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 6 / 14 The Bray-Curtis similarity [37], quantifies differences between two sets of species abundances. it is defined as follows: 2×∑p i=1 min(Nij,N ik) ∑n i=1(Nij+Nik) , where N ij is the number of individuals of species i at site j,N ik is the number of individuals of species i at site k, and p is the total number of species in the samples. It ranges from 0 (sets have different species composition or abundance markedly differ for shared species) to 1 (all species have the same abundance in both sets). Results All 15 species simulated on each video were detectable (i.e., present on at least 1 of the images) for analyses with a pro- cessing rate of 1 image per second or higher. For the 2 lowest processing rates (0.25 and 0.5 images per second), 14.3 (sd = 0.67) and 14.8 (sd = 0.42) species on average were detectable, respectively (S1 Table). For total abundance, the ground truth was 65 individuals present in each simulated community. Estimated abundance increased with the frame processing rate for all combinations of deep learning model performance, reaching a maximum of 187 individuals for a detection model with a recall of 0.99 and a precision of 0.6 (fifth row on Fig 1), and an identification model with an accuracy of 0.6 (red dot on Fig 1). For all processing pipelines for which the processing frame rate was 2 and above, the estimated number of species was 30, which corresponds to the number of species learned by the classification model. For lower processing rates (0.25,

0.5 or 1), estimated species richness was lower than 30 and even reached 16 for a recall of 0.6, a precision of 0.99 and an accuracy of 0.99. Species diversity was 8.71 for the ground truth. For all combinations of model parameters, the estimates of this metric were all above the ground truth and increased with the processing frame rate. It was at a minimum of 10.0 for a detector with a recall of 0.6 and a precision of 0.99 (second row on Fig 1) and a classification model with an accuracy of 0.99 (pink dot on Fig 1) for a processing frame rate of 0.25. Estimated diversity even reached a maximum of 27.8 for a detector with a recall of 0.6 and a precision of 0.6 (top row on Fig 1) and a classification model with an accuracy of 0.6 (red dot on Fig 1) for a processing frame rate of 30. The Jaccard similarity index between actual species composition and those estimated after automated processing was constant at 0.5 except for 36 models with a processing rate of 0.25 or 0.5 frames per second, for which the Jaccard simi- larity was between 0.32 and 0.49 (S2 Fig). The Bray Curtis similarity between actual abundances and those estimated after automated processing decreased with increasing processing rate from 0.72 for 1 frame processed by second to 0.64 for 30 frames processed by second (Fig 2). Given these results, we focus hereafter on the effects of algorithm performance for a processing rate of 1 image per second to avoid missing out on rare species. For total abundance, using a 0.99 confidence threshold post-processing, it was possible to get closer to the ground

truth for all the detection models. A detection model with a recall of 0.6 and a precision of 0.6 (top row on Fig 3) estimated 7% more abundance than the ground truth, regardless of the accuracy of the classification model (colors of dots on Fig 3), when the post-processing threshold was between 0.8 and 0.95. For a detection model with a precision of 0.6 (first and third rows on Fig 3) the post-processing threshold had no impact on the richness for a threshold under 0.95. Setting threshold above 0.99 led the species richness being closer to the ground truth [15], averaging 23.9 for a threshold of 0.99. For detection models with a precision of 0.99, the impact of threshold post-processing was greater when the accuracy of the classification model was close to 1 (Fig 3). The Jaccard similarity between actual species composition and those estimated after automated post-processing decreased regardless of the classification model performance (S3 Fig). With a post-processing threshold of 0.99, the max- imum increase was 0.45, going from 0.51 to 0.96 for a model with a recall of 0.99, a precision of 0.99, and a classification accuracy of 0.99. Using a post-processing threshold increased the Bray-Curtis similarity (Fig 4) between actual abundances and those estimated after automated processing for all performances of detection or classification models. With a post-processing