Automated detection of sea cucumbers in turbid subtidal marine habitats: An explainable approach

Deposit-feeding sea cucumbers perform important ecosystem functions, enhancing nutrient cycling and maintaining ecosystem health. Understanding this function over large spatial scales requires substantial population surveys, which can be logistically challenging. While deep learning has shown promise for sea cucumber detection, factors like object occlusions, limited ambient light, and turbid environmental conditions hamper model generalizability. Yet, detection models that can generalise across a broad range of environmental conditions are key to conducting large-scale in-situ population surveys of sea cucumbers. Here, we demonstrate the pipeline using YOLOv8 for large-scale population surveys of subtidal organisms. Specifically, we use sea cucumbers as a model species to showcase a series of photometric data augmentation processes tailored for small datasets, collected under various turbid environmental conditions with a consumer grade waterproof camera (GoPro Hero9 Black). We further interpret the model detection sensitivity with Eigenvalue-based class activation map (Eigen-CAM) to visualise the model perceptual fields on unseen in-situ image samples. Our sea cucumber detection model achieved 94.1 % mean average precision, with the augmentation processes effectively enabling generalisation when applied to unseen sites despite limited training data. Our study includes Eigen-CAM, a novel visualization tool that enhances model interpretability and transparency during model prediction. This enables researchers to visually validate the model outputs and effectively identifying areas for refinement. Overall, our detection model offers an efficient solution for subtidal ecological surveys using simple cameras, particularly in low-visibility environmental conditions. • Efficient holothurian detection under turbid marine condition with machine learning. • Tailored photometric augmentation processes increased data quality and quantity. • Visualising model prediction with Eigen-CAM provides explainable detection. • Video based sampling and machine learning supports efficient ecological surveys.