Calculation of fishing vessel gross tonnage based on machine learning methods

Vessel gross tonnage can represent the overall internal volume of a ship in the maritime industry, including all spaces, such as cargo areas, living quarters, and on-board spaces. However, the vessel gross tonnage can be limited to a highly complex calculation and low accuracy in general. In this study, an optimal formula was proposed to integrate feature engineering with multiple algorithms using machine learning. Initially, the feature variables were mined using the structural parameters of the ship. A correlation analysis was also conducted on the feature variables. Their correlation was also determined for the Pearson correlation coefficients between the feature variables and the gross tonnage. Three types of nonlinear regressions were performed on the gross tonnage: The linear multiplicative model, the sub-item exponential model, and the hybrid model. The nonlinear relationship between the feature variables and the gross tonnage was then obtained after optimization. Subsequently, the dataset of 1 913 fishing vessels in the South China Sea region was divided into the training and testing sets in a ratio of 8:2. Specifically, the dataset included 448 trawlers, 479 purse seiners, 440 gillnetters, 237 setnetters, and 309 longliners. Regression of the vessel gross tonnage was also performed using Backpropagation Neural Network (BPNN) and Random Forest (RF). Model fitting and validation were then conducted using the nonlinear least squares method (LSM) and particle swarm optimization (PSO). Ultimately, the robustness tests were carried out to verify the models. Different models were also selected with the various noise intensities of 0%, 3%, 5%, 10%, and 20%. The results showed that the Pearson correlation coefficients between the feature variables (length L, width B, and dapth D) and the gross tonnage (GT) were 0.9377, 0.8204, and 0.9327, respectively. These values were all close to 1, indicating a strong correlation between L, B, D, and GT. Therefore, the L, B, and D were selected as the feature variables. Furthermore, RF outperformed BPNN in various error metrics, including mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and the coefficient of determination (R²), indicating a higher accuracy of the prediction on the gross tonnage of fishing vessels. The nonlinear LSM outperformed the PSO in various evaluation metrics, including mean bias error (MBE), MAE, MAPE, and computational efficiency. The hybrid GT prediction model also outperformed the linear multiplicative model and the sub-item exponential model. Moreover, there was significantly better interference resistance under different noise intensities, compared with the linear product and the sub-item exponential model. The values of the MBE, RMSE, MAE, MAPE, and R² were -1.244 4, 32.036 2, 24.481, 9.94%, and 0.961 9, respectively, indicating better generalization and robustness. Vessel gross tonnage was calculated to integrate the feature engineering with multiple algorithms. A comparison was also made on the actual ship gross tonnage, according to the International GT measurement formula, and the prediction. The higher accuracy was achieved to validate the effectiveness of the prediction. Artificial intelligence and machine learning were employed to optimize the prediction for the gross tonnage of the fishing vessels, significantly improving the accuracy of the calculations. This finding can provide valuable references for the digital and intelligent management of the fishing vessels in the fisheries.