Machine Learning and Aquatic Remote Sensing

WHEN: Sunday, September 13, 09:00 – 13:00
WHERE: Wintercircus Auditorium & Foyer
LED BY: Roy El Hourany (LOG, Université du Littoral d’Opale) and Mortimer Werther (Eawag)

This workshop brings together machine learning and aquatic remote sensing (open ocean, inland, and coastal waters). The workshop is split into three parts and will last approximately 3 hours (including a break).

Part A will start with a concise “flash course” introducing key machine learning concepts and algorithms commonly used in this field. Participants will work with satellite and in-situ datasets centered on the retrieval of phytoplankton pigments and community structure from space using remote-sensing reflectance and contextual datasets. Through this exercise, they will learn how to prepare and harmonize the datasets, build and validate models, and interpret the results. Part A will take about 1 1/2 hours.

Part B builds on the models and concepts introduced in Part A to extend the analysis to model evaluation. It examines how models generalize beyond the data seen during training, how to quantify estimation uncertainty, and how to assess and improve calibration so that estimated values correspond accurately to observed outcomes. We will explicitly cover how methodological choices (e.g., training strategy, validation design) influence reported model performance, and how we can provide a measure of uncertainty during applications to satellite imagery. We will also demonstrate how to report these elements in a transparent and reproducible manner. Part B will last about 1 hour.

Part C will focus on distinguishing machine learning from artificial intelligence (AI) in this field. This part will provide participants a brief review of current research topics and trends, and examines how AI may influence the field in the future. This last part is designed to take approximately 30 minutes.

For Part A and B programming knowledge is highly recommended and appreciated, but not needed to participate. The course materials will be provided in Python (open source). Only Part A will also be available in MATLAB. Each participant should bring their own device (e.g., laptop) to follow along. The Jupyter notebooks will also be provided for the session for you to access the datasets and all involved code (and, if possible, shared in advance as pre-workshop resources). Setting up a Jupyter notebook (https://jupyter.org/install) is required before the beginning of the workshop (contact us beforehand if you require assistance).

SHORT COURSE

Working with Aquaverse and STREAM: Atmospheric Correction, Water Quality, and Uncertainty Products for Inland and Coastal Waters

WHEN: Sunday, September 13, 09:00 – 13:00
WHERE: Wintercircus Catacombs
LED BY: Ryan E. O’Shea, Arun M. Saranathan, Akash Ashapure, and Will Wainwright
(Science Systems and Applications, Inc., and NASA Goddard Space Flight Center)

This course will cover the fundamentals of working with Aquaverse, a machine-learning-centered processing workflow to generate downstream products from remote sensing observations. Aquaverse is an aquatic inversion scheme for the estimation of biogeochemical parameters and inherent optical properties, as well as their associated prediction uncertainties, from remotely sensed optical data over inland and coastal waters. Pre-processed chlorophyll-a, total suspended solids, and Secchi disk depth products are available from Sentinel 2’s Multispectral Instrument (MSI) and Landsat-8/9’s Operational Land Imager (OLI/OLI-2) for CONUS and select regions around the globe via our web interface https://ladsweb.modaps.eosdis.nasa.gov/stream/. Pre-processed products will be available for download, for global inland and coastal waters, from the Ocean Color Instrument (OCI) aboard the PACE mission. Alternatively, similar water quality products can also be estimated for most multi- and hyperspectral missions, including Sentinel 2’s MSI, Landsat-8/9’s OLI/OLI-2, Sentinel-3’s Ocean Land Colour Instrument (OLCI), the Earth Surface Mineral Dust Source Investigation (EMIT), Planet’s SuperDove, the Moderate Resolution Imaging Spectroradiometer (MODIS), the Visible Infrared Imaging Radiometer Suite (VIIRS), Medium Resolution Imaging Spectrometer (MERIS), the Hyperspectral Imager for the Coastal Ocean (HICO), PACE’s OCI, and the PRecursore IperSpettrale della Missione Applicative (PRISMA), via user-installable machine learning models. Aquaverse machine learning models for atmospheric correction of MSI and OLI imagery are also supported via these tutorials.

First, the course will introduce use cases of Aquaverse products via maps, scatterplots, and timeseries covering dynamic inland and coastal regions, including Lake Erie, the Chesapeake Bay, and small ecosystems across Europe. Second, the course will lead users through downloading, loading, and timeseries products of the available online products. Third, we will demonstrate installation and example scripts for the user-installable machine-learning model, termed Mixture Density Networks (MDN), intended for AC and WQP retrieval from supported multi- and hyperspectral sensors. Finally, there will be time for users to interact with course instructors to (1) learn the fundamentals of MDNs, (2) install the MDN, (3) run example scripts, and (4) discuss potential projects, applications, and user needs. Users are encouraged to bring their own laptops to the workshop to install the MDN toolbox (available at https://github.com/STREAM-RS/STREAM-RS).

SHORT COURSE

Polarization Optics in the Aquatic Environment

WHEN: Sunday, September 13, 13:30 – 17:30
WHERE: Wintercircus Expo #2
LED BY: Tristan Harmel (Magellium), Olivier Burggraaff (NPL), Robert Foster (NRL), and Amir Ibrahim (NASA)

As an electromagnetic wave, light can be defined by its amplitude (intensity), frequency (wavelength), and state of polarization. Each of these components provides important insights on the interaction of light with matter. In the natural world, certain species of insects and crustaceans can sense polarization, while other animals utilize it for camouflage. Within aquatic environments, polarization is everywhere—from atmospheric skylight to in-water radiation, to reflection and refraction by the air-sea interface.

This course attempts to give a brief overview of recent developments on the use of polarization for aquatic environment monitoring including assessment of aerosols, atmospheric correction, reflection and transmission at the water surface, distinct signature of optically active water constituents. After a short historical review of the successive discoveries that punctuated our understanding of light polarization in the aquatic environment, the following topics will be addressed:

• Basic polarization theory in light scattering and radiative transfer
• Ways that polarization is generated in atmosphere-water systems (scattering, Umov effect, reflection/refraction)
• Modeling of polarization inherent properties through the scattering (Mueller) matrix of phytoplankton, suspended sediments, and other artificial matters (micro-plastics)
• Polarization to help disentangle elastic and inelastic (e.g., fluorescence) light contribution
• Correction of reflected light for above-water radiometry
• Simulations from the water column to the top-of-atmosphere level with vector radiative transfer model (e.g., OSOAA)
• Visualization of the polarization parameters from in-situ measurements (e.g., polarimetric cameras)
• Download and use of PACE polarimeter data, comparison with OSOAA model and/or measurement data from cruises/airborne instruments

Skills in Python programming and use of Jupyter notebooks will be beneficial to practical examples.