Instructor notes#

Notes from the theoretical courses taught during the Machine Learning in Glaciology workshop, at the Finse research station (Norway).

Lecture 1: Physics-based machine learning for glacier modelling#

Authors: Jordi Bolibar, Facundo Sapienza

The presentation introduces students to the general concepts of a machine learning pipeline. How to properly design a dataset, how to correctly train models and how validate, test and understand the capabilities and limitation of the model(s).

The following contents are covered:

  • Modelling the glacier system

    • Glacier evolution models

    • Local vs Global glacier modelling

  • Physics-based machine learning

    • Machine learning pipelines

    • Regression for physical processes

      • Respecting physics

        • Feature selection

        • Data driven machine learning

        • Physical losses or Physics-Informed Neural Networks

        • Neural/Universal Differential Equations

      • Trustworthy models

        • Testing and validation

        • Physical interpretation

      • Being mindful about model limitations

  • Project description

Lecture 2: Deep Learning for remote sensing and glacier mapping#

Authors: Benjamin Robson, Konstantin Maslov and Thomas Schellenberger

The three presentations will cover:

  • Remote Sensing in Glaciology Remote Sensing in Glaciology – the traditional basics

    • Intro Optical and SAR remote sensing and their applications in Glaciology

    • Glacier extend mapping

    • Glacier zone mapping

    • Challenges

  • Random Forest and Deep learning image classification for Glacier Mapping

    • Intro to ML image classification

    • Random Forest

    • Deep learning

  • Cryospheric Mapping with Remote Sensing - an overview of the problems, data and methods (with a focus on OBIA and debris-covered glaciers

    • Object based image analysis

    • Mapping debris-covered glaciers

  • Project description The students will have the opportunity to work to mapping glaciers in High Mountain Asia using Sentinel-1 and Sentinel-2 data and pre-trained random forest and DL models as well as OBIA. They are encouraged to tune and train additional random forest models with a number of different input features and to compare the performance of the three approaches statistically.

Lecture 3: Introduction to JupyterHub and Git#

Authors: Facundo Sapienza and Ellianna Abrahams

Lecture 4: Understanding Statistical Methods for the Sciences#

Author: Ellianna Abrahams