What do an oil spill, invasive species and routine fish stock surveys have in common? They all have to use an expert to manually identify and count the species involved. Changes occur in plankton distributions following an oil spill, invasive species introduced through discharge of ship’s ballast water can seriously impact on ecosystems and fish larval identification must be carried out for routine stock assessments of a fishery. These are all examples where the necessary expertise requires years of training before competence is achieved.

  There has been a serious decline in scientists who have this biological expertise over the last 50 years. Training of taxonomists and specialized training programs has been cut back in Universities and Research Laboratories^[1]. As a result there are only about 4,000–6,000 professional taxonomists worldwide and the demand for skilled analysts is escalating. Therefore, there exists a bottleneck^[2] in the identification of natural organisms due to the lack of qualified taxonomists and analysts across a very wide range of taxa from agriculture, marine sciences, aquaculture and border control inspection to paleontological analysis of oil-bearing rocks.

  Computer vision and artificial intelligence offers a way of reducing the burden on these skilled analysts. Advances over the past decade in computer-based identification mean that some classes of identification are possible without human operator intervention in routine analysis. The steps involved are broadly: first obtain a set of specimen images that can be used for training and testing purposes (an expert must identify and label each specimen with a high degree of accuracy). Then use computer vision techniques to extract meaningful features from the set of images. Then by using machine learning techniques, patterns in the data can stored and used for later identification of previously unseen images. Performance on training data and test sets can approach 90%, when individual specimens are present in an image form. Performance in cluttered scenes is often poor, but progress is being made.

  Typically objects are detected in each image through a monochrome conversion and features processed for numerical descriptors of invariant moments, contour shapes and grey level histograms. A numerical vector containing all the descriptors is then assembled for each specimen image. Sets of these are fed into machine learning methods such as support vector machine^[3] and random forest^[4]. All current methods require at least 40 examples of each class. Repeated randomized training continues until a minimum learning performance is achieved on both the test and the training set. The trained classifier is then evaluated on previously unseen data and applied to field collected specimens for example^[5-17]. A serious shortcoming of current automatic methods is the poor performance at identifying rare organisms.

  It is important that the biological oceanography community embrace these new techniques for specimen identification, in the same way that the geophysical oceanographers have welcomed instrumentation, and benefitted from widespread acceptance of Conductivity Temperature and Depth (CTD) measurements that have revolutionized our understanding of the physical dynamics of our seas and oceans. However, there is much resistance to new methods in the biological and ecological community, as these new tools are not refined enough to replace plankton analysts per se, but they can assist people to increase analytical throughput.

  What is often not understood is that humans suffer from cognitive shortcomings that limit their performance in identification tasks. A number of studies led by Culverhouse and others^{[18-24,26-28]} reveal that plankton identification by experts varies from better than 90% to worse than chance, according to the cognitive biases extent at the time of the activity. Boredom and exhaustion are significant factors affecting performance, so too positivity bias and recency effects^[25]. Day to day repeatability is also in question, with 14 out of 21 experts in one study^[26] performed as experts, being self consistent to better than 90%, the remainder ranged from 80% down to less than 20% repeatability. Experts often work alone, with no quality assurance or cross checking. At the present time, machine performance is subject to closer scrutiny than expert performance.

References