You are all warmly invited to the October EON seminar where we will hear from Sarah Maebius and Theo Ogier, two recent Environmental Data Science and Machine Learning MSc graduates from the Dept. of Earth Science and Engineering, Imperial College. Sarah and Theo will both be discussing exciting new findings from their recently completed MSc Projects, details of which are provided below.
Sarah Maebius: Sentinel-1 SAR L2 OCN wind product modelling using auxiliary ERA5 satellite maps and buoy data
Abstract – This study uses satellite imagery to obtain coastal wind speed and wind direction estimates with a high spatial and temporal resolution, which is necessary for offshore wind farm providers. A Pix2PixGAN was trained to generate Sentinel-1 SAR L2 OCN wind product with the availability of ERA-5 wind predictor variables and buoy wind speed. Wind estimates were validated against a test dataset for each of three site locations. The study found that Pix2PixGAN-generated wind speed and wind direction images had a higher accuracy estimating SAR L2 OCN wind product compared to ERA5 wind estimates. With these results, offshore wind farm providers can obtain wind estimates with a spatial resolution of 1 km at an hourly rate, which helps to optimize wind turbine location, maintenance, and performance.
Bio – Sarah completed her undergraduate in Mathematics and Statistics at Reed College, where she conducted a research project implementing satellite imagery data with ground level tree data to train supervised machine learning models and classify tree species in Portland, Oregon. An aspiring data scientist Sarah is enthusiastic for all things nature related and has a passion for wrangling, drawing conclusions from, and visualizing data, especially in an environmental context.
Theo Ogier: Mapping glacier volume loss in mountainous environments and quantifying melt with deep learning
Abstract – The aim of this study was to leverage remote sensing and deep-learning to create a glacier mapping solution that could be used to quantify glacier melt. A model was trained using exclusively open-access data, with multispectral remote sensing data from Sentinel-2 and Digital Elevation Models (DEMs) from the Advanced Land Observing Satellite (ALOS), and large amounts of target data were created rapidly using glacier labels from the Federal Office of Topography swisstopo database. The semantic segmentation model developed was based on the UNet architecture and was comfortably able to outperform conventional methods in the task of identifying glacier extents. The model was able to quantify glacier retreat on a region of the Swiss alps in line with established values, validating that the model can be used to quantify glacier melt in mountainous environments.
Bio – Theo is motivated by the vast potential of AI to solve the biggest challenges facing the planet and his primary research interest is the combined use of deep learning and remote sensing in application to the earth sciences. Theo holds a degree in Physics with Astrophysics from the University of Leeds and his professional background includes Technology and Finance.
Please e-mail Johnny Adams (j.adams19@imperial.ac.uk) if you would like to be sent the calendar invite