Data for "Machine Learning-based Prediction of Passive Gears from Vessel Tracking Data in Small-Scale Multi-gear Fisheries"

This is the shared framework described in the paper "Machine Learning-based Prediction of Passive Gears from Vessel Tracking Data in Small-Scale Multi-gear Fisheries" by Lattanzi et al.
A Machine Learning approach was tested to predict passive gears from vessel tracking data in multi-gear small-scale fisheries, by focusing on hauling events instead of entire fishing trips. This is crucial for SSF fleets that often use multiple types of passive gear during a single trip, as traditional trip-level analysis would miss important details about gear usage. By examining each time a gear is retrieved (hauling), the researchers can more accurately identify what gear was used.

The method involved the pre-processing of data from 1634 fishing trips (recorded for 10 vessels from Ancona, Italy), keeping only the 7,164 hauling events that had soak time information. This data was then split into training (70%), validation (20%), and a held-out test set (10%). To find the best model, a nested cross-validation approach was used for hyperparameter tuning and performance estimation. The top-performing model was then evaluated on the unseen test set, and sensitivity analysis (using Permutation Feature Importance) helped understand which features were most important. Finally, the best model for each level of analysis (i.e., specific gears or gear category) was selected based on strong performance and minimal overfitting, and, then, fine-tuned by re-training it with an optimized split and refined hyperparameters to maximize its predictive ability.