sar-transfer-glaciers — Do frozen encoders pretrained on optical images transfer to SAR? Glacier calving-front delineation on the CaFFe benchmark with frozen DINOv3 (satellite and photo), C-RADIOv4-H and two SAR-pretrained encoders, only small probes and heads trained on top, a U-Net from scratch as reference. A 0.57 M-parameter decoder on frozen C-RADIO features matches the 7.8 M-parameter U-Net on front error on the test glaciers (979 ± 31 m vs 979 m). Photo-pretrained encoders beat the SAR-pretrained ones we tried, and satellite pretraining gave no clear advantage over photo pretraining.

sar-transfer-domains — Part two of the SAR study. The same frozen pipeline, nothing tuned per dataset, on four Sentinel-1 datasets (glacial lakes, snow, forest types, alpine glaciers). With a small decoder, the satellite-pretrained DINOv3 is ahead of C-RADIOv4-H on every dataset and test split (snow IoU 0.735 vs 0.602), the reverse of the glacier study. In the one like-for-like published comparison, frozen features reach a snow F1 of 0.847 against 0.897 for a U-Net trained on the same radar channels. From one lake patch, 83 % of the 100 most similar test patches are lake for the satellite DINOv3, against 37 % for C-RADIO.

sar-transfer-distillation — Part four of the SAR study. Can a small model learn to read Sentinel-1 radar by copying several large frozen models at once? A C-RADIO-style ViT-B student is distilled from SigLIP2, satellite DINOv3, PE-Spatial and, as a fourth teacher, TerraFM-B reading the radar itself, then scored with frozen features on the four radar datasets of part two. The student does not reach the frozen satellite DINOv3, behind on all four by 0.030 to 0.055. The radar-reading teacher helps, snow IoU 0.659 against 0.603 without it and alpine glaciers 0.508 against 0.477, with ties on lakes and forest.

vggt-omega-aura-benchmark — Failure analysis of the released VGGT-Ω 3D reconstruction checkpoint on FZI-AURA, a driving dataset published after the model and so outside its training data. Depth and pose error are broken down by weather, lighting and object motion instead of averaged. Daytime depth on the held-out test split reaches AbsRel 0.082; error concentrates on thin objects, people and moving vehicles. Python reference implementation with a C++ core for the scoring.

uav-lidar-vs-satellite — Raw drone sensor logs of two MARS-LVIG flights in Armenia, a river valley and a small town, turned into DSM, DTM and land cover at 0.5 m with our own ROS 2 readers, gyro-aided poses and a ground filter written in numpy, then checked against Copernicus GLO-30, ESA WorldCover, a global canopy height map, Sentinel-2 and OpenStreetMap. GLO-30 meets the drone DSM to 0.80 m RMSE on bare ground in the valley and 0.45 m in the town, sits 0.74 m below it over buildings and metres below it over trees. WorldCover agrees on 82 % of the valley but 32 % of the town, where it calls the gardens trees and grass.

Change-Detection-Using-Dinov3 — Building change detection on frozen satellite-pretrained DINOv3 features, on SpaceNet-7 monthly imagery and LEVIR-CD. Decoder designs compared on identical features, feature differencing at 0.56M parameters and cross-attention at 0.83M, finished within 0.002 F1 of each other at 0.910 on LEVIR-CD, so the smaller head was the one to keep. Both beat the frozen embeddings read directly by roughly ten times on SpaceNet-7. Trained on a free Colab T4.

owl-caribou-overhead — Cross-herd, cross-year evaluation of four overhead animal detectors on 2,607 aerial survey patches of the Central Arctic caribou herd (Alaska, 2022), with error bars from resampled mosaics, a failure analysis and a threshold sweep. The two best models tie threshold-free (average precision 0.978 vs 0.977); the gap at the fixed setting is an operating-point effect. Half of the remaining errors sit in a 16 px band at the patch border, and most of those are real animals the patch's ground truth does not list.

bavaria-wheat-sentinel2-oco2 — Monthly NDVI and NIRv for winter wheat across the 96 NUTS-3 regions of Bavaria, 2017–2024, from 1.13 TB of Sentinel-2 L3A composites masked with yearly 10 m crop type maps, aggregated with exact partial-pixel zonal statistics on the GPU. The Sentinel-2 half of a two-person MSc capstone. A side analysis matches Sentinel-2 reflectance to OCO-2 solar-induced fluorescence footprints.