Forecasting mussel settlement using historical data and boosted regression trees

Abstract

Many aquaculture sectors internationally, most notably for the cultivation of bivalves, rely almost completely on wild-caught juveniles (‘spat’) to stock farms, with poor ‘catches’ representing one the biggest constraints on global production. An example of this practice is green-lipped mussel Perna canaliculus aquaculture in New Zealand, where the industry in the main growing region has been monitoring P. canaliculus settlement for almost 40 yr. This practice involves deploying settlement arrays across the region to guide the places and times to place spat-catching rope. Using a subset of these data spanning 25 yr (1993-2018), we identified regional spatio-temporal patterns of P. canaliculus spat settlement. Boosted regression tree (BRT) models were used to forecast settlement at 2 different sub-regions with consistent high catch yields. BRT models confirmed a strong seasonal influence on settlement, with highest predicted settlement levels coinciding with the main P. canaliculus spawning period (late summer to autumn). Positive relationships were detected between settlement and the occurrence of positive temperature anomalies, easterly winds, periods of large tidal range and Southern Ocean Oscillation Index values associated with La Niña episodes. The models were able to forecast P. canaliculus settlement with excellent prediction accuracy based on time of year and environmental conditions 1 mo prior to collection. This study highlights the benefit of undertaking long-term monitoring of spat settlement and the related environmental factors that affect this ecological process. In combination with advance modelling techniques that enable forecasting of settlement densities, such knowledge can help to overcome challenges in spat supply and enable production upscaling.