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 4 / 14 We simulated fixed remote video surveys with a total duration of 10 minutes. This length was chosen as a trade-off between the ability to simulate contrasted temporal abundance of species and the size of simulated datasets. We considered that each video survey allows recording 15 species (out of the 30 from the regional pool) representing the 8 groups with 1–3 species, sorted randomly (Table 1). Then, the simulation of abundance distribution through time of each species was led through 4 steps. First, the average duration of a presence slot was computed as the ratio between temporal occupancy (min) and the number of presence slots. Second, the duration of each presence slot was sorted from a normal distribution with the mean being the average presence time and the standard error as the square root of the mean time. Third, the time interval between two consecu- tive presence slots was iteratively sorted from a uniform distribution, ranging from the number of remaining time slots to the total remaining time. Lastly, the number of individuals within each presence slot was sorted from a binomial distribution with the duration of the time slot, incorporating the species’ abundance. These temporal abundances were simulated with a resolution of one second, which corresponds to the unit of mea- surement usually used during diver counts. They were reduced to 30 frames per second to simulate video recording at the site studied.

To demonstrate the robustness of our results, 10 video surveys were simulated to account for variability in species abundance through time. The number of replicates was limited to 10 to balance computational cost with reproducibility. Simulation of video processing with detection and identification algorithms For each video, we simulated an automated processing with a detection algorithm and then a classification algorithm with varying performances. For the detection model, we used recall to assess false negatives, ensuring that all present individuals are detected, and precision to evaluate false positives, minimizing incorrect detections. For the classification model, we relied on accuracy to measure the proportion of correctly classified images, providing an overall assessment of classification performance. This approach follows standard evaluation practices, where recall and precision are essential for detection tasks, while accuracy is commonly used for classification. We first varied the processing rate of video processing with 7 levels from 0.25, 0.5, 1, 2, 5, 10, and 30 frames per second. The two first processing rates mean 1 image processed every 4 or 2 seconds, respectively. The last level is the processing of all the frames from a standard video recording. The subsampling consisted in keeping the first frame of each second, then evenly choosing among the next ones depending on the target processing rate. The first step of the automated analysis was the detection of all individuals using a single-class detection algorithm.

The recall measures the ability of a model to identify all relevant positive instances. The recall score of the algorithm varied from 0.60 (similar to the least efficient algorithm published) to 0.99 (i.e., hypothetically quasi-perfect algorithm) with 5 intermediate levels (0.70, 0.80, 0.85, 0.90, 0.95). To simulate this, each individual present in the image has a detection Table 1. Values used to describe the groups of species considered in this study. group abundance relative temporal occupancy (%) number of time slots number of species in the regional pool number of species in each 10-min video 1 20 20 5 3 1 2 2 90 15 3 1 3 2 30 15 3 2 4 8 10 1 3 2 5 8 5 1 3 2 6 1 5 1 5 2 7 1 1 1 5 3 8 1 0.5 1 5 2 https://doi.org/10.1371/journal.pone.0327577.t001