Impact of deep learning and post-processing algorithms performances on biodiversity metrics assessed on videos

PLOS One | https://doi.org/10.1371/journal.pone.0327577 August 11, 2025 7 / 14 threshold of 0.8, the minimum augmentation was 0.003, going from 0.791 to 0.794 for a model with a recall of 0.6, a pre- cision of 0.95 and a classification accuracy of 0.99. With a post-processing threshold of 0.99, the maximum augmentation was 0.24, going from 0.61 to 0.85 for a model with a recall of 0.99, a precision of 0.6 and a classification accuracy of 0.6. Although only 10 simulations were performed, the results were highly consistent across replicates, as indicated by the very low standard deviations around the mean values shown in the figures. Discussion In our simulations, the scarcest species were not always detectable by the models processing an image every 2 seconds (processing frame rate = 0.5) or an image every 4 seconds (processing frame rate = 0.25). Such species moving briefly across the camera field-of-view, such as fast-swimming predators (tuna, trevally) in the marine environment contribute Fig 1. Effects of processing rate and algorithms performance on estimates of biodiversity. Total abundance, species richness and species diversity estimates as a function of video processing rate (fps, frame per second) for 42 automated analysis models resulting from 6 detection models, with their respective recall (“rec”) and precision (“prec”) performances (rows) and 7 classification models (accuracy (“cla_acc”) as colors). Each dot rep-

resents the average over the 10 simulated videos with corresponding standard error as vertical bars. The red horizontal bar represents the ground truth (i.e., diversity facet present on each video). Total abundance is the sum of the maxN (maximum number of individuals of a species seen in one image) of each species, richness is the number of species seen, and species diversity is the exponential of the Shannon entropy index. https://doi.org/10.1371/journal.pone.0327577.g001