consistency of length measurements Mackerel: Average size (mm) Saithe: Average size (mm) Standard deviation Standard deviation SD compare d to length (%) SD compar ed to length (%) Measurements within a +/- 10 mm range of average length(%) Measurements within a +/- 10 mm range of average length(%) Number of measurements Number of measurements Number of different pictures Number of different pictures Difficult fish pictures (pictures to the right) seems to have a higher influence on the SD compared to fish length Minimum length measured Minimum length measured Maximum length measured 240,6 7,0 2,9% 88,9 126 14 228,0 265,7 338,9 15,9 4,7% 57,4 155 31 246,2 380,3 Maximum length measured 476,5 17,9 3,75 59,5 84 14 447,4 576,8 563,7 9,2 1,63 77,5 80 8 541,8 586,3
Length distribution Length distribution from Deep Vision was compared with length distribution from data collected from technicians onboard R/V G.O. Sars DV length data from 200 fish Put through an equation to transform fork length to standard length (SL = 1,0956 * FL + 0,8256) (Shale Rosen, 2012) Length data from 100 fish taken by technicians Subsampled on deck and collected from a codend with split Compared by plotting the data in a cumulative curve and in a histogram Statistically tested by using Kolmogorov-Smirnov test FL SL
Length distribution Cumulative curves P = 0,048 P = 2,2e^-16 P = 2,5e^-14 P = 2,2e^-16 P = 1,8e^-5
Length distribution Histogram
Recap 1. For 30 minute haul: 72 minutes to count 466 mackerel, 42 minutes to count 335 saithe (sampling 20 seconds of each minute) 2. 88 minutes to measure lengths of 100 random mackerel within a haul 3. On average, 95,2 % of mackerel; 71,8 % of saithe; 68,3 % of cod can be measured 4. Percentage of measurements that was +/- 10 mm of average length varies from 57,4 % to 88,9% 5. Statistically significant difference between Deep Vision lengths and lengths collected by technicians for 4 of 5 hauls (mackerel) Deep Vision measurements show larger fish
Thank you for your attention Any questions?