Predicting Compliance with Fuel Sulphur Content Regulations to Improve Inspection-Targeting by Port State Control Authorities

Abstract

With a continuously growing maritime shipping industry comes an increase in emission created by ships. These emissions are a danger to our environment and health, which is why the International Maritime Organization (IMO) places strict regulations on the amount and type of emissions ships are allowed to emit. Among these emissions are Sulphur oxides, created by burning Sulphur rich fuels. It’s the task of the Human-Environment and Transport Inspectorate (ILT) to enforce compliance with Sulphur emission regulations for ships coming into port in the Netherlands through inspections. With the ILT’s push to become more data driven, our research explores the applicability of artificial intelligence (A.I.) for the purpose of improving the ship-selection process for inspections. The aim of this thesis is first, to test whether machine learning models are capable of predicting which ships are non-compliant, and second, to compare different models to find which type of model is best suited for this task. To accomplish this, we were provided with ship-inspection data collected between 2015 and 2017 from member states of the Paris Memorandum of Understanding on Port State Control. From this data we created ship representations that we used to train and test three different models with the goal of predicting non-compliance based on ship details and a ships inspection history. We tested a Random Forest, AdaBoost, and LogitBoost implementation in a repeated hold-out validation process with downsampling of the training set, and carried out a statistical analysis of the models’ results. In our test we found that the Random Forest has a significantly higher precision than AdaBoost and LogitBoost, and that for specific discrimination thresholds LogitBoost has a significantly higher precision than Random Forest but also excludes more items during the prediction process than Random Forest does. We estimate that Random Forest’s precision, if applied in the real-world, would be between 8.4% and 21%. We conclude that A.I. has the potential to increase the precision of inspection targeting, and that Random Forest models are well suited for that task. In order to improve the models, more data should be made available and integrated. The European Maritime Safety Agency, the ILT’s air quality measurement stations, and the Sentinel 5P satellite are all promising data sources that future research should explore. Considering the importance of environmental protection and the success of our research, we recommend to invest more recourses into researching the utilization of A.I. to enforce compliance, both, by exploring different A.I. methodologies, as well as optimizing existing approaches.