species identification, neural network, citizen science, Odonata

Dragonflies and damselflies are suitable animals for bioindication, also aided by their easy recognition based on a detailed description or photograph. Thus, it is beneficial to involve the public through citizen science to obtain data on their occurrence. However, to identify the species or even the sex of individuals correctly, amateurs need help and automatic identification proves to be a promising mechanism. The Dragonfly Hunter CZ application helps with species identification using an algorithm based on environmental and expert data. However, this algorithm needs to be improved with additional knowledge of specialist odonatologists. Therefore, we attempted to enhance the algorithm with typical characteristics of individual species or groups of species that would improve the identification of the species and their sex. Using a testing dataset containing 50,000+ records, subjects were classified into 148 classes (74 Odonata species in Czechia), based on features such as colour etc. The basic model orders all classes for each subject by relevance, with the correct class being on average at 12.27th place. To further improve the accuracy, a model for dynamic questions was developed that selects 5 additional questions (from 17 available) by their influence on results after a basic classification is made, which yields an average correct class place to be 7.62. Even more promising improvement in the identification can be achieved by image recognition. Based on 22,000+ photographs of individual Czech species, we were able to increase the accuracy of species recognition to the sex level substantially, using a neural network with the SEResnet-50 architecture. The probability of correct identification in the first position of the resulting “species identification list” was 84%, within the first three positions we reached 95% and within the first five positions we even reached an accuracy of 99.89%. ...