A ROBOT submarine that can be taught to recognise any fish species could soon be helping conservationists find out if fish populations really are as close to collapse as some suspect.
Fish stocks are currently estimated by measuring the size of sample catches. But the samples are often caught at points up to 100 kilometres apart, leading to potentially misleading assumptions about how many fish there are lurking between the sampling points.
But there might be a better way. Daniel Doolittle and his colleagues at the Virginia Institute of Marine Science in Gloucester Point have developed an autonomous underwater vehicle (AUV) that takes sonar pictures of passing fish shoals and uses an artificial intelligence system to recognise the fish species in question and count them.
Advertisement
The idea is that a fisheries ship could deploy a flotilla of the subs to count populations in many different areas at once, and also search rough seabed terrain where trawling is impossible. The AUVs move at a leisurely 1 or 2 metres per second on a pre-programmed course, and can travel up to 100 kilometres before their batteries need charging.
Doolittle鈥檚 subs use sonar rather than light because the sea is often too murky for optical cameras, and although strong lights can help these often scare away many of the fish. So Doolittle bounces sound waves off the fish to create sound-based images.
The sonar system emits sound pulses, which fan out from either side of the sub and bounce off anything swimming nearby. The returning waves vibrate the sonar transducers, generating electrical signals that form the images.
The resolution of sonar images increases with the frequency of the vibrations, but higher frequencies cannot penetrate as far through the water. The sub鈥檚 system vibrates at 600 kilohertz, which gives a resolution of just 2 centimetres within 10 metres either side of the AUV.
This resolution is better than that of commercially available fish-finders, but conventional pattern recognition software still cannot tell a cod from a haddock, says Doolittle. So he and his colleagues have designed neural network software that can be trained to recognise any number of different species by their shape and the way they move.
The neural network learns which combinations of inputs, such as shape details, lead to a particular output, such as a positive species identification. During training, the network鈥檚 inputs are weighted, so the influence of each can be increased or decreased depending on its significance to the outcome. The institute鈥檚 neural network already recognises four types of fish and will be trained to spot more.
Michael Kaiser, a fisheries researcher at the School of Ocean Sciences at the University of Wales in Bangor, agrees that Doolittle鈥檚 device may provide an answer to the fishing industry鈥檚 criticism of current methods of fish stock assessment. The USNavy is also interested in the smart subs, which could be put to work patrolling harbours or shipping lanes on the lookout for mines or other weapons.
