The Impacts Of Different Spatial Resolutions On The Fishing Ground Prediction Accuracy Of LSTM And The Comparison Between Its Prediction Accuracy And That Of The Quantile Regression Model Under The Optimal Spatial Resolution

As an important part of tuna fishery,Albacore tuna(Thunnus alalunga)is widely distributed in temperate and tropical waters of all oceans in the world.It is one of the main target species of tuna longline fishing.Its resource fluctuation is an important content of pelagic fishery research in China.The Western and Central Pacific,where the Cook Islands are located,is the largest tuna fishing area in the world.It is very important to improve the fishing ground prediction accuracy of target tuna and explore the relationship between the temporal and spatial distribution of tuna resources and the marine environment.Because the marine environment data are usually presented with different spatial resolutions,it brings different errors and affects the accuracy of tuna fishing ground prediction.To establish a better prediction model of Albacore tuna fishing ground,scientific and accurate marine environment data with large sample size and corresponding catch data are required as preconditions,so as to further understand the temporal and spatial distribution characteristics and habitat environment of Albacore tuna resources,and to finally achieve the goal of sustainable development and tuna fishery resources conservation and management.Based on the data of vessel monitor systems(VMS)of China’s pelagic fishery enterprises from January 1,2017 to May 31,2021,this paper calculates and matches the catch in number of Albacore tuna and longline hooks to the unit grid,and obtains the nominal CPUE(catch per unit effort,unit: inds./thousand hooks).Based on the data of more than 4 years,the time distribution changes of Albacore tuna fishing ground are analyzed by month and quarter.Twenty six spatio-temporal and environmental factors e.g.chlorophyll a concentration,sea surface height,temperature,salinity,dissolved oxygen concentration of 0～300m water layer,year,month,longitude and latitude,are selected as variables taking days as the time resolution.The K-S test(Kolmogorov Smirnov test)and Q-Q distribution map are used to detect whether the CPUE data set at each spatial resolution obeys the normal distribution.The correlation test is conducted between CPUE and spatiotemporal environmental factors,the factors that have little impact on CPUE are eliminated,and the non collinear spatiotemporal and environmental factors are extracted based on the VIF variance expansion factor and used for modeling analysis.Based on the long short term memory neural network model(LSTM),the fitting degree and accuracy of the model are analyzed to determine the differences of the four different spatial resolutions(0.5 ° × 0.5°,1 ° × 1°,2 ° × 2 °,and 5 ° × 5 °)on the fishing ground prediction results;Combined with quantile regression,the model is established for the data set corresponding to the spatial resolution with good prediction accuracy,and the integrated habitat index(IHI)is introduced.By comparing the Poisson correlation coefficient between the habitat index of Albacore tuna obtained by LSTM and quantile regression and the measured value of CPUE,the prediction accuracy difference between the two models is compared;Finally,the spatial distribution characteristics of CPUE of Albacore tuna are predicted in combination with the integrated habitat index(IHI).The results show that:(1)During the period from January 1,2017 to May 31,2021,the CPUE of Albacore tuna in the first half of the year was lower than that in the second half of the year,and the average CPUE reached the highest in June,and from October to December of the year;Based on the quarterly observation,the boundary of the Albacore tuna fishing ground in the second quarter was 12 ° S.The CPUE of the fishing ground in the south of 12 °S was significantly higher than that in the north of 12 ° S.In the third quarter,the fishing ground was scattered and the CPUE value was not high,and the fishing ground was shrink in the fourth quarter.(2)LSTM results showed that different resolutions had different prediction accuracy for Albacore tuna CPUE,and the best prediction resolution was 1 ° × 1 °,the mean absolute error and root mean square error were 0.0268 and 0.0452 respectively,and the prediction resolution was followed by 0.5 ° × 0.5°,2° × 2°,5° × 5°.The data of2° × 2 ° spatial resolution,its verification set and training set were relatively stable.The results of LSTM on the data with spatial resolution of 1 ° × 1 ° showed that the correlation between the verification results and the measured results was very significant,and the Person correlation coefficient was 0.904.(3)The quantile regression has a good prediction accuracy on the data with spatial resolution of 1 ° × 1 °,and the Person correlation coefficient reached 0.801.The Poisson correlation coefficients between the integrated habitat index and the measured CPUE value obtained by the quantile regression and the LSTM were 0.6874 and 0.8796 respectively,indicating that the prediction ability of the LSTM was stronger than that of the quantile regression,but the quantile regression was more accurate in screening the environmental factors affecting CPUE.(4)The habitat integrated index could better reflect the distribution of CPUE of Albacore tuna.The areas with high IHI value appeared in 9 °～13 ° S and 158 °～165 °W,and the sea areas with low IHI distribution were 7 °～ 8° S,156 °～ 168 ° W;14°～17°S,162°～168°W;9°～14°S,165°～168°W.Based on the above results,it was suggested that tuna longline fishing vessels of China should operate in areas with high IHI value,so as to improve the fishing efficiency.