Image recognition
Naturalis Biodiversity Center has developed a special AI algorithm that is used to analyze the photos from the DIOPSIS insect camera to count, recognize and determine the insects on each photo.
Count
In the first step of the process, the algortime analyzes where in the picture the insects are. This involves distinguishing between insects and other objects in the image, such as mud, grass and shadows. Once the insects are located, their numbers can be easily counted.
The DIOPSIS camera v2 takes pictures when motion is detected on the yellow screen. The image recognition software tracks the insects that are in view. Insects that remain on the screen longer are therefore likely to be photographed multiple times, but the results clearly indicate which detections are of the same individual to avoid double counting.
Recognize
The image recognition algorithm was trained from a dataset of some 4,000 photographs on which over 4,800 insects were identified by experts. Similar to the well-known "face recognition" in humans, the algorithm looks for certain specific features, which it extracts itself from an annotated dataset. The algorithm is currently able to classify insects at the level of 7 species, 4 genera, 41 families and 11 orders. Most identifications are performed at the family level.
The algorithm also estimates the reliability of the determinations. If desired, uncertain identifications can be excluded in an ecological analysis of the results. The image recognition 'pipeline' is shown to the right.
Determine weight
By measuring the length of an insect and relating it to an existing dataset of weights of similar species, a broken power function is used to estimate the biomass of the insect. This method is most accurate when the exact species is known.
ARISE
Naturalis Biodiversity Center, in collaboration with the University of Twente, the University of Amsterdam and the Westerdijk Fungal Biodiversity Institute, is building a research infrastructure(ARISE) to identify species faster and more efficiently. Within ARISE, the data pipeline is built to send images from the DIOPSIS cameras to the supercomputer running the algorithm.
Researchers can use the ARISE Digital Species Identification (DSI) portal to view their own data, run image recognition analyses and download the results. Other users can request a license for image recognition analysis via a message to diopsis@naturalis.nl.