Tutorial: Machine learning with ICESat-2 data (Room 345)

[JessicaS11]

Lead: Wei Ji Leong
Date: 21/08/2024
Start Time: 0900
Duration: 60
Description:

Details

Learning Outcomes

• understanding how to convert ICESat-2 point cloud data into an analysis/ML-ready format
• learn the different levels of complexity of ML / AI approaches and the benefits/challenges of each
• learn the potential for using ML / AI for ICESat-2 photon classification

People Developing the Tutorial (content creation, helpers, teachers)

Wei Ji Leong

Summary Description

A machine learning pipeline from point clouds to photon classifications

Dependencies (things people should know in advance of the tutorial)

• Knowledge of how to download ICESat-2 (ATL03?) data, and maybe also get coincident imagery data

Technical Needs (GPUs? Large file storage? Unique libraries?)

• NVIDIA T4 GPU

[weiji14]

Looked through some of the survey responses last week, and did a bit of thinking on what might make sense for an ICESat-2 Machine Learning tutorial. There's a range of interests (many research topics) and skill levels (from Masters/PhD students to Postdocs with years of experience), so a little tricky to gauge what an interesting topic and difficulty level would be. Doing a braindump of possible combinations for now:

Task:

• General topic like photon classification of different surface types, relevant for:
  • Land ice vs supraglacial lake
  • Sea ice vs polynya water
  • Land vs shallow water (coastal areas)
  • Vegetation types (grassland/forests) using ATL08
• More niche topics like:
  • Climate/weather forecasting (maybe better for GeoSMART track?)
  • Snow depth estimation (SnowEx track?) or bathymetry estimation (ICESat-2 track?)

Methods:

• Using scikit-learn / CuML (easy)
  • Supervised learning - Random Forest or other decision tree-based methods
  • Unsupervised Learning - K-means, DBSCAN
• Using pure Pytorch (medium)
  • Feedforward artificial neural network
  • Convolutional Neural Network
• Using Pytorch with a framework like Lightning, Huggingface, etc (hard)
  • Graph-based neural network
  • Transformers
  • State Space Models

[NCristea]

Hi Wei Ji! Thank you for your valuable comments and ideas. We will be having an ML tutorial on Tuesday, that will cover a decision tree method (probably SWE) and ANNs (probably snow depth). If you would like to further develop your ideas, please don't hesitate to contact us. We are reaching out to tutorial leads to discuss/refine the tutorial content.

[ZihengSun]

Thanks Wei Ji. I would vote for the Using Pytorch with a framework like Lightning, Huggingface, etc and try to make it easier by using AutoML and wrap all the math and calculations in a few lines, like calling transformers library of hugging face. I can help if needed.

[weiji14]

Thanks for chiming in! I was looking at #39 and didn't want there to be too much of an overlap in terms of tasks/methods. I've already scheduled a meeting with Jessica actually to discuss about what ICESat-2 topic would be suitable, but could set up another one with you @NCristea and @ZihengSun to maybe talk about what ML method to teach.

[ZihengSun]

@weiji14 yeah sure, I would love to chat about it.

[JessicaS11]

Also happy to expand our meeting invitation if that day/time works for all!

[NCristea]

Great - I am usually available in the morning - would Monday 9am work?