FROM PAGE C1
of Moore’s Law – the exponentially falling cost of any computer-based technology – the world suddenly finds itself flooded with cheap sensors, cheap processing boards, cheap communications.
There is an abundance of off-theshelf tech waiting to be taken advantage of. You don’t need a corporate R&D department building entire systems from scratch, just a team with the smarts to glue sophisticated components together, he says.
‘‘It’s the start-up mentality. Agile, or ‘fail fast’. Try it and see if it works. Then move on to the next idea if it doesn’t.’’
Finlay-Smits says Cacophony breaks down into three subprojects, each of which will be a useful predator control achievement in itself.
The first is the AI recognition system. This has been knocked up quickly using available hardware.
The thermal camera is a USmade $230 Flir Lepton 3 – a single chip device smaller than your finger-nail. The computer is a $50 Raspberry Pi board with built-in wireless. The software is an open source version of Google’s TensorFlow neural network programme.
Finlay-Smits says it is just a question of integration. ‘‘Really, 80 per cent of the work is getting the data into a format the system can use for its machine learning.’’
To train the Cacophonator, as they are now calling it, the team has installed units in a number of Banks Peninsula locations to film animals rummaging about at night.
Finlay-Smits says the system can be very dumb. At the beginning, it started to label every tree as a possum.
‘‘Early on, the model realised any time there was a tree visible in the frame, there was probably going to be a possum. It’s like a child almost. It will find whatever it is that there is in common with all the videos you’ve tagged for it to look at.’’
But quickly, as the parameters were adjusted so it only tracked small, moving, hot objects, it has learnt to distinguish possums by their signature shape and movement. Likewise stoats, weasels and rats.
Finlay-Smits says the system is also getting reliable on cats, hedgehogs and birds. A camera was set up at the Willowbank Wildlife Reserve to train it on kiwis in the nocturnal enclosure.
The Cacophonator should be completely trained by Christmas. ‘‘It’s come along much faster than we expected.’’
And DoC is keen to use it to help with the problem of how to accurately count the presence of predators in an area of land.
Currently rangers rely on chew cards – bits of hollow board baited with peanut butter. Or for mustelids like stoats, tracking tunnels where the animals crawl across an ink pad and leave paw prints on paper as they exit.
The Cacophonator would also help with live trapping. Traps mostly rely on a pest species sticking their heads inside something small and confining and often animals are just too wary.
So a large open capture trap is more effective. But weka or pukeko can blunder into them as well.
Radio transmitter versions can tell DoC staff that a trap has been triggered. But they still need to go check what they have caught.
A Cacophonator could realise it was a native bird nosing about, says Finlay-Smits. ‘‘If it sees a weka, it says don’t activate. If it sees a possum, then do.’’
So a cheap AI recognition device would already have many uses as part of Predator Free 2050.
The project’s second major goal is to develop the technology of sound lures. Traditional control measures, whether a poison pellet or trap, lure with food odours.
But Finlay-Smits says social sounds, like mating and distress calls, or food-like noises, such as the rustle of a nest of chicks or the scratching of wetas, could be used to draw animals towards a device. Even the curiosity value of a faint electronic beep might work.
And immediately, that increases a device’s range. ‘‘The maximum you’d get out of a food lure is 3 to 5m, whereas it’d be tens of metres or more out of sound,’’ he says.
So the plan is to equip the Cacophonator with the ability to play sounds from a library of recordings. Finlay-Smits says Zip and others will be running field trials around New Zealand to discover what works in the wild.
Finally, the killing bit. The ultimate goal. Finlay-Smits grins.
As project leader, he says his job is to ensure Cacophony moves ahead in sensible stages. And right now, less sexy stuff, like finding a reliable battery, is a more pressing matter for him.
But yes. They have done some Heath Robinson trials with a laser sighting system.
A heat pack – the kind you microwave for a sore back – strapped to the top of a toy remotecontrol car, makes for a good surrogate stoat. ‘‘We got the prototype to happily follow that around.’’ Fun is clearly being had in someone’s backyard.
