Identifying Purse Seine Vessel Operations Through Machine Learning Models for Better Spatial Fishing Effort Estimates
Abstract
Spatial fisheries management requires precise, spatial explicit information on species distribution and fishing impacts. In the NW Mediterranean, the distribution of small pelagic fish (SPF) stocks is currently assessed through annual acoustic surveys. Although accurate, these surveys might lack the temporal resolution needed to capture population dynamics. Analysing fishing fleet spatial behaviour allows estimation of fishery impacts and can uncover patterns of target species at high temporal resolution. In this study we apply machine learning models to classify multiple vessel operations on vessel monitoring system (VMS) data for purse seiners targeting SPF populations. Three main vessel operations were defined based on onboard observations: Fishing, Tracking and Cruising. Then, random forest models were trained to predict vessel operations in VMS data using seven predictive variables. Machine learning models highly improved predictions accuracy (81% and 73%) compared to a classical speed filter method (60%). Fishing effort metrics were computed and compared across predictive methods resulting in a high overestimation of fishing activities when using a speed filter approach. The effect of spatial resolution in fishing effort metrics was also tested revealing a good performance of random forest predictions at 2–3 km2. The methods developed allow quantification of three vessel operations activities improving purse seiners effort metrics compared to classical binary (fishing/non-fishing) approaches. Consequently, spatially explicit catch per unit effort (CPUE) estimates for SPF will also be improved as well as the accuracy of the information needed for spatially manage of this fishery.
