open access

https://hdl.handle.net/2115/95297

Predicting catch of Giant Pacific octopus Enteroctopus dofleini in the Tsugaru Strait using a machine learning approach

英語

GRU

2023-05

Elsevier

Fisheries research

261

106622

We attempted to forecast catches of giant Pacific octopus Enteroctopus dojeini using past catch data and the machine learning approach. The algorithms tested were the Seasonal Autoregressive Integrated Moving Average (SARIMA), Light Gradient Boosting Machine (LightGBM), a type of Gradient Boosting Decision Tree (GBDT), and the Recurrent Neural Network (RNN), the latter of which is represented by Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The catch of giant Pacific octopus on the southern coast of the Tsugaru Strait was backcasted for 2001-2019, using data of at least 20 years (1981-2018) for training. In this study, a longer training period led to better performance of SARIMA; the model trained from all past years performed better than models trained from the past 20 or 5 years. Moreover, for LightGBM, fishing-season accuracy of the basic L-GBM's was equal to that of L-GBM-T, which adds environmental factors to the explanatory variables. As a result of 19 backcasting trials, the SARIMA model obtained the best agreement with the actual catch data (R2 = 0.848) followed by L-GBM and LSTM (both R2 = 0.847), and GRU (R2 = 0.839). The root mean square error (RMSE) was within a narrow range for all four models (28.5, 28.0, 27.9, and 28.2 metric tons [t] for SARIMA, L-GBM, LSTM, and GRU, respectively). The SARIMA model was unstable when performance was assessed by the mean absolute percentage error (MAPE), which was standardized by average catch per year, whereas LSTM and GRU remained relatively stable. However, the latter two models were difficult to interpret due to their black box-like structures. By contrast, the SARIMA model was advantageous for interpreting the results. LightGBM showed good predictive performance for time-series data, despite being a static model. Thus, each model has advantages and disad-vantages, so it is necessary to choose the best model depending on the purpose.

学術雑誌論文

Accepted Manuscript

Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International

© 2023. This manuscript version is made available under the CC-BY-NC-ND 4.0 license http://creativecommons.org/licenses/by-nc-nd/4.0/

0165-7836