Postdoc: Weakly supervised ML-based Earth observation

Project Overview

The postdoc will develop an advanced Earth observation (EO) machine learning (ML) framework to identify, geolocalize, and quantitatively estimate the physical and socio-economic impacts of climate extremes such as floods, storms, wildfires, and landslides, using satellite image sequences.

A primary bottleneck in climate adaptation is that historical impact data is plagued by annotation sparsity: detailed pixel-level ground truth of damage is rare, while rich textual data from news, insurance disclosures, and humanitarian reports is unstructured, noisy, and spatially coarse. Furthermore, models trained in data-rich regions may face data distribution shifts when deployed across different geographic, infrastructural, or socio-economic contexts.

To address these challenges, the postdoc will develop annotation-efficient, weakly supervised ML-based Earth observation models that leverage unstructured data as a source of scalable distant supervision. The framework will incorporate uncertainty quantification techniques to ensure that the AI-derived impact estimates are reliable and actionable for decision-makers.

Core research task: Architecting data-efficient change detection for geospatial foundation models

Design and develop specialized, data-efficient architectural modules, built on top of frozen geospatial foundation models. The goal is to automatically detect and isolate large, extreme-event-related physical changes resulting from floods, storms, or wildfires, by modeling the contrast between pre- and post-event observations in heterogeneous, multi-sensor image sequences (Sentinel-1/2, Landsat/HLS, VIIRS/MODIS).

The development will focus on two major machine learning challenges:

• Weak supervision under annotation sparsity. Engineering models to learn robust representations of rapid environmental transitions using minimal labeled instances and sparse, noisy text reports treated as distant supervision.

• Distribution shifts and uncertainty quantification. Integrating strict geometric constraints (e.g., Spectral-normalized Neural Gaussian Processes [SNGP]) directly into the adapter layers so the model can inherently flag high epistemic uncertainty when encountering out-of-distribution geographical features or unprecedented extreme-event severities.

About Climes

The Swedish Centre for Impacts of Climate Extremes (climes) is a research and training platform building an interdisciplinary field on how climate extremes affect people, ecosystems and infrastructure in a changing world. The Centre brings together physical, medical, social and engineering sciences—disciplines rarely combined—to advance knowledge, shape the Swedish research landscape and strengthen societal resilience. Our work centres on three themes: compiling high-quality data on the impacts of climate extremes; analysing the physical–societal interactions that drive consequences; and developing policy-actionable scenarios to prepare for future extremes.

Who You Are

Beyond the specific requirements for each role, we are looking for curious, analytical individuals with a genuine passion for artificial intelligence and its potential to solve real-world problems in cross-disciplinary settings. You are a proactive problem-solver, motivated by contributing to impactful research with tangible outcomes. You combine methodological curiosity, the drive to innovate algorithmically amid real-world constraints like annotation sparsity, with the interdisciplinary agility to bridge machine learning and climate and environmental science, all in pursuit of using technology to tackle urgent challenges in climate adaptation and sustainability.

Requirements:

• A PhD in Computer Science, Machine Learning, Data Science, Applied Mathematics, Physics, or a highly quantitative equivalent.

• Strong theoretical understanding and hands-on coding mastery of deep learning architectures using PyTorch or TensorFlow, with a proven track record in computer vision or spatio-temporal modeling. Experience working with machine learning for remote sensing data.

• Excellent communication skills in English (written and spoken) are required.

Nice-to-haves:

• Prior experience handling multi-spectral geospatial formats, radar (SAR) data, or standard satellite pipelines (Sentinel, Landsat/HLS, MODIS/VIIRS).

• Familiarity with Parameter-Efficient Fine-Tuning (PEFT/LoRA frameworks), out-of-distribution detection, change-detection tasks, or single-pass Uncertainty Quantification methods (SNGP, Evidential DL, Conformal Prediction).

• Peer-reviewed publications in top-tier ML conferences or major remote sensing journals.

• Proficiency in Swedish or other Nordic languages is a welcome plus for stakeholder communication.

Are We a Good Match?

At RISE, we strive to create a workplace where diverse expertise and backgrounds converge to solve societal challenges. We offer a dynamic research environment where you will work on exciting projects with strong connections to industry and societal needs. With us, you'll have ample opportunities to learn, develop professionally, and contribute to meaningful innovation alongside experienced colleagues both within machine learning and application areas.

Welcome with your Application!

Curious to know more? For questions about specific roles, please contact Dr. Olof Mogren (olof.mogren@ri.se). The last application date is September 15th. Selection and interviews will be conducted continuously during and after the application period

Our union representatives are: Ingemar Petermann, SACO, +46 10 228 41 22 and Linda Ikatti, Unionen, +46 10 516 51 61.

Other information

Project Title: Weakly supervised ML-based Earth observation for climate extreme impact quantification

Duration: 1.5 Years

Hosted by: RISE / Climes

Location: Gothenburg