This content is not published

Using Deep Learning Methods for Plastic Litter Detection from Satellite Remote Sensor

Wed, 06/21/2023 - 13:01
Start Date: 
2020
Programme: 
Discovery
End Date: 
2021
Programme Reference: 
20-D-T-TEC-01-g
Country: 
France
Contractor: 
Capgemini
Status: 
Closed
Description: 

The goal of the project was to generate synthetic data sets of marine debris accumulations and to train on them AI models in view to automatically detect plastic litter accumulations in real EO images.

In order to achieve this goal, the project aimed at the following objectives:

1. To simulate plastic patches at different concentrations and shapes on real sea images coming from satellites such as Sentinel-2. The model is able to simulate the water leaving reflectance at different plastic concentrations and its effect on the spectral bands of Sentinel-2.

2. To detect marine plastic litter from simulated images from generated Machine Learning (ML) and/or Deep Learning (DL) models. The idea is to use ML/DL methods to automatically extract the specific information in spectral bands in order to detect marine plastic litter from simulated data.

3. To validate the simulated model with real plastic litter images through satellite images with marine plastic litter spots. The model, trained on simulated plastic litter images, will enable the validation with real EO images of plastic litter.

A synthetic generator software developed during the activity is available below for download under the name "Synthetic_windrows_SW.zip_.pdf" :
- right click on the link and choose "Save Link As.." 
- once the file is downloaded, remove the final file extension " .pdf" from the file name and unzip the file.
The sofware code is in Python.
Application Domain: 
Earth Observation
Technology Domain: 
16 - Optics
Competence Domain: 
5-Radiofrequency & Optical Systems and Products
Keywords: 
Remote Sensing
Marine Litter
plastic litter