Portfolio

Crop yield prediction under domain uncertainty

Field-level yield models for crops that must predict a year nobody has seen before — and say honestly how wrong they might be.

Method

• Evaluation. Leave-one-year-out CV only. Random k-fold leaks across seasons and inflates R² for a model that will always be deployed into an unseen year.
• Uncertainty. Split-conformal wrapper; the calibration year is picked by nearest seasonal red-edge profile instead of a random split, and residuals are collected leave-one-year-out so no training data is sacrificed.
• Structure. National in-season forecast supplies the year-level prior; the field model predicts the residual around it, separating the climate signal from the field-to-field signal.
• Features. Sentinel-2 red-edge time series, Sentinel-1 backscatter, ERA5-Land (VPD, soil moisture, water balance), MODIS NDVI/EVI, phenology-aligned aggregation windows, yield lags.
• Models. CatBoost, XGBoost, Random Forest and TabPFN

Satellite-based soil erosion modelling

A reproducible RUSLE pipeline on Google Earth Engine that turns Sentinel-2 time series into per-parcel soil-loss rates — validated against European reference layers, not against itself.

Method

• Cover factor from measurement. Fractional vegetation cover derived from Sentinel-2 surface reflectance and weighted over the season, replacing the per-crop constants that standard assessments read from a lookup table at kilometre resolution.
• Cloud-gap reconstruction. Harmonic fit of the fCover series, benchmarked against linear interpolation and resample-plus-harmonic on real gapped series with the ungapped record as truth.
• Crop context. JRC EUCROPMAP land cover drives crop-aware C assignment.
• Validation. Four C-factor methods scored on mean, standard deviation and Pearson correlation against the JRC European C-factor layer; the R factor validated against the GloRESatE reference.
• Engineering. YAML-config driven, Dockerised, parallel over locations and years, batch processing of GeoJSON areas, exports straight to cloud storage. Output: soil loss in t · ha⁻¹ · yr⁻¹ at 100 m, EPSG:3857, plus the individual factor layers.

Embeddings under domain shift

A crop-type classifier trained on two seasons and asked to work on a third one that was 1.4 °C warmer and drier — and what self-supervised pre-training does, and does not, fix.

Method

• Shift. Sentinel-2 L1C field time series over Upper Bavaria for six crops (corn, winter wheat, winter barley, rapeseed, sugar beet, potato), 2016–2018. Train on 2016–17, predict 2018: a drought year, 1.4 °C warmer than 2016 with below-normal spring rainfall, so the spectral fingerprint of the plants moves.
• Representation. SimSiam with a transformer encoder, symmetrised negative cosine loss, then a linear classification head fine-tuned on top.
• Augmentation. Three schemes: real same-crop pairs drawn across years as positives, synthetic noise and drift, and a cloud-offset simulation. The cross-year pairing is the one that treats weather variation itself as the augmentation.
• Collapse control. The ℓ2-normalised output is tracked against 1/√dim throughout pre-training. Training loss alone did not reveal collapse; without this check the fine-tuning silently stops improving.
• Baselines. Random Forest and the same transformer trained supervised, over four label budgets for the target year: none, 5 %, 10 %, and all years pooled.

Results