Then he says in about six months, the plan is to start looking at adapting a paint-ball gun. Technically, nothing seems too difficult as it is all existing hardware. Someone just has to put it together as a practical system.
Of course, he realises that the creation of an ‘‘autonomous killbot’’ must instantly raise health and safety fears. Even project members make the joke – what could possibly go wrong? The Cacophony Project will eventually face regulatory hurdles.
But think about it realistically. The heat signature of a human or other large animal is going to be easy to pick out. The devices will be set 1m high, pointed down towards the ground, and only operating after dark.
‘‘Most of the risks can be mitigated. At worst, you’re only going to hit people below the knees. It might break the skin, but as long as people don’t lick themselves, they should be OK,’’ Finlay-Smits answers with a laugh.
There will be also the chance of error in hitting a native bird by mistake. Or someone’s beloved cat.
But he says compare that risk to what already happens with 1080 being rained from the sky by the truckload. The wrong things often get killed as it is.
Then there is the bang for buck factor – Ryan’s elevator pitch that these devices could be 80,000 times more cost effective.
The argument is sound lures should represent a 100-fold improvement over regular traps in terms of labour hours at least. No need to ever reset baits, or clear dead bodies from traps.
And only a hundredth of the actual number of traps would be needed if sound lures can attract animals into killing range over a distance ten times greater.
The fact the system is universal – taking out possums, rats, mustelids and feral cats all at the same time – spells another factor of four cost reduction. Then the kill rate should be nearer 100 per cent rather than the 50 per cent of conventional methods.
Tot it up and you get Ryan’s eyecatching figure.
And drones? Finlay-Smits says Cacophonators could march as a line across the landscape, clearing the bush as they go.
This would be cheaper than having to sprinkle them about. They would just sit in one spot for a week or so, then be moved along.
And a drone version would save on paying humans to do the moving. Finlay-Smits says you could have a system that even returns to base for recharging and reloading before flying off to its next location.
It sounds crazy. But modern tech is evolving so fast that you can really try just about anything you can imagine.
‘‘Because the project is of such national interest, and there’s a lot of smart technologists in New Zealand, once you start showing some success, people are going to be saying, hey, I can help you with that.’’
By its nature, the Cacophony Project could still go off in surprising directions. It is not following some exact blueprint but making it up as it goes along. The final outcome may be different from what is currently envisaged.
But Devon McLean, the ex-forest industry boss who now chairs Zip and is a director of Predator Free NZ Ltd, says Cacophony’s contribution should be significant.
‘‘It could be the game-changer for us in the end. Especially if it allows us to chase down the last few percentage of the pests.’’
McLean says the incremental approach is also looking like it can extract far more out of existing control technology than expected.
Working with DoC, Zip has been conducting landscape-scale trials on the more effective use of 1080 and trapping.
McLean says it might be thought New Zealand already must know how to do the old school stuff. However eradication programmes have often had one hand tied behind their backs by public opinion and uncertain government funding.
Knocking back the numbers was often enough. Only now is the research going into finding what kind of control regimes could lead to complete elimination of pests in an area.
Zip has been experimenting with more intense 1080 campaigns – doubling the coverage – while also taking advantage of ‘‘defensible’’ geographic features like a peninsula, or tract of bush isolated between two rivers.
McLean says they are calling it ‘‘1080 to zero’’. Once an area is cleared, it then needs to be maintained by a virtual border of traps along the boundaries or bridge crossings.
‘‘It’s more expensive. But not if it’s also your last 1080 drop needed.’’
Then Cacophony-style technology is inevitably going to slash the labour costs of even conventional trapping. DoC is talking of 90 per cent reductions.
And in the long term, McLean says genetic engineering may be added to the arsenal.
Gene-editing – adding a gene for sterility to a population to let a species breed itself to death – looks about 10 years off.
Although, as that will be something developed internationally, it is likely only to be on offer for rats and mice – the kinds of animal considered pests elsewhere.
But generally, says McLean, the Predator Free 2050 promise is looking hopeful. And smart technology is going to have a lot to do with that.