Automated Segmentation of Bryozoan Colonies and Prediction of Growth in Kelp Farms

Abstract

In the aquaculture industry, the cultivation and harvesting of seaweed play a pivotal role, of‐ fering substantial economic and ecological benefits. Seaweed biofouling, which includes the growth of undesirable organisms on seaweed, significantly diminishes the quality of the kelp (brown seaweed), lowering the market value of the product. In Norway, the growth of bryo‐ zoan colonies, which are encrusting small (few mm to cm) invertebrate aquatic organisms, is the major bottleneck of the industry, where the growth period is reduced (harvest in May‐ June). Therefore, monitoring and managing bryozoan colonies is crucial to maintaining healthy kelp for the cultivation of high‐value products. This research, which leveraged state‐of‐the‐art instance segmentation techniques using the Detectron2 framework, has provided significant insights. It detects, segments, and quantifies bryozoan colony coverage on kelp that assist in the estimation of how fast they spread in the farm, and allow farmers to manage the produc‐ tion. Several processes have been applied to the dataset, including annotation marking of the precise locations and extents of bryozoan colonies on the kelp lamina and performing a resizing and augmentation for the dataset to achieve good results. Different Mask R‐CNN models with different backbones are compared to find the model that fits the research application. The results are evaluated using the Average Precision (AP) metric. The APBryozoan= 44% and APkelp = 88% achieved is comparable to other papers for Mask RCNN and the model’s benchmark. Additionally, this study introduces a time series model to enhance the understanding of bryozoan colony growth. It uses an Auto‐Regressive with exo‐ genous inputs(ARX) statistical machine learning model to predict future data points by consid‐ ering past values and external influences, achieving a fit percentage of 92% for predicting the area of bryozoan colonies based on historical data and external inputs. This study concludes that applying cutting‐edge technologies, such as deep learning al‐ gorithms, enhances accuracy and significantly saves time in the seaweed cultivation process, marking a significant step forward in the field. Based on this research, the optimisation of Sea‐ weed Production can be implemented in future work.