Many fish stocks have shown to display nonlinear dynamic behavior. In the nonlinear forecasting framework, the ability to make useful predictions depends on the nature of time series. However, ecological time series are usually short and noisy while repeated measurements in spatially replicated sites are common. In this study, we developed a spatial time delay embedding method via Gaussian process modeling to evaluate whether information on spatial structure would improve predictive ability. The method was then tested against simulated data as well as the larvae survey dataset in the CalCOFI database to forecast recruitment.
We found that spatial information greatly improves the forecast ability of recruitment in sardine. This is particularly important when spatial time series are available but each of the time series is not long enough to make meaningful predictions. Our method also provides the ability incorporate habitat variables in space and time.