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 12 / 14 accuracy, particularly for rare classes. However, this comes at the cost of reduced recall, which may lead to nondetection of some rare species, a trade-off that should be carefully considered depending on the monitoring goal. Findings from our simulations involving a succession of two models — detection and classification — are also valid with a detection model alone or a classification model alone. A detection model alone serves as a processing pipeline where an expert performs classification with near-perfect accuracy (0.99), while a classification model alone represents detection by a human expert with near-perfect precision and recall (both 0.99). Multiclass detection models [23] provide an alterna- tive to a single-class detection followed by a classification approach [22], sharing an overall recall for the detection and the classification and species-specific accuracy, thus aligning with our approach and conclusions. However, post-processing these models is challenging due to difficulties in independently training them twice to establish confidence thresholds for all species, especially rare ones with limited images [29]. Moreover, classification accuracy in multiclass detectors is gen- erally lower compared to separate-step pipelines, which reinforces current recommendations to separate detection and classification tasks [26]. In conclusion, selecting a processing rate of 1 image per second optimizes the trade-off of minimizing false positives

and subsequent misidentifications while ensuring the detection of rare species. Using a detection model with a high recall then applying a post-processing on outputs of the model classification using a strict confidence threshold eventually allows faithful estimates of key biodiversity metrics. Overall, these results demonstrate that automated processing of video with deep learning models having high, although not perfect performances, will ease the assessment of disturbance gradients and the benefits of protection on biodiversity. Supporting information S1 Fig. Example of the distribution of species abundances within a video. The number of individuals of each spe- cies is illustrated with shades of green (white slots indicate absence of species). (TIF) S1 Table. Number of detectable species according to the processing rate of videos. Values are the number of sim- ulations (out of 10) yielding 12, 13, 14, or 15 detectable species (i.e., at least 1 individual present on at least one frame). For each of the 7 processing rates (frames per second). (DOCX) S2 Fig. Effect of video processing rate and algorithms performance on the estimation of species composition of simulated communities. Estimation of the species composition as a function of video processing rate (fps, frame per second) for 45 automated analysis models resulting from 9 detection models, with their respective recall (“det_recall” – columns) and precision (“det_prec” – rows) performances and 5 classification models (accuracy (“cla_acc”) as colors).

Each dot represents the average over the 10 simulations with corresponding standard error as vertical bars. The Jaccard similarity index was used to measure the difference in composition of species present on the videos and the composition estimated after automated processing using algorithms. (TIF) S3 Fig. Effect of post-processing threshold and algorithms performance on the species composition estima- tion. Estimation of the species composition as a function of post-processing threshold for 45 automated analysis models resulting from 9 detection models, with their respective 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. Each panel gathers results for a detec- tion 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