Popular Science: New methods of insect identification
Over a million insect species are currently known. Besides that, considerable variability can often be observed within a single species due to sex, life stage or color morphing. However, its reliable identification is very important. Insects form a large component of biodiversity, while playing an important and often irreplaceable role in the functioning of individual ecosystems on the Earth. On the other hand, there are many insects that are disease vectors or are pests of agricultural crops. The identification of a species requires years of training and specialization. Therefore, there have been efforts in the past to automate and simplify the identification of insects. Unfortunately, they were often inaccurate and required a lot of input data. With the development of a new method called convolutional neural networks (CNNs), most of the problems with the automatic identification of insects were solved.
CNNs are a practical tool for image recognition, for which the need for input parameters is radically reduced. Briefly, in the convolutional part, the image is automatically preprocessed into vector form that enters the neural networks which process it, and the resultant data are classified.
Since this is a promising method, researchers have focused on determining the accuracy of automatic insect identification using convolutional neural networks. They used four different photo data sets for their research.
The first two datasets, consisting of taxonomically diverse images, were used to determine whether the automatic system can "learn" to recognize different higher groups of beetles and flies. The other two datasets were focused on the recognition of visually similar species. The first dataset contained 884 images of flies from 11 families photographed from the front, which were identified with 92.7% accuracy. Beetles (2936 different images of 14 families photographed from the dorsal) were successfully identified in 96.1% of the cases. For the other two datasets, scientists tried to test the success of the automatic recognition of species that are very similar and difficult to separate even for human experts. In this case, 97.3% were correctly recognized for three related Oxythyrea species (339 individuals, dorsal view) and 98.6% of nine species of Plecoptera larvae (3845 images from one sample from different views).
The main results of the research are highly satisfying, as they have shown that with modern technologies it is possible to identify insect species with an accuracy of more than 90%. This was achieved even on very small datasets and the results could be compared with the entomology experts’ evaluations. Using convolutional neural networks, it seems to be possible to fully automate the taxonomic identification system.
The use of convolutional neural networks in the field of biodiversity is still in its infancy, but considerable success can already be observed. The authors point out that this method could not only be used to identify existing species, but also unspecified species, or “cryptic” species, which only differ by less noticeable features and are primarily identified today by DNA.
Valan, M., Makonyi, K., Maki, A., Vondráček, D., Ronquist, F. (2019): Automated Taxonomic Identification of Insects with Expert-Level Accuracy Using Effective Feature Transfer from Convolutional Networks. Systematic Biology 68(6): 876–895.