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 10 / 14 All 30 species present in the environment were always detected and identified when analyzing 1 frame per second, regardless of the recall and precision scores of the detection model, and the accuracy of the classification model. There were hence 15 false positive species, arising from classification errors, including those from false positives from the detection model. However, even with a near-perfect detection model (recall and precision = 0.99) and a near-perfect classification model (accuracy = 0.99), there were on average 29.4 (sd = 0.7) species identified at the end of the process- ing pipeline. With a near-perfect classification model (accuracy = 0.99), the errors were mostly due to false positives in the detection models. Indeed, background objects erroneously detected as objects of interest only lead to a misclassification, likely evenly picked among the list of putative species (i.e., present in the training set of the classification algorithm). In our simulations, we considered that errors in the confusion matrix were distributed randomly. In real cases, there are some Fig 4. Effect of post-processing threshold and algorithms performance on the communities abundance-structure estimation. Estimation of abundance-structure as a function of post-processing threshold for 45 automated analysis models resulting from 9 detection models, with their respec-

tive recall (“det_recall” – columns) and precision (“det_prec” – rows) performances and 5 classification models (accuracy (“cla_acc”) as colors) for 1 frame per second processing rate. Each dot represents the average over the 10 simulations and the error bars are shown. The Bray-curtis similarity measures the difference in species abundances estimated after automated processing using algorithms and species abundances actually visible on videos. Each panel gathers results for a detection algorithm (recall in column and precision in row) and a classification algorithm (color corresponding to its accuracy). Post-processing applies a confidence threshold to outputs of the identification models, discarding those with the lowest confidence scores to minimize misidentifications. https://doi.org/10.1371/journal.pone.0327577.g004