UruQuantumSpace

Montevideo, Uruguay | Leveraging AI/ML for plastic marine debris

NO VACATION ISLAND

High-Level Project Summary

The main objective of this project is to propose an End-to-End solution using satellite information and Machine Learning to identify the most densely populated areas of debris and map clusters of debris to then map an optimal cleaning route and finally using aquatic drones to detect and classify debris both on the surface and in deeper parts.

Link to Project "Demo"

https://drive.google.com/file/d/1-ny2NkzZJvHf6JBzyoreBd3yhr1p9DAy/view?usp=sharing

Link to Final Project

https://drive.google.com/file/d/177ckdOoAXdWh9VQmhO9BBcUF3SRSH3HC/view?usp=sharing

Detailed Project Description

Satellite information and machine learning:

Our project includes 3 solutions using Machine Learning to attack the problem.

One of the ideas to be able to quantify ocean pollution is to generate a first map where you can see a significant accumulation of garbage and the flows of it since it accesses the sea. It seems to us that through satellite images, radars, measurement of wind speed among other data can be a good way to have this screenshot, using machine learning techniques for recognition of these foci of greater pollution or more critical foci related to more life in that area.

We have observed that through Nasa's CYGNSS constellation we have been able to obtain results, mainly about microplastics, searched in smoother waters (where the wind speed is not so high), where it is assumed that they could be and so it was.

These studies can be useful for example to be predictive, since the flows of the streams are divided monthly, repetitive behaviors can be observed and routes can be made based on this. We would know that in such a period of time garbage migrates to certain areas, and we could intervene.

There is also, for example, the company Fragata Space, specialized in transforming satellite images into high-resolution information and predictive value. It uses IoT sensors to detect the volume of waste.

The satellite images are obtained from European satellites (sentinel 1 and 2 among others). These data confirm that it is feasible to use this technology to monitor mainly large masses of waste.

For example, in the following image we can see an example from the SENTINEL-2 satellite where marine debris can be visualized.

Using the Skimage library we convert the image to black and white and together with a threshold we manage to define the following number of dots

We were able to identify up to 21 points in this example.

Obtaining location and generation of optimal routes:

Having counted them, we proceed to generate a cluster and obtain the latitude, longitude and radius of the garbage cluster and the number of objects per cluster.

These data are stored in a database in charge of mapping them with geospatial data and planning an optimal route to pass through the garbage clusters.

Then, approaching the garbage piles using boats, we proceed to deploy the aquatic drones for collection.

Deep Learning monitoring systems:

They have 2 detection systems depending on the conditions.

Firstly, an object detection model using Computer Vision, specifically using the Deep Learning architecture YOLOv5-S which obtained an average mean accuracy of 0.851 and an F1 Score of 0.89 while maintaining a near real-time speed, with this model we will be able to capture marine objects without trapping living beings.

Finally, in case of low light conditions due to the depths of the debris, it is possible to implement as an option to identify the debris by means of sonar images, which have given results of more than 90% accuracy in studies carried out.

Mapping with other information: Another part of the idea is to be able to generate another map, with information on the main migrations or concentrations of the species that are most affected annually, through the multiple studies of marine life that already exist and that can continue to be generated, and then proceed to compare both maps and observe whether the movements of these species coincide with the flows and concentration of waste in order to implement a quick solution to preserve their health and vital development.

It seems viable in these cases, to send ships with aquatic drones to be able to have a series of better detailed images, and to classify the waste, as well as a team prepared to extract the garbage.