With some time passing since we gave any update on the Nethermills counter on the Ayr, I thought now was as good a time as any.
What became obvious to us after commencing this project back in 2022 was that it was never going to be as simple as download an accurate figure that the counter generates automatically. Neither was verifying the accuracy of the counter due to so many variables and limitations. I expect I was rather naive and perhaps slightly misled when I agreed to undertake the counter maintenance and downloads etc. What we believed would be relatively simple and quick has become a monster with legs that commands a huge amount of time from us. That is time that drains the trust of resources as there is no remuneration at all for this work and never has been. That is something we are hoping will change soon.
We are continually refining whatever we can along with Marine Directive (MD) to make things more streamline and accurate however, we realised that probably the best way to improve accuracy and speed the process was to develop AI software, ultimately replacing the need for laborious manual verification etc. Verification will still be required but should reduce as the AI learns.
For the last 8 months or so, we have been working with partners at MD, East Anglia University Fishery Management Scotland, and Sean from SFCC/FMS to develop the AI. ART’s task was identifying trout, salmon and smolts and then upload this so that the software could effectively teach a computer to identify fish and species. This is still work in progress with the team devoting many hundreds of hours to this. In fact, I’ve spent much of this weekend scrutinising video from the last few weeks that will also be run through the software, then used to generate a count. This is all very experimental as yet but early results are encouraging. We are nearing the point where we can trial this AI system as soon as the final touches to it are made. The more files the AI processes, the more accurate it should become. We will need to continue manual identification and verification at the moment to ensure the results are good and check progress with accuracy.
The variables in the system are immense and must all be factored in when attempting to gain accuracy. Sunlight/shadows, time of day, turbidity, water level, clarity, direction of travel, angle of fish to camera, detritus etc. etc. etc. all influence the ability of the AI and those identifying the fish (ART).
A large sea trout returning to the river or is it a grilse? In my opinion it is a trout but it can be very difficult when they move through at speed under the cover of darkness. It was only as the fish left the frame that the thickness of the wrist at the tail gave it away as a trout. There will always be a small factor of error.
The AI results and manual verification will then be cross referenced to the original resistivity counter data and hopefully then MD will be able to come up with a more accurate count each year.
Counting smolts is something that the original counter just can’t do. Using AI to achieve this would be an incredible win for us particularly if we can design a system to guide smolts into the fish pass rather than through the turbine. We have a few ideas on this but feel free to make suggestions. Again, this is a fair challenge but what a tremendous benefit it would be to use the fish pass cameras and AI software as an automatic smolt counter able to provide a minimum estimate of smolts leaving the river.
A shoal of smolts leaving the river in good viewing conditions. This should be relatively easy count for the AI under such conditions.
In the next week or two, further Infra Red panels and a new IR overhead camera and light source will be installed in the fish pass completing the physical upgrades to the system. The rest of the improvements will come by allowing the AI to do its stuff and teaching it when it makes errors.
Bear with us. This is still work in progress and there’s no one keener than ourselves to be able to come up with a meaningful minimum estimate count.
A large salmon going through the counter in good visibility. There’s no mistaking what this fish was under suck clear viewing conditions
