| Description: |
Sea urchin overgrazing can transform vital kelp forest ecosystems into less productive 'urchin barrens'. In southeastern Australia and northeastern New Zealand, overgrazing by three sea urchin species, Evechinus chloroticus, Centrostephanus rodgersii and Heliocidaris erythrogramma, has caused habitat losses across 1000s of kilometres of coastline. Underwater imagery is now a critical tool for monitoring urchin impacts on large scales, however image manual annotation is prohibitively time-consuming, delaying crucial decision-making and necessitating an automated solution. Using underwater imagery collected from multiple imagery platforms from northeastern New Zealand and southeastern Australia, a sea urchin detection and classification dataset was developed. This dataset was used to train a machine-learning object detection model. Imagery and annotations were made available and standardised through the Squidle + online data management platform and the resulting model, 'Urchinbot', is also deployed within Squidle+. The development procedure for the model is described, including iteratively improved detection accuracy, modifications and hyperparameters for best performance. 'Urchinbot' can accurately and rapidly detect urchins within underwater imagery (F1 score = 0.879, mean Average Precision [mAP50] = 0.908 across all species) and is precise in distinguishing between the three key species, with less than 1 % confusion. Validation on real-world datasets found similarly strong model performance and high practicality. 'Urchinbot' can dramatically improve the processing speed of underwater imagery data (∼24x faster than manual annotation), aiding large-scale assessment of urchin populations and impacts. Its application will accelerate management decisions on urchin overgrazing and serve as a foundation for further machine learning development to address time-sensitive environmental challenges in kelp forest ecosystems. |