The Identification and Prediction in Abundance Variation of Atlantic Cod via Long Short-Term Memory With Periodicity, Time–Frequency Co-movement, and Lead-Lag Effect Across Sea Surface Temperature, Sea Surface Salinity, Catches, and Prey Biomass From 1919 to 2016

Abstract

The population of Atlantic cod significantly contributes to the prosperity of fishery production in the world. In this paper, we quantitatively investigate the global abundance variation in Atlantic cod from 1919 to 2016, in favor of spatiotemporal interactions over manifold impact factors at local observation sites, and propose to explore the predictive mechanism with the help of its periodicity, time–frequency co-movement, and lead-lag effects, via long short-term memory (LSTM). We first integrate evidences yielded from wavelet coefficients, to suggest that the abundance variation potentially follows a 36-year major cycle and 24-year secondary cycle at the time scales of 55 years and 37 years. We further evaluate the responses of Atlantic cod abundance to the external impact factors, including sea surface temperature (SST), catches, prey biomass, and sea surface salinity (SSS), in aid of the wavelet coherence and phase difference, which allows us to identify the dominantly correlative factors and capture the leading roles along the time domain and then divide the responses around the recent 60 years into three stages: before 1985, 1985–1995, and after 1995. At the first stage, the reason for the decline in abundance could be mainly attributed to the rapid rise of fish catches. At the second stage, the impact of SST and SSS also provides significant indices, besides overfishing; meanwhile, the mortality of primary producers and forced migration of fish species indirectly cause the decline. At the third stage, warming SST and growing SSS directly led to the decrease of abundance. Finally, we establish one ensemble of LSTM-SAE architecture to comprehensively reflect the predictive patterns at each stage. It has been demonstrated from experimental results that the models behaved better when intentionally feeding with the dominantly correlative multivariate inputs, instead of either all factors or only the abundance. The proposed scheme provides opportunities to symmetrically identify the underlying predictive attributes of Atlantic cod abundance and potentially perform as the quantitative references in reasonably making fishing decision. With the rapid development in deep learning capabilities, it is hopeful to expect better predictions of the responses to global changes, not only for Atlantic cod but also for other fish species and the ecosystem as a whole.