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 5 / 14 probability calculated using a Bernoulli distribution with recall as a probability. False negatives, i.e., undetected individuals, are not treated in the rest of the simulations as will be the case in real-life pipelines. The precision is the proportion of correctly labeled individuals among all those detected as present. The precision score of the algorithm also varied from 0.60 to 0.99 with 5 intermediate levels (0.70, 0.80, 0.85, 0.90, 0.95). We determined the number of false positives using the precision rate and the number of detections, then randomly added these false positives across all frames. We acknowledge that in real-world applications, false positives could be clustered, typically caused by recurring background features that resemble target objects. Second, a species-level classification algorithm was simulated for each bounding box from the detection algo- rithm. The accuracy is the proportion of correctly labeled individuals among all individuals of this class. The algorithm had an accuracy score ranging from 0.60 to 0.99 with 5 intermediate levels (0.80, 0.70, 0.85, 0.90, 0.95). We thus simulated misidentification at random using 1-accuracy as a probability of misidentification and the identity of the erroneous species was randomly chosen among the remaining species from the regional pool (i.e., 29 species). We hence assume that the misidentifications were even across species (i.e., there was no species pair more likely to be

confounded to each other). We acknowledge that in real-world applications misidentifications are often not random with on average lower accuracy for the classes with the fewer images in the train set and across classes with similar visible features. Third, we implemented a post-processing step inspired by the confidence threshold approach of Villon (2020) that transfers some identifications (expected to be mostly misidentifications) to an unsure class. For each misidentification, there was a probability p (5 levels: 0.80, 0.85, 0.90, 0.95, 0.99) that it was eventually considered as “unsure”. Meanwhile, for each good identification, there was a probability p/20 that it was eventually considered as “unsure”. We also analyze results from detection and identification steps without this post-processing. We thus simulated a total of 14,406 automated processing pipelines (7 processing rates * 7 detection recall * 7 detec- tion precision * 7 classification accuracy * 6 post-processing thresholds) for each of the 10 simulated videos. For each processing pipeline, we eventually computed the abundance of each species as the maximum number of individuals present in a single frame (i.e., MaxN metric commonly used in video-based surveys). Assessing the performance of automated processing to estimate biodiversity metrics As a preliminary analysis, we checked whether all individuals were detectable (i.e., present on at least one of the images) on the frames kept for analysis with each processing rate.

We assessed the influence of algorithms’ performance on three biodiversity metrics frequently used in the monitoring of ecosystems. First, we computed total abundance as the sum of species abundances. Second, we computed species richness as the number of species seen along the video (i.e., at least one individual in at least one frame). Third, we computed species diversity using the Hill number version of the Shannon entropy [35]: exp ( – ∑S i=1 pi × log (pi) ) , where S is the total num- ber of species present and pi is the abundance of species i relative to the total number of individuals. Species diversity increases with increasing number of species and increasing evenness of their abundances. These 3 indices were computed for data from ground truth (simulated communities) and for each output of each pro- cessing pipeline (species abundances provided by detection, classification, and post-processing algorithms). To further investigate the outputs of the processing pipeline we computed Jaccard and Bray-Curtis similarity indices between species composition or abundance from each processing pipeline and species composition or abundances from ground truth. The Jaccard index [36] quantifies the similarity between two sets of species by dividing the number of shared species by the total number of distinct species across both sets. It ranges from 0 (no species shared) to 1 (all species shared).