基于机器学习的金枪鱼随附群集群存在状态与规模分级预测OA

To accurately estimate the biomass of tuna aggregations associated with drifting fish aggregating devices(DFADs)in the Western and Central Pacific Ocean,this study developed binary and multi-class classification models using fishing log records and acoustic buoy data to identify the presence of tuna aggregation and school size,respectively.Missing values were input via Lagrangian interpolation,while dataset optimization incorporated spatiotemporal grid matching and filtering of data closest to local sunrise time.Classification was performed using multilayer perceptron(MLP)neural networks.Key findings revealed:1)Near sunrise,tuna aggregations around DFADs predominantly inhabited depths below 80 m,with biomass distribution proportionally increasing at deeper layers;2)The MLP classifier achieved 84%overall accuracy in detecting aggregation presence/absence,demonstrating high sensitivity for presence identification(0.98)but low specificity for absence detection(0.27);3)The overall accuracy of MLP classifier was 0.48 in identifying the scale of tuna aggregations under DFADs,with a relatively low recognition error rate for the category below 30 t(0.54 for precision),similar classification error rates between the 30-55 t and over 55 t categories(precision of 0.45 and 0.48,respectively),and higher recognition accuracy for the 30-55 t category(0.68 for sensitivity)than for categories below 30 t(0.35)and over 55 t(0.37).The results can provide a scientific basis for exploring tuna abundance indices based on acoustic data and support the resource assessment of tropical tuna stocks.