Towards Global Monitoring of Glaciers with Deep Learning.

Overview

Glaciers are a key indicator of climate change and have shrunk rapidly across the world in recent decades as a result of global warming (Zemp et al., 2019). Quantifying changes in glacier dynamics is not only vital for monitoring the impacts of climate change, but also for assessing their potential to generate hazards, such as glacial lake outburst floods, and to sustain water supplies during dry periods. Thus, mapping global glacier change provides vital information for estimates of global sea level rise and regional to local scale impacts of glacier shrinkage (Stokes et al., 2018). This project aims to deploy Deep Learning models to comprehensively scan Sentinel-2 and landsat-7,8 images (30 year timespan) and develop an approach to rapidly and efficiently generate glacier inventories over the last few decades and into the future.

Methodology

This project will combine Google Earth Engine and Deep Learning in order to develop an automated method to generate glacier inventories, primarily focussing on mountain glaciers and ice caps. Google Earth Engine (GEE) delivers an unprecedented access to the data archives for the Landsat and Sentinel programs. Using the GEE API it becomes possible to programmatically access satellite tiles for the entire world and for any data available in the archive. The resulting volume of available data will require a combination of automation and reliability that can only be delivered by Deep Learning (DL). This project will use a 2-part DL workflow: 1) the project will develop a DL object localisation algorithm that is capable of identifying the presence of glaciers in a given satellite tile; 2) the project will refine an existing DL workflow, SEE_ICE (available on GitHub), in order to perform semantic classification of detected glaciated landscapes. Training and validation data for the DL architectures will be derived from previous regional glacier inventories (e.g. Stokes et al., 2018) and the Randolph Glacier Inventory (RGI: Pfeffer et al., 2014) combined with extensive manual digitisation. The code will be developed in Python using the Tensorflow DL library.

Project Timeline

Year 1

Assembly of training and validation data
Preliminary work on the object localisation network architecture
Refinement of the semantic classification network architecture.
Integration of the GEE API in to deep architecture pipelines.
Python and machine learning training delivered by the Durham University Advanced Research Computing facility.

Year 2

Student further training at the Karthaus Glaciology Summer school.
Finalisation and training of the DL models.
Global scale validation of the models

Year 3

Global scale deployment of DL workflow
Compilation and dissemination
EGU conference in the spring.

Year 3.5

Writing and AGU conference in December

Training
& Skills

Deep learning, Machine Learning, Python

References & further reading

Pfeffer, W.T. et al. (2014) The Randolph Glacier Inventory: a globally complete inventory of glaciers. J. Glaciol. 60, 537–552. https://doi.org/10.3189/2014JoG13J176

Stokes, C.R. et al. (2018) Widespread and accelerating glacier retreat on the Lyngen Peninsula, northern Norway, since their ‘Little Ice Age’ maximum. Journal of Glaciology, 64 (243), 100-118.

Zemp, M. et al. (2019) Global glacier mass changes and their contributions to sea-level rise from 1961 to 2016. Nature, 568, 382-386.

Marochov, M., Stokes, C. R., and Carbonneau, P. E.: Image Classification of Marine-Terminating Outlet Glaciers using Deep Learning Methods, The Cryosphere Discuss. [preprint], https://doi.org/10.5194/tc-2020-310, in review, 2020.

Useful websites:

The Randolph Glacier Inventory website: https://earthobservatory.nasa.gov/images/83918/the-randolph-glacier-inventory.

Google Earth Engine: https://earthengine.google.com/
The SEE_ICE Github project : https://github.com/PCdurham/SEE_ICE

Further Information

Apply Now