Abstract
With the advent of the big data era in ocean remote sensing and fisheries, there is a growing demand for finer temporal scales to predict spatial distribution of the jumbo flying squid (Dosidicus gigas). This can help reduce fuel costs and provide higher quality and faster decision-making. Therefore, this study employed a deep neural network (DNN) model, using sea surface temperature, sea surface height, sea surface salinity, and photosynthetically active radiation as input factors, with catch per unit effort as the output factor. We construct five cases with temporal scales of 3, 6, 10, 15, and 30 days using data spanning 10 years (2012–2021). Additionally, the performance of DNN was compared with those of traditional methods such as generalized additive model (GAM), extreme gradient boosting (XGBoost), and artificial neural network (ANN). The results demonstrated that the DNN model had the best performance. As the temporal scale decreased, the mean squared error and the mean absolute error increased, whereas the area under the precision−recall curve decreased, indicating a decline in model performance. The interpretability analysis indicated that spatial and temporal factors significantly contributed to the model, with longitude exhibiting the highest contribution. To improve the accuracy of finer temporal scales, future research should focus on reducing noise in the data and address the presence-only nature of fishery data, particularly by cleaning the unsampled portions.