High-resolution density assessment assisted by deep learning of Dendrophyllia cornigera (Lamarck, 1816) and Phakellia ventilabrum (Linnaeus, 1767) in rocky circalittoral shelf of Bay of Biscay

Abstract

This study presents a novel approach to high-resolution density distribution mapping of two key species of the 1170 “Reefs” habitat, Dendrophyllia cornigera and Phakellia ventilabrum, in the Bay of Biscay using deep learning models. The main objective of this study was to establish a pipeline based on deep learning models to extract species density data from raw images obtained by a remotely operated towed vehicle (ROTV). Different object detection models were evaluated and compared in various shelf zones at the head of submarine canyon systems using metrics such as precision, recall, and F1 score. The best-performing model, YOLOv8, was selected for generating density maps of the two species at a high spatial resolution. The study also generated synthetic images to augment the training data and assess the generalization capacity of the models. The proposed approach provides a cost-effective and non-invasive method for monitoring and assessing the status of these important reef-building species and their habitats. The results have important implications for the management and protection of the 1170 habitat in Spain and other marine ecosystems worldwide. These results highlight the potential of deep learning to improve efficiency and accuracy in monitoring vulnerable marine ecosystems, allowing informed decisions to be made that can have a positive impact on marine conservation.

New interactive machine learning tool for marine image analysis

“…Alternatively, the convolutional neural network termed 'You Only Look Once' (YOLO) [79] is also capable of automated segmentation and is becoming increasingly user-friendly through the production of thorough GoogleColab notebooks [80,81] and the development of accompanying graphical user interfaces [82,83]. Similarly to RootPainter, YOLO can automatically apply data augmentation to enhance model performance [44], so non-trivial user-controlled image pre-processing is not required [84]. Users are encouraged to begin training using one of YOLO's pre-trained models, but de novo development is possible.…”

“…The capability of YOLO to automatically detect objects from marine images has been shown; YOLO version 4 [85] was used to develop a model to identify the Xenophyophore, Syringammina fragilissima (Brady, 1883), within 58 000 AUV video frames, requiring less than 10 days for complete analysis, and achieving a final precision of 0.91 and recall of 0.84 [86]. Additionally, the recently released YOLO version 8 [87] has been used to develop a model to simultaneously quantify the coral Dendrophyllia cornigera [44] and sponge Phakellia ventilabrum [44] within 5201 transect images [44]. The fully trained YOLOv8 model required just over 2 h to process the data, with detection metrics depending on species and study sight, but all surpassing 0.85 [44].…”

“…Additionally, the recently released YOLO version 8 [87] has been used to develop a model to simultaneously quantify the coral Dendrophyllia cornigera [44] and sponge Phakellia ventilabrum [44] within 5201 transect images [44]. The fully trained YOLOv8 model required just over 2 h to process the data, with detection metrics depending on species and study sight, but all surpassing 0.85 [44]. As the multiclass annotation and transfer learning capabilities of YOLO (version 4) have also been demonstrated [88][89][90], model sharing and optimization with this algorithm may serve to ease the image analysis bottleneck for future studies of marine species' distributions and biodiversity.…”

“…This may be reflected in the smaller standard deviation for the average area error of Mini (0.45 ± 0.86 cm 2 ) as predicted by Model 2 within smaller images, than for Magnus (−0.06 ± 2.87 cm 2 ) as predicted by Model 1 . Training strategies may, therefore, require adaptation for datasets with significant class imbalances to ensure adequate model performance is achieved [44,94].…”

“…However, this incurs inherent difficulties depending on varied species presence and visual complexity within datasets [43]. While the latter does not appear to have hindered benthic object detection models [4,35,37,44], there has been limited success with models capable of simultaneously, and differentially, extracting the area of multiple species [10,45]. When extracting deep-sea coral and sponge areas, Purser et al found that the large variation in texture and colour of sponges at the site limited model performance; this may have been improved through increased exposure of the algorithm to sponges within the training dataset [10].…”