published on 09.02.2021
Artificial intelligence makes our CollectiX particularly efficient when it comes to collecting rubbish. The algorithm "APLASTIC-Q" was developed at the German Research Center for Artificial Intelligence.
In an interview, developer Mattis Wolf explains what makes the technology so unique, how the algorithm allows conclusions to be drawn about the cause - and why drones are so important in the process.
Mattis Wolf, in the summer of 2020, we were in action in Slovakia with the algorithm "APLASTIC-Q" that you developed. How exactly can you imagine artificial intelligence (AI) on a rubbish collection boat?
Wolf: The big advantage of this technology is the efficiency gain, also in cleaning up plastic. Cameras on the rubbish collection boat and camera drones in the surrounding area provide a picture of what we are dealing with: Are there a lot of PET bottles floating in the water? Or is it mainly LDPE bags? How much organic material - dead wood, for example - is mixed in with the waste? We record this to make sorting easier and to better identify the cause of the waste.
How exactly can the algorithm recognise such details from the large mass?
Wolf: We analyse images step by step through two "convolutional neuronal networks": these are neuronal networks that work particularly well in image processing. The first neural network analyses whether the image tiles contain rubbish or not. If so, the AI estimates the number of objects. If not, it registers what can be seen instead - for example, water, vegetation or wood. Then the second neural network comes into play and evaluates the tiles that the first network has already classified as "rubbish" in a finer resolution. It then classifies the images into different types of rubbish, for example PET bottles or food packaging made of Styrofoam. This is because it is particularly important in the later efficient recycling of the plastic waste.
Welche Ergebnisse förderte die Auswertung zutage, die wir ohne den Algorithmus nur schwer oder gar nicht bekommen hätten
Wolf: First of all, we were able to record the amount of waste in numbers, which is of great value for later operations. One particular result was the waste composition in the Hron. If you look there as a human being, at first it looks just like Asian rivers. But the results showed us that there are sometimes big differences.
What are the differences?
Wolf: In European rivers, there are indications that wood is more likely to float mixed with bottles, for example, while in Asian rivers it is mainly food packaging that is common. These are "good" conditions for training artificial intelligence - the more diverse the data, the better. And the different light conditions are also helpful. While there were a lot of clouds in the sky in Slovakia at the time, we had very strong light in Southeast Asia near the equator. This poses new challenges for the algorithm because the images look different.
So you have already been on missions in Asia?
Wolf: We developed "APLASTIC-Q" as part of a project funded by the World Bank and used it as a pilot for the first time in Cambodia. There we analysed drone images of rivers, beaches and urban canals for waste and types of waste. Other projects World Bank projects in the Philippines, Vietnam and Myanmar are now building on this. The results of APLASTIC-Q are getting better: when the AI has been trained on more data, it becomes more powerful because it learns from the data.
We too will soon be cleaning up on Asian rivers and using "APLASTIC-Q" - and thanks to your technology, we will be one of the first to be able to not only clean up, but also approach local companies to establish prevention measures. So is your technology one of the first of its kind?
Wolf: Well, we haven't reinvented the wheel, artificial intelligence has been around for decades. But the approach of using two-stage AI for drone images is comparatively new. Fortunately, there are other scientists who are developing algorithms for plastic waste detection. Previous approaches have mainly focused on litter lying around on beaches. But our algorithm stands out especially when it comes to dense and very diverse litter, as is often the case in rivers.