Bulk arthropod abundance, biomass and diversity estimation using deep learning for computer vision

Abstract

Arthropod abundance, biomass and taxonomic diversity are key metrics often used to assess the efficacy of restoration efforts. Gathering these metrics is a slow and laborious process, quantified by an expert manually sorting and weighing arthropod specimens. We present a tool to accelerate bulk arthropod classification and biomass estimates utilizing machine learning methods for computer vision. Our approach requires pre‐sorted arthropod samples to create a training dataset. We construct a dataset considering 18 terrestrial arthropod functional groups collected in southern Ontario, Canada. The dataset contains 517 high‐resolution images with approximately 20 individuals per image taken from either a petri dish or a bulk tray. Our tool uses the watershed algorithm to obtain precisely cropped individuals without any object annotations. After manually sorting cropped images of biological ‘debris’ and petri dish edges, three classifiers, DenseNet121, ResNet101 and MobileNetv2, each with trade‐offs of computational efficiency versus accuracy, are trained and compared to predict arthropod functional groups for each cropped individual. To calculate biomass, we compare seven linear and nonlinear models considering the arthropod pixel masks obtained using the watershed algorithm, in combination with images of a single function group with recorded weights, to calculate the per pixel density per functional group. From our experimentation, we recommend using DenseNet121 as it had the highest top‐1 functional group classification accuracy, likely a result of being the model with the largest number of parameters, with 86.14% considering the 20 labelled classes (18 arthropods plus debris and petri dish edge) in comparison to ResNet101 (85.10%) and MobileNetv2 (84.94%). For biomass estimation, we recommend using the average per pixel density which had the highest ranked performance considering both total error, 0.043 g (0.855% error), and cumulative class‐specific error, 1.62 g (40.67% average error across all classes), in comparison to the total ground truth biomass of 5.10 g. Our estimated Simpson's Index of Diversity was 0.9404 in comparison to the ground truth 0.9408. Our method simultaneously classifies >1,000 arthropods to functional groupings while estimating total and class specific biomass, without any computer vision bounding box or mask annotations, all from a single photo. We release our code and dataset to further research efforts in computer vision for arthropod classification.

Emerging technologies revolutionise insect ecology and monitoring

“…Computer vision can be applied to both live and dead insects to count and classify insects with less human labour and observer bias, reducing the necessity for taxonomic expertise and creating opportunities for the engagement of citizen scientists (Box 3). When applied to live insects, advantages are that the method is nondestructive and can be completely automatized, providing information on species' occurrences, abundances, individual size, biomass, and movement [22,23], as well as behaviour and interactions [15]. Imaging of dead specimens allows control of lighting conditions and minimises background variation to achieve impressive classification performance and biomass estimation, and allows independent validation of species identity [24,25].…”

Global monitoring of soil animal communities using a common methodology

“…As an alternative high-throughput approach, we adopt image analyses of mixed invertebrate community samples (Schneider et al 2021). Image analysis is based on visual animal identification and allows for direct body size estimations which can be used to estimate biomasses and energy fluxes.…”

“…Animals are extracted and photographed in mixed samples at Local hubs. Abundances and body sizes of taxonomic groups are estimated by the Central team using manual image annotations and a trained image analysis algorithm(Schneider et al 2021). Biomasses and energy fluxes are estimated using allometric regressions and food-web reconstruction approaches(Ehnes et al 2011, Sohlström et al 2018, Jochum et al 2021, Potapov 2021) and used in statistical analyses and modelling.…”

Extracting abundance information from DNA-based data

“…However, this method is difficult to apply to very large numbers of individuals and cannot be applied to trace DNA or microbial taxa. Note that this approach could also be carried out via machine-learning-accelerated visual identifications of photos of arthropods (Schneider et al 2022).…”