TU Delft and India collaborate on self-supervised Learning for Plastic Hotspots Identification from Satellite Images

Plastic hotspots refer to regions where plastic waste accumulates, often in large quantities, due to factors such as water currents, tides, and human activities. These hotspots represent significant environmental concerns because they contribute to pollution in aquatic ecosystems, endanger aquatic life, and impact human health through contamination of the food chain.

The original idea behind SPIHSI was to merge the research done at IIT Delhi and TU Delft to develop a state-of-the-art AI-based detector of plastic hotspots that could be used to localize and quantify hotspots everywhere around the globe. These works include TU Delft's AIdroLab (PI: Dr Riccardo Taormina) research on self-supervised Deep Learning models for plastic detection using camera images, and the Yardi School of AI at IIT Delhi's work on change detection from satellite imagery (PI Dr Sudipan Saha). 

Typically, models for change detection leverage deep learning frameworks pre-trained on large, generic datasets like ImageNet. Our approach aimed to improve performance by using self-supervised learning to pre-train models with more relevant high-resolution, unlabelled satellite imagery. 

A crucial part of the project was validating the satellite-based plastic hotspot detection by comparing it with on-the-ground data. To achieve this, TU Delft sponsored a project involving three MSc students from the Faculty of Civil Engineering to characterize plastic pollution in the Saigon River, specifically in Ho Chi Minh City. With the assistance of local researchers, the students gathered high-fidelity ground truth data, correlating with requests for high-resolution satellite images (30 cm/pixel resolution) from the same locations and timeframes.

Unfortunately, due to the monsoon season, no clear-sky satellite imagery was available on the dates the students conducted their field surveys. While the students successfully quantified and characterized plastic pollution in the Saigon River, the lack of corresponding satellite imagery prevented us from developing the model with this dataset. This is an example of challenges we face in this field.

Despite this setback, alternative plans allowed the project to move forward. IIT Delhi developed an unsupervised change detection method capable of identifying floating deposits in water, though not specifically plastic hotspots. They tested this methodology in various locations across India and also experimented with promptable segmentation techniques, utilizing state-of-the-art language-vision models like the Segment Anything Model. However, the reliance on lower resolution (3m/pixel) imagery has limited the detection capabilities. 

Meanwhile, TU Delft is using the camera-based dataset collected in Ho Chi Minh City to refine and test an improved version of its self-supervised model for camera-based plastic detection. The model and dataset will be released later this year, with plans for presentation at either the European Geosciences Union (EGU) or American Geophysical Union (AGU) 2025 conferences.