Midway evaluation in Electrical and Computer Engineering - Kennedy Adriko

Event on Zoom

Title of thesis: Learning Embeddings for Earth Observation Applications with Supercomputing

Student: Kennedy Adriko

Doctoral committee:

Dr. Morris Riedel, Professor, Faculty of Industrial Engineering, Mechanical Engineering and Computer Science, University of Iceland and Jülich Supercomputing Centre, Forschungszentrum Jülich (Germany)

Dr. Gabriele Cavallaro, Associate Professor, Faculty of Electrical and Computer Engineering, University of Iceland and Jülich Supercomputing Centre, Forschungszentrum Jülich (Germany)

Dr. Claudia Paris, Assistant Professor, Faculty of Geo-information Science and Earth Observation, University of Twente

Dr. Rocco Sedona, Postdoctoral Researcher, Jülich Supercomputing Centre , Forschungszentrum Jülich (Germany)

Abstract

Crop yield forecasting at an operational scale requires extracting phenological signals from large volumes of satellite Earth Observation (EO) data. However, data storage and compute bottlenecks pose a significant obstacle. Geospatial foundation models (GeoFMs) promise a general-purpose, compressed representation of these data, but whether their embeddings remain useful for downstream forecasting across heterogeneous conditions, crop years, spatial scales, and climates remains largely untested. This PhD project asks whether compressed, general-purpose GeoFM embeddings can generalise across time, spatial scales, and climates for operational crop forecasting. This PhD research develops scalable methods for analysing large volumes of high-resolution, multi-source EO data using pretrained GeoFMs and their compact representations. Specifically, it will design phenology-aware temporal encoders that structure multi-temporal embeddings according to crop growth stages and support few-shot learning; use neural controlled differential equations to model irregular acquisition times and cloud-induced observation gaps; and develop climate-conditioned, parameter-efficient adaptation strategies to transfer GeoFMs across geographic regions and climate regimes. This work is carried out within the Earth Observation & Weather Data Federation with AI Embeddings (Embed2Scale) project, funded by the EU’s Horizon Europe programme, and is guided by the principle that embeddings should serve the application: they should provide compressed representations that remain useful for real-world crop forecasting at scale.