Multi Objective AutoML Towards Accurate and Robust EO models

On 17 September 2025, the ESA Φ-Lab Collaborative Innovation Network hosted a technical seminar featuring Dr Jan N. van Rijn, Assistant Professor at Leiden University. He co-leads the Automated Design of Algorithms (ADA) group within LIACS, where his research focuses on developing methods and evaluation frameworks that help democratise the use of artificial intelligence.

Talk Overview

Earth Observation (EO) models face strict requirements: they must be accurate, lightweight enough for deployment on satellite hardware, and robust to domain shifts that occur across time, geography, and sensor types.

Automated Machine Learning (AutoML) has proved valuable in supporting data scientists by selecting suitable architectures and optimising hyperparameters. This allows experts to focus on higher-level tasks. However, as highlighted during the Horizon TAILOR project—where ESA contributed informally—there is a growing need for AutoML systems not only to improve accuracy but also to ensure trustworthiness.

In this seminar, Jan presents:

Why EO models must be robust to input perturbations and domain variability

How current AutoML techniques fall short in addressing robustness criteria

Insights from TAILOR project research on strengthening neural network robustness

Which hyperparameters matter most when optimising for trustworthiness

The session offers a clear view of how AutoML can evolve to meet the demands of real-world EO applications.

🔔 Subscribe for more technical seminars from the ESA Φ-Lab Collaborative Innovation Network.

#AutoML #EarthObservation #TrustworthyAI #Robustness #MachineLearning #NeuralNetworks #Philab #CIN #pischool