This article is reproduced with the permission of New Scientist for exclusive use by Nova users.

Species-spotting for beginners
24 September 2005
From New Scientist Print Edition.
James Randerson

"I feel like I am standing in a field in the 11th century, trying to show a ploughman that literacy would be of value to him."

So says an exasperated Daniel Janzen, a tropical ecologist at the University of Pennsylvania in Philadelphia, trying desperately to communicate his grand vision for revolutionising the way people interact with nature. He wants to give every one of us a device for identifying any living thing we find, which he hopes will lead to a new age of universal "bioliteracy".

He is not alone. Taxonomy, the science of naming and describing living things, is undergoing a transformation. In the era of expensive genome projects and flashy genetic engineering, the meticulous, eye-crossing job of cataloguing and describing the planet's wonderful natural beauty has found itself out of favour. The number of taxonomic experts is in seemingly terminal decline.

And yet we need them more than ever. Biologists have named around 1.8 million of the species on Earth, but between 75 and 90 per cent still remain to be logged - and that's not including bacteria.

Even when a species has already been described, getting hold of someone who can identify a specimen is difficult. A shipment of fruit at US customs can sit in quarantine for three days while customs officers send a suspicious insect that might infest local crops to the Smithsonian Institution in Washington DC for identification.

Now the species-namers are turning to technology to revitalise their field. "This is taxonomy's finest hour," says Quentin Wheeler of the Natural History Museum in London. He thinks that scientists are in a race against time to chart biodiversity before humans kill most of it off. "We are the last generation with the opportunity to investigate the tree of life," he says.

The vision is to create a universal "life-reader" for on-the-spot identification. This will have myriad practical uses. With the ability to identify any species you find, you can plug those two Latin words into the internet and access a wealth of knowledge about the species. A reader that can scan DNA would allow shoppers to check whether processed meat really is what it says on the packet, or tell nut-allergy sufferers whether their meal is safe. And image-processing software could help birdwatchers instantly identify what they are looking at, for example. It would also help geologists identify the plankton in sediments that can guide them to oil deposits.

To get off the ground, the life-reader will have to be practical. One prototype system is DAISY, or Digital Automated Identification System. Its software compares images of the animal or plant you want to identify with sets of training pictures of known species.

DAISY transforms the images into less detailed versions, shrinking them from 500-by-500 pixels to 32-by-32. By massively reducing the extraneous information in each picture, it speeds up the identification process. "It is a lot easier to process the downsized images," says Stig Walsh of the Natural History Museum.

At the same time it conducts a rearrangement of pixels called a "polar transformation". Rather than using x and y coordinates, each pixel is represented as a distance from a central point combined with an angle. This effectively downgrades information about the subject's outline shape and makes the identification software concentrate on the object's pattern. Although shrinking and transforming the image throws out lots of information there is still enough there to tell species apart.

Getting it wrong

DAISY's training sets comprise as few as 10 to 20 images per species, although closely related or physically similar species might require 100 images for DAISY to tell them apart. The images are fed into a neural network algorithm which learns how to put them into species groups.

So far, Mark O'Neill, one of DAISY's developers at the University of Newcastle, UK, has tested it out on insects, plants, fish, pollen and even human faces. Challenged to identify specimens from Britain's 60 butterfly species, DAISY gets it right 92 per cent of the time. That might not sound great, particularly to a doctor or a customs officer for whom mistakes could be disastrous, but would a person do any better?

Phil Culverhouse of the Centre for Interactive Intelligent Systems at the University of Plymouth in the UK set 16 experts in marine plankton taxonomy and general marine ecologists the task of identifying 310 images of a group of marine plankton. The taxonomists agreed with each other on their assessment of the specimens in 83 per cent of the identifications, and the ecologists agreed on 72 per cent. "Physicists and chemists would look at that and say 'hmm, there appears to be a problem'," says Norman MacLeod, keeper of palaeontology at the Natural History Museum.

Add to that the fact that it is generally difficult and expensive to retrain people, for example to recognise a different group of species, and software begins to look very attractive.

Another advantage of a software-based system is that it could collate the assessments of various experts, rather than relying on one opinion. Culverhouse is embarking on a web-based study in which experts can vote on what species they think a specimen belongs to. By using this consensus opinion as the training data rather than the opinion of one taxonomist, an automatic system could effectively use the combined wisdom of the taxonomic community all at the same time.

Another prototype system is SPIDA (Species Identified Automatically), which has been tested on a family of 121 species of - you've guessed it - Australasian spiders. SPIDA, developed by Kimberly Russell and colleagues at the American Museum of Natural History in New York, uses a neural network algorithm to learn from a training set of anatomical pictures.

Before being fed into the neural network, the images to be identified are simplified by a process called wavelet transformation. This reduces the image to a coarser version of itself, eliminating unnecessary details such as the spine and hairs, while preserving the overall shape.

Another simplification, which DAISY and SPIDA share, is that they are effectively told roughly what it is they are looking at, such as the species group to which the creature belongs. Developing a visual system to recognise any creature from scratch would be next to impossible, says Russell. "Creating a system to identify all life, unsorted, would be messy, inefficient and difficult to manage."

One major problem for image-based systems is working out what is animal and what is background. The DAISY system gets around this by asking you to first outline the object of interest using a computer mouse, so that it can ignore the rest of the image. But that would be no good for a birder watching a finch through binoculars, for example.

An alternative solution investigated by David Chesmore in the department of electronics at the University of York, UK, is to identify animals by the sounds they make. He taught his system to recognise the grasshopper noises in a Yorkshire nature reserve. But cars, singing birds, aircraft and even wind all confused the system. So he programmed these in as separate "species". He thinks there could be a market among birdwatchers for a system that identifies birds automatically from their song.

But probably the most broadly applicable system would be a device to analyse DNA. That would allow you to identify your species long after it has been turned into pâté, or even if it were too small to see (see "Barcoding life"). With sequencing becoming faster, cheaper and simpler, Janzen thinks DNA barcoders will be available to everyone in less than five years.

He hopes that will engender a renewed respect for nature. "The ability to 'read life' will change forever the way humans interrelate with wild biodiversity."

From issue 2518 of New Scientist magazine, 24 September 2005, page 28

Barcoding life The best bet for a universal "life-reader" will probably be something that analyses DNA. And although the technology for sequencing DNA is not yet small enough or cheap enough for mass use, it is rapidly heading that way. Importantly, it would not be necessary to sequence a whole genome to recognise a creature. The Consortium for the Barcode of Life (CBOL), a group of zoos, museums, research and government organisations from 27 countries, has set itself the task of making DNA identification practical by finding a region of the genome that can act as a barcode. The sequence they are using is a stretch of the gene cytochrome c oxidase subunit 1, which is found in all eukaryotes. The 648-letter sequence is ideal because it has enough variation between species to tell them apart (New Scientist, 26 June 2004, p 32). CBOL has set out to barcode most of the 28,000 fish species and 10,000 bird species over the next five years. Other groups of organisms will follow. Meanwhile, the Global Biodiversity Information Facility aims to make existing knowledge about species identification more widely available. The plan is to allow electronic access to the 1.5 to 2 billion museum specimens that underpin taxonomic expertise by putting images and data on the web.

For the latest from New Scientiist visit www.newscientist.com