A method uses sparse autoencoders to split opaque Earth observation model descriptions into simpler features, ideally one concept each, such as water or vegetation health. It makes powerful AI models more transparent and understandable, building trust for monitoring a changing world.
Decomposing Earth Embeddings with Sparse Autoencoders · Vitus Benson, Fanny Yang, Markus Reichstein · arXiv:2611.11398
KuKa-Mod is a physically based model that uses radar signals in two bands, Ku and Ka, to measure ocean wind speeds from satellites. It offers a more robust and versatile tool for improving wind data from current and future Earth observation missions.
A Dual-Frequency-Constrained Structure-Function Kirchhoff Model for Near-Nadir Ocean Backscatter -- Part II: The Practical GMF, KuKa-Mod · Alexey S. Mironov, Frédéric Nouguier, Bertrand Chapron · arXiv:2611.12649
MarsNet learns how the Martian atmosphere behaves purely from spacecraft observations, instead of complex physics-based simulations. It gives planetary scientists a more accurate, independent reference for Mars's near-surface weather and climate.
MarsNet: Learning the Martian Atmosphere from Observations Alone · Sebastian Mengwall, J. Michael Battalio · arXiv:2611.12787
RainFlow uses infrared satellite images to estimate ground rainfall and produces multiple realistic rain scenarios with a measure of certainty. It is valuable for improving flood warnings and water management in regions lacking ground radar as extreme weather becomes more common.
RainFlow: Flow Matching for Infrared-Only Quantitative Precipitation Estimation · Matthieu Meignin, Cécile Mallet, Nicolas Viltard · arXiv:2611.09663
A physics-based model explains how radar signals reflect off ocean surfaces when looking almost straight down, using one description of centimeter-scale waves for two radar bands. It offers a more consistent, physically grounded way to interpret data from emerging dual-frequency radars and study ocean roughness.
A Dual-Frequency-Constrained Structure-Function Kirchhoff Model for Near-Nadir Ocean Backscatter -- Part I: Theoretical Framework · Alexey S. Mironov, Frédéric Nouguier, Bertrand Chapron · arXiv:2611.12614
Enhancing common lower-resolution ground elevation maps can accurately correct about 10-meter positioning errors along entire GEDI data collection paths. It makes GEDI forest insights more precise and useful, especially where expensive high-resolution ground surveys are lacking.
Super-resolved digital terrain models for GEDI geolocation correction: agreement with airborne LiDAR DTM-based corrections at the beam level, not the footprint level · Natsuki Yuda, Shinichiro Shirota · arXiv:2611.09854
A single AI model removes speckle noise from many kinds of Synthetic Aperture Radar images, from simple to complex polarization. It simplifies processing abundant, diverse SAR imagery and makes valuable Earth observation insights easier to extract.
Polarization Aware Deep Learning Framework for SAR Despeckling · Gianpaolo Passarello, Sergio Vitale, Hossein Aghababei et al. · arXiv:2611.10091
Gravity-sensing satellites plus a physics-constrained AI method produced a continuous global map of groundwater changes over the past two decades. It reveals significant worldwide groundwater depletion and offers data for water security, sea-level rise and Earth's rotation studies.
Reconstruction of Global Groundwater Storage Anomaly from GRACE Satellite Gravimetry Using a Physics-Constrained Decomposition LSTM · Arpon Sarkar, Showmitra Kumar Sarkar, Mafrid Haydar · arXiv:2611.08567
The effective angle of a polarizing filter subtly changes with position across a wide field of view, which could cause inaccurate measurements. Correcting this effect helps keep wide-field polarization sensors precise for uses such as space weather monitoring and Earth observation.
Foreshortening of Polarizer Angles in Wide Field Polarizing Imagers · Ritesh Patel, J. Marcus Hughes, C. E. DeForest et al. · arXiv:2611.11159
Newer language and vision models can represent world information in increasingly similar internal ways when given clean data, a shared understanding trend. It helps explain how AI builds internal knowledge and whether different AI capabilities are moving toward a more unified understanding.
Where Does Alignment Emerge and What Does It Mean? A Closer Look at the Platonic Representation Hypothesis · Julia Herbst, Simon Roschmann, Quentin Bouniot et al. · arXiv:2611.12869