Vision from the Skies - Machine Learning

Author: Benjamin Hahn, Engineer - The Plastic Tide Team.


These past years have brought about huge improvements in computer vision, that is the ability of machines to make a certain amount of sense of the world around them.  While this seems like a trivially simple task to us humans, even the task of differentiating an apple from a pear is quite hard for most computers.

Machine learning techniques such as deep artificial neural networks have enabled computers not only to distinguish different types of fruit from each other, they provide a method of teaching a computer to learn how to distinguish objects as well as any human can. The name stems from the fact that what partly inspired this technique were the biological neurones in humans’ brains.

The way this works is that instead of teaching the computer program to recognise a list of objects, the artificial neural network is fed a large number of images, each with labels indicating the position of an object and the type of object


In an iterative manner, the algorithm tries to recognise what is in each image and compares it to the labels of the image. Next, it calculates how close its guess of the image contents was compared to the actual contents (the ground truth) and adjusts its internal parameters in an attempt to improve its estimates. This is done thousands of times, with the algorithm continuously improving, i.e. learning to recognise the correct objects, eventually leading to a very capable image detector. Examples of what these sorts of algorithms have been used for are the ability of self-driving cars to spot pedestrians and road signs or computer programs detecting a person’s emotions based on his or her facial expression.

For The Plastic Tide, this algorithm will be trained on a large set of labelled aerial images of beaches, taken by a drone, where each piece of plastic in the image is tagged, using the online tagging software Zooniverse ( Since there are thousands of images to tag, this is done by a large group of volunteers, and anyone who is motivated can join in.

Using these labelled images, the algorithm will eventually be capable of detecting plastics in images, without any further aid from humans. This means, sufficiently high-res satellite images of beaches could be fed into the algorithm and it would spot the type and amount of plastics on beaches anywhere around the world. This would allow scientists to quantify the amount of plastic on beaches as well as laying the groundwork for more efficient, possibly automated beach cleans.

Taken together, The Plastic Tide does not only pique my interest as an engineer but simultaneously motivates me since it aims to solve one of the most pressing environmental problems that our world faces.