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 9 / 14 The near-perfect detection model (both recall and precision of 0.99) yielded on average 53 false positives (i.e., objects being actually part of the background) for 5,295 detectable fish when the 10-minutes videos were processed at 1 frame per second. There were up to 527 false positives for 52,927 detectable fish when videos were processed at 10 frames per second and 1,586 false positives for 158,789 detectable fish when processed at 30 frames per second. According to our simulation design, the number of false positives is proportional to the number of frames analyzed, which is realistic as false positives occur mostly because of textured sessile organisms (e.g., algae, corals, sponges) [39]. Therefore, we recommend using a processing frame rate of 1 frame per second for video analysis, in order to ensure all species are detected while minimizing the number of false positives. However, if fast-moving species are not present or are not in the scope of monitoring, it is possible to analyze fewer images per second. In such contexts, using a lower processing rate (e.g., 0.5 fps) does not compromise species detection, while significantly reducing the number of analyzed frames and the number of false positives, thereby reducing the proportion of error in biodiversity estimates. Fig 3. Effect of post-processing model outputs on biodiversity estimates. Total abundance, species richness and species diversity estimates as a

function of post-processing threshold for 42 automated analysis models resulting from 6 detection models, with their respective recall (“rec”) and preci- sion (“prec”) performances (rows) and 7 classification models (accuracy (“cla_acc”) as colors). Each dot represents the average over the 10 simulated videos with corresponding standard error as vertical bars. The red horizontal bar represents the ground truth. 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. 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.g003