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 2 / 14 biodiversity [1]. Such efficient monitoring requires consistently tracking the presence and abundance of species at large spatial scales, in various habitats and frequently to assess ecosystem changes accurately [2]. This challenge is particularly critical in underwater environments, where conducting biodiversity surveys is logistically complex and resource-intensive. Coral reefs, which alone host nearly one-fifth of all marine life on Earth [3], are among the most threatened ecosystems, making it essential to develop efficient monitoring techniques to assess and protect their biodiversity effectively. Various community assessment methods are convention- ally used but none of them are yet satisfactory. Sampling-based approaches such as fishing and trawling observations yield valuable data, but they often come with ecological impacts [4]. Among non-invasive methods, eDNA allows censusing a high number of species (including elusive or cryptic ones) but it does not provide esti- mates of species abundance [5]. Acoustic based approaches allow long term moni- toring and development of ecosystem health proxies, but do not census all species and do not accurately estimate abundance [6]. Visual counting methods performed by scuba divers are reference approaches to survey community composition but are notoriously time-consuming and require highly trained experts [7]. In response to these challenges, the use of video-based surveys has emerged as a promising

solution [8] with diverse approaches depending on the animals and ecosystems studied with underwater camera traps [9], baited underwater stereo-cameras for marine predators [10,11], long-duration remote underwater cameras for reef fishes [12,13], autonomous underwater vehicles for reef fishes [14], and unmanned aerial vehicle for marine megafauna [15,16]. These image recorders allow for accurate and non-invasive monitoring of marine and terrestrial ecosystems and have thus been increasingly used for the last decade [8]. However, the on-screen analysis of recorded videos by experts remains a significant obstacle to scalability. The detec- tion and identification of all targeted species are indeed demanding tasks even for highly-trained experts [17,18]. Meanwhile, the rise of the use of image recording and artificial intelligence applied to image processing have made significant progress during the last decade, mostly through the progress of deep learning algorithms [19]. Today, two of the main families of deep learning models applied to images are used: detection models and classifi- cation models. Detection models are used to find all objects of interest in an image [20]. Classification models are used to associate each image with a label [21]. While multi-class detection models do both tasks simultaneously [22,23], training them effi- ciently remains a challenge when most classes are rare, which is the case for most species assemblages [24–26]. Deep learning algorithms can identify all animals on

images considerably faster than humans [27]. However, detection algorithms produce both false negatives (individuals that are not detected) and false positives (regions of image without any individuals but detected as containing one). In addition, identifi- cation algorithms may assign the wrong species to the correctly detected individuals. Finally, all false positives from the detection step, yield a misidentification as clas- sifiers always attribute one of the learned classes to an input object. One approach to limit model errors consists in thresholding [28], i.e., setting aside raw outputs and reproduction in any medium, provided the original author and source are credited. Data availability statement: All data are fully available. Research data are available at https:// doi.org/10.5281/zenodo.15519964. Code is available at https://github.com/valentine-fleure/ deep_perf. Funding: V. F. was supported by the Association Beauval Nature (CIFRE 2022/0127). K. P was supported by the IA-Biodiv ANR project FISH- PREDICT (ANR-21-AAFI-0001-01). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist.