Convolutional-LSTM approach for temporal catch hotspots (CATCH): an AI-driven model for spatiotemporal forecasting of...

Efficient fisheries management is crucial for sustaining both marine ecosystems and the economies that heavily depend on them, such as Iceland. Current fishing practices involve decisions informed by a combination of personal experience, current data on environmental and oceanographic conditions, reports from other captains, and target species within the constraints of the fishing quota. However, the intricate spatiotemporal dynamics of fish behaviour make it difficult to predict fish stock distributions. Despite technological breakthroughs in fishing vessel data collection, much of the decision-making still relies heavily on subjective judgement, highlighting the need for more robust, data-driven predictive methods. This paper presents CATCH, a convolutional long short-term memory neural network model that forecasts fish stock probability densities over time and space in Icelandic waters to support operational planning and adaptive strategy in fisheries. The framework represents the first utilization of large-scale Icelandic fishing fleet data integrating multidimensional inputs, particularly depth, bottom temperature, salinity, dissolved oxygen and catch data, to produce accurate, multivariate forecasts. The model achieves favourable performance with average RMSE, MAE, WD, and SSI of 4.71 × 10^-3, 1.16 × 10^-3, 0.94 × 10^-3, and 0.955, respectively, for cod, while 6.13 × 10^-3, 1.25 × 10^-3, 1.04 × 10^-3, and 0.949, respectively, across other target species (haddock, saithe, golden redfish, and Greenland halibut). Additionally, Syrjala's test yielded nonsignificant P-values (P > .05) in most cases across lags and forecast horizons, indicating that the predicted and observed distributions are statistically indistinguishable. Its promising results suggest deep learning models have the potential to optimize fisheries operations, enhance sustainability, and support data-driven decision-making.