Recent advancements in cabled ocean observatories have increased the quality and prevalence of underwater videos; this data enables the extraction of high-level biologically relevant information such as species' behaviours. Despite this increase in capability, most modern methods for the automatic interpretation of underwater videos focus only on the detection and counting organisms. We propose an efficient computer vision- and deep learning-based method for the detection of biological behaviours in videos. TempNet uses an encoder bridge and residual blocks to maintain model performance with a two-staged, spatial, then temporal, encoder. TempNet also presents temporal attention during spatial encoding as well as Wavelet Down-Sampling pre-processing to improve model accuracy. Although our system is designed for applications to diverse fish behaviours (i.e, is generic), we demonstrate its application to the detection of sablefish (Anoplopoma fimbria) startle events. We compare the proposed approach with a state-of-the-art end-to-end video detection method (ReMotENet) and a hybrid method previously offered exclusively for the detection of sablefish's startle events in videos from an existing dataset. Results show that our novel method comfortably outperforms the comparison baselines in multiple metrics, reaching a per-clip accuracy and precision of 80% and 0.81, respectively. This represents a relative improvement of 31% in accuracy and 27% in precision over the compared methods using this dataset. Our computational pipeline is also highly efficient, as it can process each 4-second video clip in only 38ms. Furthermore, since it does not employ features specific to sablefish startle events, our system can be easily extended to other behaviours in future works.

As marine ecosystems experience accelerating change, there is an urgent need for efficient and scalable biodiversity monitoring tools. We present a 10-step framework for integrating computer vision (CV) tools into long-term underwater biodiversity monitoring, using a case study from coastal British Columbia. Over 9000 h of unbaited remote underwater video footage were collected from two kelp farms and reference sites between March 2022 and June 2023. The framework includes steps for creating an annotated training dataset using unsupervised and supervised CV tools, culminating in the training and validation of a YOLOv8 object detection model. This process produced over 241,000 annotations across 54 pseudo-taxonomic categories (representing both taxa and visually similar groups of fauna), with a focus on fish and gelatinous zooplankton groups. The final model achieved an overall F1 (2 × precision × recall/(precision + recall)) of 0.74 and a mean average precision at 0.5 intersection over union (mAP50) of 0.78. The model had the highest performance on fine-resolution taxa such as Phanerodon vacca (F1 = 0.88) and Aurelia labiata (F1 = 0.90), and lowest performance on pseudo-taxa with limited visual distinctiveness such as Actinopterygii (F1 = 0.60) and Cnidaria (F1 = 0.60). A re-training experiment using annotation thresholds between 25 training images to full dataset availability (~200–24,000 images per group) found that model performance was positively correlated with annotation effort, with F1 averaging 0.84 and mAP50 averaging 0.91 at the maximum training dataset size. Our results suggest that the model is most effective for abundant and visually distinctive taxa, while performance declines for groups with coarse taxonomic resolution. We recommend optimizing annotation effort by targeting genus- or species-level taxa, having at least one broad-level group to capture order-level abundances, and supplementing annotations of rare groups with common but morphologically similar groups, which may further improve model performance.