Dyrmann, M. | bioRxiv Science

bioRxiv2023

Hierarchical Classification of Insects with Multitask Learning and Anomaly Detection

Cameras and computer vision are revolutionising the study of insects, creating new research opportunities within agriculture, epidemiology, evolution, ecology and monitoring of biodiversity. However, a major challenge is the diversity of insects and close resemblances of many species combined with computer vision are often not sufficient to classify large numbers of insect species, which sometimes cannot be identified at the species level. Here, we present an algorithm to hierarchically classify insects from images, leveraging a simple taxonomy to (1) classify specimens across multiple taxonomic ranks simultaneously, and (2) highlight the lowest rank at which a reliable classification can be reached. Specifically, we propose multitask learning, a loss function incorporating class dependency at each taxonomic rank, and anomaly detection based on outlier analysis for quantification of uncertainty. First, we compile a dataset of 41,731 images of insects, combining images from time-lapse monitoring of floral scenes with images from the Global Biodiversity Information Facility (GBIF). Second, we adapt state-of-the-art convolutional neural networks, ResNet and EfficientNet, for the hierarchical classification of insects belonging to three orders, five families and nine species. Third, we assess model generalization for 11 species unseen by the trained models. Here, anomaly detection is used to predict the higher rank of the species not present in the training set. We found that incorporating a simple taxonomy into our model increased accuracy at higher taxonomic ranks. As expected, our algorithm correctly classified new insect species at higher taxonomic ranks, while classification was uncertain at lower taxonomic ranks. Anomaly detection can effectively flag novel taxa that are visually distinct from species in the training data. However, five novel taxa were consistently mistaken for visually similar species in the training data. Above all, we have demonstrated a practical approach to hierarchical classification based on species taxonomy and uncertainty during automated in situ monitoring of live insects. Our method is simple and versatile and could be implemented to classify a wide range of insects as well as other organisms.

ecology↗

bioRxiv2022

Accurate detection and identification of insects from camera trap images with deep learning

Reported insect declines have dramatically increased the global demand for standardized insect monitoring data. Image-based monitoring can generate such data cost-efficiently and non-invasively. However, extracting ecological data from images is more challenging for insects than for vertebrates because of their small size and great diversity. Deep learning facilitates fast and accurate insect detection and identification, but the lack of training data for coveted deep learning models is a major obstacle for their application. We present a large annotated image dataset of functionally important insect taxa. The primary dataset consists of 29,960 annotated insects representing nine taxa including bees, hoverflies, butterflies and beetles across more than two million images recorded with ten time-lapse cameras mounted over flowers during the summer of 2019. The insect image dataset was extracted using an iterative approach: First, a preliminary detection model identified candidate insects. Second, candidate insects were manually screened by users of an online citizen science platform. Finally, all annotations were quality checked by experts. We used the dataset to train and compare the performance of selected You Only Look Once (YOLO) deep learning algorithms. We show that these models detect and classify small insects in complex scenes with unprecedented accuracy. The best performing YOLOv5 model consistently identifies nine dominant insect species that play important roles in pollination and pest control across Europe. The model reached an average precision of 92.7% and recall of 93.8 % in detection and classification across species. Importantly, when presented with uncommon or unclear insects not seen during training, our model detects 80% of individuals and usually interprets them as closely related species. This is a useful property to (1) detect rare insects for which training data are absent, and (2) generate new training data to correctly identify those insects in future. Our camera system, dataset and deep learning framework show promising results in non-destructive monitoring of insects. Furthermore, resulting data are useful to quantify phenology, abundance, and foraging behaviour of flower-visiting insects. Above all, this dataset represents a critical first benchmark for future development and evaluation of deep learning models for insect detection and identification.

ecology↗