Estimating fishing effort in small-scale fisheries using GPS tracking data and random forests

During the last decade spatial patterns of industrial fisheries have been increasingly characterized using tracking technologies and machine learning analytical algorithms. In contrast, for small-scale fisheries, fishers' behaviour for estimating and mapping fishing effort has only been anecdotally explored. Following a comparative approach, we conducted a boat tracking survey in a small-scale reef fishery in Madagascar and investigated the performance of a learning random forest algorithm and a speed threshold for estimating and mapping fishing effort. We monitored the movements of a sample of 31 traditional sailing fishing boats at around 45 s time interval using small GPS trackers. A total of 306 daily tracks were recorded among five gear types (beach seine, mosquito trawl net, gillnet, handline, and speargun). To ground-truth GPS location data, fishers' behaviour was simultaneously recorded by a single on-board observer for 49 tracks. Typical, gear-specific track patterns were observed. Overall, the random forest model was found to be the most reliable, generic, and complex method for processing boat GPS tracks and detecting spatially-explicit fishing events regardless gear type. Predictions of mean fishing effort per trip showed that both methods reached from 89.4% to 97.0% accuracy across gear types. Our findings showed that boat tracking combined with on-board observation would improve the reliability of spatial fishing effort indicators in small-scale fisheries and contribute to more efficient management. Selection of the most appropriate GPS data processing method is dependent on local gear use, fishing effort indicators, and available analytical expertise.

Automated summary

This study found that using a random forest algorithm is the most reliable and effective method for analyzing boat GPS tracks and detecting fishing events in small-scale fisheries. Both the random forest algorithm and speed threshold method showed high accuracy in predicting fishing effort across different gear types, with accuracies ranging from 89.4% to 97.0%.