arXiv subjects

Lukas Brunner

Publications and source records attributed to Lukas Brunner.

3 recordsLinked to original sources

arXiv2026

Evaluating local climate in global storm-resolving models with the K\"oppen-Geiger classification

Global storm-resolving models aspire to become digital twins of the Earth, delivering information at the local scale at which humans experience climate. We evaluated how well two such models, ICON and IFS-FESOM, reproduce the climate as classified by the K\"oppen-Geiger system, using 30-year (2020-2049) simulations from the nextGEMS project at 9~km global resolution under SSP3-7.0 scenario. Both models capture the global distribution of the five main climate categories, encouraging given the infancy of storm-resolving climate modelling. Substantial regional biases nonetheless remain. Both underestimate tropical rainforest (Af) extent due to insufficient dry-month precipitation in Amazonia and equatorial Africa. ICON almost eliminates hot arid desert (BWh) across Australia through excessive precipitation, while IFS-FESOM reproduces it well. The two models show opposing biases along the temperate--continental boundary: IFS-FESOM winters are too cold in western Europe, ICON winters too warm. Substituting observed temperature or precipitation into the model fields reveals that precipitation errors dominate misclassification, while temperature biases play a secondary role confined to mid-latitude climate zone boundaries. Under climate change, the two models and CMIP6 projections agree on the direction of climate zone shifts: expansion of tropical savanna and hot desert at the expense of subarctic, tundra, and ice cap zones. However, inter-model differences in present-day climate exceed the 30-year climate change signal for many zones, calling for caution in regional projections and adaptation planning. Our results expose where local-scale climate representation still falls short of the digital twin ambition, while confirming that storm-resolving models already perform well across many regions. We propose K\"oppen-Geiger classification as a standard diagnostic to help track further progress.

physics.ao-ph↗

arXiv2025

Through-Foliage Surface-Temperature Reconstruction for Early Wildfire Detection

We present a method to reconstruct surface temperatures through forest vegetation by combining signal processing and machine learning, enabling fully automated aerial wildfire monitoring with drones for early fire detection. Synthetic aperture (SA) sensing reduces canopy occlusion but introduces thermal blur. To overcome this, we train a visual state space model to recover subtle thermal signals of partially occluded soil and fire hotspots from blurred data. To address limited real-world training data, we generate realistic surface temperature simulations using a latent diffusion model, temperature augmentation, and procedural thermal forest modeling. On simulated datasets, our method reduces RMSE by 2-2.5 versus conventional thermal and uncorrected SA imaging; in field experiments on hotspots, RMSE improved by 12.8-fold and 2.6-fold, respectively. Our approach also generalizes to other thermal signals, including human signatures, capturing morphology and extent -- critical where simple thresholding fails -- while conventional imaging struggles with partial occlusion.

cs.CV↗

arXiv2020

BlendTorch: A Real-Time, Adaptive Domain Randomization Library

Solving complex computer vision tasks by deep learning techniques relies on large amounts of (supervised) image data, typically unavailable in industrial environments. The lack of training data starts to impede the successful transfer of state-of-the-art methods in computer vision to industrial applications. We introduce BlendTorch, an adaptive Domain Randomization (DR) library, to help creating infinite streams of synthetic training data. BlendTorch generates data by massively randomizing low-fidelity simulations and takes care of distributing artificial training data for model learning in real-time. We show that models trained with BlendTorch repeatedly perform better in an industrial object detection task than those trained on real or photo-realistic datasets.

cs.CV↗