Collaboration with the IIASA team for marine litter tile sorting

About the project

The Marine Remote Sensing Group (MRSG) has performed extensive surveys on coastal environments in Greece, gathering valuable information about marine plastic pollution using drones. This information needs to be further classified based on the presence and quantity of litter. However, due to the sheer amount of data collected by the drones, new and innovative approaches are needed for transforming the images into a training data set suitable for ML algorithms. Crowdsourcing solutions can be used to classify these data sets in a very rapid way and hence provide calibration and validation data sets, e.g., to test ML algorithms. At the same time, they can help to raise awareness of the marine debris litter problem and thereby help to mitigate the problem directly. The use of Picture Pile, developed at IIASA, serves perfectly this purpose as it combines a crowdsourcing model with rapid image classification. This application has already been used for rapid image assessment to map cropland, deforestation, presence of night-time lights, poverty, damage to buildings after a hurricane and cloud cover on optical imagery. The application can ingest any type of imagery including that from drones.

The MRSG data sets will be formatted to be compatible with the Picture Pile application. To help users classify the images faster and more accurately, each image will be split into smaller tiles with 1024x1024 pixels. The task of the users will be to examine each tile to determine whether they can see visible evidence of plastic litter. The tiles will be geolocated, and after sorting through the pile of pictures, a large data set will be available for use with ML algorithms for the development of an automated application for marine plastic litter detection.

The GeoWiki Platform and the Picture Pile application

The Geo-Wiki platform, developed at IIASA, provides citizens with the means to engage in environmental monitoring of the earth by providing feedback on existing information overlaid on satellite imagery or by contributing entirely new data. Data can be input via the traditional desktop platform or mobile devices, with campaigns and games used to incentivize input. These innovative techniques have been used to successfully integrate citizen-derived data sources with expert and authoritative data. Since 2009, Geo-Wiki has grown rapidly, with currently over 15,000 registered users and applications in many successful citizen science campaigns, most recently crowdsourcing global agricultural field-size data, performing post-disaster damage assessment, poverty mapping and more. We have many ongoing projects that bring together the field of Earth Observation and citizen science, including several citizen observatories funded by the EU, which are developing new services such as land cover change detection, quality assurance of citizen science data, supporting local food growers and more.

Picture Pile can be used in two ways: a) to label the presence and absence of marine litter from the drone data in the red box (see Figure 1a) and b) to specify the percentage of marine litter shown in the red box (Figure 1b). Additionally, the volume of marine litter can be quantified using existing empirical relationships. The labeled data will provide a free and open reference data set of marine litter with a time stamp. The data set can also be used to provide guidance for planned clean-up events.


Figure 1a                                                            Figure 1b

Case Study - Stomio

In October 2019, the MRSG team performed an extensive survey with Unmanned Aircraft Systems (UAS) over the Stomio coastline in western Crete. In total, 500 with very fine resolution RGB images were acquired. In the field the identified various spots were litter was present and so this dataset was decided that it was suitable for the marine litter detection testing.

Because of high overlapping percentages, only one third of the images was used. The selected data were subseted 1024 x 1024 pixels tiles and at the end over 2500 tiles were produced. Those tiles were uploaded to Picture Pile application for sorting. Anyone interested could participate in the sorting process, where the options were YES if the person could identify litter in the image, NO if they could not and MAYBE if they were unsure.

The sorting results were obtained by the MRSG team for further processing. Firstly, each tile was converted to a vector point by using its central coordinates. These points were grouped whether they contained litter or not based on the number of YES and NO numbers. The end goal was to create a simple density map (Figure 2) in order to present the spatial density and distribution of the marine litter spottings.

Figure 2

Demo of MRSG pilot: