State-Space Modeling For Inter-Site Spread Of Sea Lice With Short-Term Population Predictions

Accurate predictions for sea lice abundance can have a great impact on effective management of the parasite. Sea lice data are often from multiple production cycles coming from different aquaculture sites and exhibit strong process and observation variations, presenting challenges in developing effective prediction tools. Reliable predictions of sea lice abundance on aquaculture sites can save money and efforts in organizing effective control measures. A quantitative evaluation of different methods for estimating sea lice infestation pressure on aquaculture sites was presented in a preparatory paper (Elghafghuf et al., 2020). In the current work, we used the favored method for infestation pressure and predicted future sea lice abundances on different parasitic life stages at the site level using a multivariate state-space model fitted to data from aquaculture sites located near Grand Manan Island in the Bay of Fundy, New Brunswick, Canada. We employed a rolling prediction procedure that did not require the prediction model to be refitted to the data at each time of prediction and compared our results with those from a similar approach involving re-estimation of the model at each prediction. The accuracies of predictions were assessed based on hold-out samples of data, whereby the predictions were computed without prior knowledge of these observations. The results showed that the non-updated-model procedure gave short-term (six weeks) predictions with reasonable accuracy, albeit lower prediction errors could be obtained with the updated-model approach. The performance of the non-updated-model prediction, as well as its practical and computational advantages, suggests it to be well suited for user-friendly implementation linked to underlying database. Additionally, the research quantified the contributions of predictive covariates such as infestation pressures to the sea lice predictions.