Publikation
RegDiff: Regression Diffusion for Earth Observation
Miro Miranda Lorenz; Ashutosh Dinesh; Duway Nicolas Lesmes Leon; Francisco Mena; Marcela Charfuelan Oliva; Andreas Dengel (Hrsg.)
IEEE International Geoscience and Remote Sensing Symposium (IGARSS-2025), August 3-8, Brisbane, Australia, IEEE, 8/2025.
Zusammenfassung
Diffusion Models have emerged as powerful tools for modeling complex distributions, excelling in applications such as image generation. However, their potential in Earth Observation remains underexplored, particularly for analyzing long, multivariate time series to monitor environmental changes. These tasks are further complicated by various challenges, including sensor failures, missing data, and noisy or sparse ground truth labels. Effectively addressing these issues necessitates robust uncertainty modeling, an essential aspect often overlooked in existing research.
This work introduces Regression Diffusion (regDiff), a novel approach based on denoising diffusion for Earth Observation regression tasks using time series data. We derive the necessity for conditioning and demonstrate performance that matches or exceeds state-of-the-art methods in crop yield prediction, a fundamental task in Earth Observation. Furthermore, we demonstrate that stochastic sampling within regDiff enables robust uncertainty estimation around the conditional mean. By addressing key limitations of existing approaches, regDiff offers a promising framework for advancing time series regression in Earth Observation. We demonstrate its generalizability across three subfield-level yield datasets, encompassing multiple crop types, coupled with multivariate satellite imagery.