How are reseachers using AI in Earth Observation workflows today?

Science ICESat-2
1

Hi everyone,

I’m an early-career researcher working primarily with ICESat-2 and other cryosphere datasets, and I’ve been thinking a lot about how AI is changing Earth observation research.

I’ve seen several exciting projects recently (AI-assisted coding, literature review agents, reviewer agents, etc.), but I’m curious about how researchers are actually using these tools in practice, rather than what they could do.

For a typical workflow involving remote sensing or climate datasets, there are many stages:

  • Finding appropriate datasets
  • Understanding documentation and variables
  • Downloading and preprocessing data
  • Harmonizing multiple products
  • Writing analysis code
  • Visualization and statistics
  • Literature review and interpretation

I’m curious:

  • Which of these stages consumes the most time for you?
  • Where have tools like ChatGPT, Claude, Gemini, Copilot, etc. genuinely helped?
  • Where do they consistently fall short?
  • What do you still find yourself doing manually that you wish you didn’t have to?
  • If you could automate one part of your scientific workflow (without sacrificing scientific rigor), what would it be?

For context, in my current work on Southern Ocean sea ice, some of the biggest bottlenecks which take time to resolve are identifying suitable datasets for a scientific question, understanding their conventions, and harmonizing variables across products with different grids, projections, and resolutions.

I’d especially appreciate examples from recent projects (e.g., “I spent two days figuring out projections” or “AI helped with coding but struggled with dataset-specific preprocessing”).

Thanks!

2

good timing on this, i think about this a lot from the ML side.

where AI genuinely helps me: boilerplate plumbing. xarray/dask/zarr glue code, dask cluster setup, rechunking patterns, plotting. copilot-style tools are fast at that, and theyre decent at debugging tracebacks too. drafts of docs and paper summaries as well, though i treat those as first drafts.

where it falls short is everything product-specific. itll hallucinate variable names for a dataset it never saw, invent API calls, and worst of all do something like pick a fill value or a reprojection default that silently corrupts your numbers. you still have to read the actual product docs and sanity-check against the metadata yourself. no way around that yet.

the harmonizing-grids-and-projections thing you mention is my biggest time sink too, and its the part these tools help with least, because the conventions live in scattered docs the models clearly never read.

if i could automate one thing: sanity checks after preprocessing. range checks, nan maps, quick before/after plots of the regridding, so silent errors like the ones above get caught instead of me eyeballing every step.