1. Introduction
According to the Sustainable Development Goal #14 (SDG 14) identified by the United Nations (UN) [
1], overfishing is a global challenge which “reduces food production, impairs the functioning of ecosystems and reduces biodiversity”. Improving the sustainable use of marine resources requires monitoring tools which can provide long-term observations of (i) the world’s fish stocks and (ii) fishing fleets’ activity. The latter can now be monitored by several systems. In particular, Automatic Identification System (AIS) and more recently Vessel Monitoring System (VMS) provide huge volumes of data on the geographical position and the identity of vessels around the world. AIS and VMS are cooperative systems that rely on vessels to report their information. These systems are compulsory for vessels of the European Union (EU) which are larger than 15 m [
2], and vessel data can be collected from shore-based or satellite receivers.
Satellite Remote Sensing (RS), especially Synthetic Aperture Radar (SAR), can also be used to monitor vessels’ position and is complementary to AIS and VMS as it does not rely on vessels’ cooperation [
3,
4]. Several research groups have developed operational automatic vessel detectors [
5]. In parallel, the availability of satellite radar data has also greatly improved with the launch of the Sentinel-1A and Sentinel-1B (S-1) SAR satellites in 2014 and 2016, respectively. This satellite mission, part of the European Space Agency’s (ESA) Copernicus program, continuously generates about 6 TB of freely available imagery per day [
6]. These free data have already shown value for monitoring marine activities. For example, [
7] applied the Search for Unidentified Maritime Objects (SUMO) vessel detector to ∼11,500 S-1A images collected over the Mediterranean Sea between October 2014 and September 2016, and obtained more than 600,000 vessel detections.
Although vessel detection with SAR is already operational, vessel classification (i.e., attributing a vessel type to the detected vessels) remains challenging. This is especially true at the spatial resolution of S-1 (20 m), because it is difficult to extract geometric and radiometric features representative of the different vessel types [
8]. Previous research on vessel classification has focused on the use of commercial high-resolution (1–2 m) SAR imagery from TerraSAR-X and COSMO-SkyMed [
9] and on three vessel types, namely bulk carriers, container ships, and oil tankers, which represent 70–80% of the vessels worldwide [
10]. Several authors used both geometric (length, width, aspect ratio) and radiometric (mean intensity, standard deviation, fractal dimension) features as input of the vessel classifier, and obtained classification accuracy above 90% [
11,
12,
13,
14,
15]. In particular, in [
13] the authors tested their approach on high-resolution (3 m) TerraSAR-X images and medium-resolution (10 m) Radarsat-2 images, and obtained classification accuracy above 90% and below 50%, respectively. They found that for low-resolution images, geometric features played a dominant role in the classification, and radiometric features were only useful at higher resolution, resulting in higher classification accuracy [
13,
14]. More recently, [
16] classified cargo, tanker, windmill, platform, and harbor structure, with an overall accuracy of 94% using a Convolutional Neural Network (CNN). [
17] also reported high classification accuracy (98%) with TerraSAR-X images (1 m resolution), using features obtained from Histogram of Oriented Gradients (HOG) combined with Manifold learning for dimensionality reduction, and a Task-Driven Dictionary Learning framework.
To the authors’ knowledge, there have not been studies on the classification of SAR detections as fishing or non-fishing vessels. Although fishing vessels are not as numerous as bulk carriers, container ships, and oil tankers, classifying them could support the monitoring of fisheries and help tackle problems around overfishing. In this paper, we present an approach based on Random Forest (RF) trained to classify fishing and non-fishing vessels based on their geographical location (longitude, latitude), an estimate of their length, their distance to the nearest shore, and the time of the measurement (am or pm). The training phase relies on labelled AIS data. After cross-validation, we apply the RF classifier to a collection of vessel detections obtained by processing S-1 images of the North Sea using the SUMO detector. We show that the variation in fishing-vessel count obtained with our technique agrees with Global Fishing Watch data on fishing-vessel presence. Apart from the initial training phase which relies on commercial AIS data, our approach uses open-source software and free S-1 imagery, making it a cost-effective method to monitor fishing patterns.
6. Conclusions
We presented a method for the classification of SAR vessel detections into fishing and non-fishing-vessel classes. The method relies on a RF classifier with 5 input features: longitude, latitude, length, distance to nearest shore, and time of measurement (am or pm). The classifier was trained and tested on labelled AIS data, and was then applied on S-1 detection provided by the SUMO detector. The length, latitude, longitude, and distance to nearest shore are the only important features in the classification. When tested on datasets with realistic sample proportions (12.6% fishing, 87.4% non-fishing), the overall classification accuracy is 91%, but the precision of the fishing class is only 58%. Precision, however, varies with the location and is only low in areas where the two vessel classes overlap, and where the non-fishing class dominates. When applied to the 2017 S-1 data, there is a good agreement between our monthly fishing-vessel count and data from GFW on fishing-vessel presence in 2016. Although our vessel count is not a reliable estimate of the actual number of fishing/non-fishing vessels, the results suggest that our method still captures the correct temporal variation in fishing activity. In particular, our approach can be used to indicate intensification or reduction of fishing effort in a given area, which is critical in the context of the global overfishing problem.