Teaching Neural Networks to See Fine Urban Greenery

This talk focuses on the practical and scientific challenges of detecting fine-scale urban vegetation from Sentinel-2 satellite imagery (https://www.esa.int/Applications/Observing_the_Earth/Copernicus/Sentinel-2) under noisy and imperfect real-world conditions.

Specifically the following assumptions will be discussed:

 

•     Assumption 1: AI can solve complex problems
• Assumption 2: Deep learning math is complicated
•     Assumption 3: BIG data is needed
•     Assumption 4: Foundational models are generalizable
•     Assumption 5: AI makes Data Science superfluous

Here is a publication on the topic for those who would like to read ahead:

   -  EarthShift: a benchmark for measuring robustness to real-world distribution shifts in Earth observation (https://arxiv.org/abs/2605.29330)
    - Code and data: https://earthshift.github.io

Our speaker Dorothea is a PhD trained Biochemist who worked in the intersect of Screening and Material Sciences with industry experience. She is a bit of a #Data4Good (/en/projects) enthusiast and try to make time for it. Here is a perspective piece which came out from a project Dorothea lead:

  -   Blog Series on Learning with Uncertain Multi-Band Images — Part 1: Why Mapping Urban Greenery Was Still Hard in 2025 (https://medium.com/@edp_2023/blog-series-on-learning-with-noisy-multi-band-images-part-1-why-mapping-urban-greenery-was-72511378a3c0)

For those who are new to CorrelAid: you are very welcome to join! We can always chat 1:1 after the main event on all you'd like to know about CorrelAid and how to participate in our volunteer projects to build up your portfolio while helping others.

Please note the following:

 This event follows the CorrelAid Code of Conduct (/en/coc)

This event will be conducted in English.

There will be NO recording of the session.

AI bots for transcription are not allowed.

There are multiple channels to enroll in this event. As such the number of persons signing up on Meetup does not reflect the eventual number of persons participating.