Abhishek Singh

Machine learning for Earth observation and remote sensing. Berlin, Germany.

Looking for a PhD position in Earth observation, remote sensing or GeoAI, or a research role. Available immediately.

Projects

Do optical foundation models transfer to SAR?

Glacier calving-front delineation on the CaFFe benchmark with frozen encoders: satellite-pretrained DINOv3, photo-pretrained DINOv3 and C-RADIOv4-H, and two SAR-pretrained encoders. Only small probes and heads are trained on top. A U-Net trained from scratch is the reference.

A 0.57M-parameter decoder on frozen C-RADIO features matches the 7.8M-parameter U-Net on front error: 979 ± 31 m vs 979 m on the two test glaciers
Per glacier: Columbia 1,067 vs 1,080 m, Mapple 464 vs 474 m; the decoder misses 7–9 of 122 fronts per seed, the U-Net 3
Zones, 3-class patch mIoU on test: a linear probe on frozen photo DINOv3 0.610, C-RADIO 0.602, U-Net 0.594
Photo-pretrained encoders beat the two SAR-pretrained ones we tried by 0.22–0.25 mIoU; satellite pretraining gave no clear advantage over photo pretraining
Linear probes alone draw no front on 30–57% of test images: fronts need a learned head

Sentinel-1 SAR image of Columbia glacier with predicted zones coloured, the real calving front in red and the model's front in orange. — Columbia glacier, Sentinel-1, 2020. The frames cycle through the dataset’s labels and each model’s zones (rock, glacier, ocean); red is the real calving front, orange the model’s. Data: CaFFe, Gourmelon et al. 2022, CC BY 4.0. Repository.

Where VGGT-Ω fails on real driving data

A stratified failure analysis of the released VGGT-Ω 3D reconstruction checkpoint on FZI-AURA, a dataset published after the model and therefore outside its training data. Error is broken down by weather, lighting and object motion instead of averaged, with predictions written down before the runs.

Held-out test split: AbsRel 0.082, δ<1.25 = 94.3%, across 107 unseen scenes
Moving objects 0.203 against parked 0.099 and background 0.077
Counting only object interiors, moving is still worse by +0.062 in 70 of 76 scenes
Dry daylight 0.081, wet night 0.159, wet twilight 0.245
394 scenes analysed; the 20 scenes every rule was tuned on are excluded from all tables

Four frames of a test scene in four columns: camera image, predicted depth, LiDAR ground truth, and signed relative error at the LiDAR pixels. — A typical test scene at the median depth error. Left to right: camera image, predicted depth, LiDAR ground truth, signed relative error — red is predicted too far, blue too near, black rings mark moving objects. Ground truth from FZI-AURA (CoCar NextGen). Full size.

Change detection with frozen DINOv3 features

Building change detection on frozen satellite-pretrained DINOv3 embeddings, over SpaceNet-7 monthly imagery and LEVIR-CD, with decoder designs compared on identical features.

LEVIR-CD test F1 0.910, IoU 0.835 — difference head, 0.56M parameters
Cross-attention head, 0.83M parameters: F1 0.908. The smaller head was the one to keep
SpaceNet-7 monthly change F1 0.178, building IoU 0.704 — a far harder task
Training-free embedding baselines reach F1 0.013–0.019, so the training carries the result

Two LEVIR-CD test pairs in four columns: the before image, the after image, the ground truth change mask, and the prediction coloured green for true positives, red for false positives and blue for missed change. — Two LEVIR-CD test pairs. Left to right: before, after, ground truth change mask, and the difference head’s prediction — green is correct, red a false alarm, blue missed change. Notebook.

Overhead caribou detectors across herds and years

Cross-herd, cross-year evaluation of four overhead detection models (one caribou-specific) on aerial survey imagery of the Central Arctic Herd, Alaska, 2022, after development on the Porcupine Herd, 2017. Error bars from resampled mosaics, a failure analysis and a threshold sweep.

Locked test domain: 2,607 patches from 43 mosaics, 12,456 annotated animals
The two best models tie threshold-free: average precision 0.978 vs 0.977, paired difference +0.001 [−0.004, +0.008] over 2,000 resamples
The gap at the fixed setting is an operating-point effect: one model holds F1 above 0.95 for any relative cut from 0.15 to 0.55, the other only from 0.35 to 0.60
A 16 px border band is 12% of the area but holds 53% of the errors; at full resolution most are real caribou the patch’s ground truth does not list
The two remaining models trail clearly: average precision 0.927 and 0.919

Three panels: precision-recall curves of four models on full patches, the same on patch interiors, and the share of empty patches with a false alarm as recall rises. — Left: precision-recall of the four models on full patches, dots mark the fixed evaluation setting. Middle: the same with the border band left out. Right: false alarms on empty ground stay near zero for the two best models until recall passes about 0.96. Repository.

Winter wheat vegetation indices for Bavaria

Monthly NDVI and NIRv for winter wheat across every district of Bavaria, from Sentinel-2 composites masked with yearly 10 m crop type maps and aggregated with exact partial-pixel zonal statistics. The Sentinel-2 half of a two-person MSc capstone.

96 NUTS-3 regions, March to June, 2017–2024
1.13 TB of input imagery, 4,768 GeoTIFFs across 29 MGRS tiles per month
1.64 billion winter wheat pixels at 10 m, processed in 5.5 h on one A100
Over 151,793 OCO-2 footprints, solar-induced fluorescence tracks NIRv (r = 0.54) better than NDVI (r = 0.42)

Line chart of winter wheat NDVI for March to June, one line per year from 2017 to 2024, peaking in May. — Winter wheat NDVI across Bavaria, March to June, one line per year from 2017 to 2024, with error bars over the 96 districts. Every year peaks in May. Notebook.

Master's thesis

Carried out at the DLR Earth Observation Center, Oberpfaffenhofen: detection of greenhouses and plastic-covered parcels in southern Germany from 20 cm aerial orthophotos, with no hand-drawn annotation anywhere in training. Labels were derived from EU parcel-level crop declarations and refined against the imagery. The detector combines a frozen satellite-pretrained DINOv3 backbone, decoded back to 20 cm by guided feature upsampling, with a trainable ResNet-34 branch and a small fusion head. Supervised by Dr Ursula Gessner (DLR) and Prof Dr Iftikhar Ahmed (University of Europe for Applied Sciences). A paper is in preparation.

How I work

Most of my projects derive training labels from registers or benchmarks that were never made for the purpose, then deal with the biases those labels carry. A result that has come back in every project so far: past a small model budget, improving the data moves accuracy further than adding model capacity does, and I measure both sides rather than assume it. Pipelines are numbered stages with pinned environments so someone else can rerun them, and evaluation is on sites and conditions held out from training, not on random splits.

Contact

GitHub: saverin0
LinkedIn: abhishekzsingh