Improved deep learning based litter detection in aquatic environments in Indonesia using drones

Millions of metric tons of plastic waste enter the ocean every year, posing a stress to the marine environment. Out of the rivers that are assessed to be responsible for 80% of the riverine plastic emissions, more than hundred are located in Indonesia. Indonesia is estimated to be the fifth most relevant riverine plastic waste source, indicating the importance of the country as a research area. Within this research, an open source deep learning based plastic waste assessment system was further improved, and applied to the Indonesian rivers Citarum, Cisadane and Tukad Saba. The systems key improvements were (i) training of the neural networks with much larger and more diverse plastic waste data sets, (ii) an adjusted classification system to make the waste assessment results more easily comparable to other waste monitoring methodologies, and (iii) waste assessment results were georeferenced for facilitating comparisons with other plastic litter monitoring methodologies and complementing net trawl surveys, field campaigns and clean-up activities. The improved deep learning based litter detection system had an overall accuracy of 83% for detecting litter in aquatic environments. The key findings of the research show that the system can be used for assessing waste type compositions, potentially identifying waste sources or plastic litter accumulation zones and empowering stakeholders with actionable information on a local and regional scale.