Fishing Techniques Classification Based on Beidou Trajectories and Machine Learning

Abstract

Beidou navigation information is widely used in Chinese fisheries. It helps fishermen to locate themselves and hedge. Compared with the data ashore, analysis of satellite data on the ocean is still not enough. Inspired by vehicle trajectory analysis, we want to do some analysis on boat trajectories. Fishing techniques are various, so their trajectories. In this paper, we choose 3 fishing techniques (Trawling, seining and gillnetting). We implement classification with two different algorithms. One of them is ResNet which belongs to image classification with deep learning. The other one is LightGBM which is a kind of decision tree algorithm. The result shows that although deep learning has made great success in daily life images classification, it adds too much redundant pixels and ignore the speed and direction parameters in this task. This leads to a lower precision and more calculations. In contrast, LightGBM can use information effectively and has a higher score with a higher speed. This work shows traditional machine learning algorithm can achieve better result than deep learning algorithm in some circumstance. It will also contribute to establish a smart ocean system.