Vessel Behaviour Classification from AIS Without Geographical Biases

Abstract

The automatic detection of vessel behaviours from Automatic Identification System (AIS) data is a challenging aspect of designing intelligent systems and aiding maritime situational awareness. The development of such systems remains limited to some activities like fishing, and by geographical biases that prevent systems to generalise to other areas than that used for training.To contribute to these questions, we investigate how to treat raw data or engineered features so that they do not convey such biases at training time and we propose methods for point-wise behaviour detection in the context of container vessels with four target behaviours. Several systems are studied, with raw data or engineered features as inputs, followed by shallow or deep learning classifiers. While good performances are obtained by several of them, we observe that a decision tree classifier with engineered features outperforms an LSTM in areas where no labelled data is available for training.

Beyond Geofencing: Behavior Detection Using Ais

“…For this we introduce new unbiased features that can be calculated from the base features. This was introduced in [43] and is reformulated here.…”

“…They have been selected to be unbiased and to carry useful information for behavior detection. As in [43], the following features have been selected : 𝑠𝑜𝑔, 𝑑𝑟𝑖𝑓 𝑡, Δ 𝑠𝑜𝑔 , Δ 𝑐𝑜𝑔 , Δ ℎ𝑒𝑎𝑑𝑖𝑛𝑔 , 𝜎 𝑐𝑜𝑔 , 𝜎 ℎ𝑒𝑎𝑑𝑖𝑛𝑔 , 𝜎 Δ 𝑐𝑜𝑔 , 𝜎 Δ ℎ𝑒𝑎𝑑𝑖𝑛𝑔 , 𝜎 𝑝𝑜𝑠 , and 𝜎 𝑠𝑜𝑔 .…”

“…AIS has been used for vessel behavior detection mainly in the context of fishing vessels, in order to detect illegal fishing activities [8,10,20,24,23,25,22,42,46]. Fewer works have been done on other types of vessels [7,28,43]. A recent trends in literature is to use some image generation methods coupled with image recognition technics, usually Convolutional Neural Networks (CNNs) [6,7,28].…”

“…This paper is a continuation of[43]. The main differences are: in section 2.1 we detail the notion of geographical biases; in section 2.2 we introduce a new data augmentation technique; section 2.3 we reformulate the design of unbiased features already introduced in[43]; in section 3 we introduce new models: random forests, hidden markjov models and transformer networks; finally, in section 4, all experiments are new and are conducted using new data.…”