Detection and tracking of marine litter based on time series of high-resolution multispectral satellite images using machine learning algorithms

We suggest machine learning analysis of time series of daily PlanetScope or SkySat data to detect, quantify and track large floating plastic litter and accumulation zones. The VNIR spectral range is chosen because our spectrometer measurements have shown that a water film of 1 mm absorbs all radiation > 900 nm whereas visible light penetrates water and is reflected by submerged objects. Our goal is to obtain precise and reliable data on floating macroplastics regarding their quantity, position, floating depth, material properties and sources which may serve as a basis for later recovery and source elimination. A focus will be put on convolutional neural networks (CNNs) which are particularly suitable for the analysis of images. Our first purely data-driven approach has been tested in the Mediterranean Sea (MS) and is able to distinguish drifting items from actively moving objects like ships and from stable objects like rocks or buoys. Future tests will be conducted in the MS in cooperation with CMCC (Italy). Since material identification is limited by the 4 multispectral bands, knowledge-based computation of spectral and spatial features (e.g. spectral band ratios or shape-based descriptors) will be tested to improve object identification. Planned tasks to advance our prototype: -Compilation of additional training data -Integration of pattern recognition for marine litter signatures in SAR data (isardSAT) -Verification of detection accuracy (GFZ, isardSAT) -Integration of a drift model to forecast potential trajectories and determine best time for recovery (CMCC) -Acquisition of hyperspectral PRISMA images at the forecasted positions to identify materials -Web-based visualisation of identified items, properties and trajectories; density map -Expert interviews to develop strategies for ship-based recovery (e.g. side-/upward-looking radar/sonar) As a side effect, tracked items can be considered as drifter experiments that can improve current models through data assimilation.

Idea domain

a) Technologies
b) Data processing
f) Transport models
g) Potential services

Participants agreement

I accept all general and special conditions to participation

Idea rights

I confirm to be in possession of all required rights to submit this idea

Novel Idea

I confirm this idea is novel and has not been submitted by me in the past to ESA

Principle author

Mathias Bochow

Submitter

on behalf of my affiliation (e.g. Industry, Academia, etc.)

Affiliation Name

Helmholtz Centre Potsdam - GFZ German Research Centre for Geosciences

ESA Entity Code

1000001618

Implementation if selected

Yes

Suggested Implementation Scheme

Early Technology Development

Background

https://www.gfz-potsdam.de/en/staff/mathias-bochow
https://www.researchgate.net/profile/Mathias_Bochow
http://twei55.github.io/

No Data to Display

Jun 3, 2019
2,592 Views
115 Visitors
3 Comments
4 Followers

Detection and tracking of marine litter based on time series of high-resolution multispectral satellite images using machine learning algorithms

Detection and tracking of marine litter based on time series of high-resolution multispectral satellite images using machine learning algorithms

All authors

Abstract

Tags

2nd Round idea

STATISTICS

Detection and tracking of marine litter based on time series of high-resolution multispectral satellite images using machine learning algorithms

Detection and tracking of marine litter based on time series of high-resolution multispectral satellite images using machine learning algorithms

All authors

Abstract

Content from Previous Campaigns/Channels

Tags

2nd Round idea

STATISTICS