YieldFormer-DE
Checkpoints of YieldFormer-DE, a model that estimates the official district yields (NUTS-3) of ten field crops in Germany at the end of each season, before the statistics are published.
Praktikum project "Transformer-based Crop Yield Prediction using Multivariate Timeseries Datasets", Chair for Computer Science IV (Computer Vision), Julius-Maximilians-Universität Würzburg. Authors: Roshan Kumar Gupta, Manisha Samal, Sreedeepthi Ramanath. Supervisor: Prof. Dr. Radu Timofte. Advisor: Dr. Fayaz Ali.
- Code: https://github.com/roshangupta00750/transformer-based-crop-yield-prediction
- Live dashboard: https://yieldformer-de.streamlit.app
- Training data: https://huggingface.co/datasets/Roshangupta007/YieldFormer-DE-data
The model
YieldFormer-DE averages three transformer networks for multivariate time series with equal weights:
| Network | Files |
|---|---|
| iTransformer | checkpoints/itransformer_v12Q/ |
| Temporal Fusion Transformer | checkpoints/tftnet_v17T/ |
| PatchTST | checkpoints/patchtst17_v17P/ |
The average is then corrected by a district error memory, each district's recent over- or underestimate. The memory is computed from earlier forecasts and has no checkpoint.
Each network was trained once per rolling season (rollYYYY, test year 2010 to 2025, and 2026 for the sealed
forecast) and seed (s42, s43, s44), using only data at least two seasons older than the test year. Every
season has two stages:
<name>_rollYYYY_sNN.ptand.pipeline.json: the network trained on all crops, with its input list and scaling;<name>F_rollYYYY_sNN_<crop>.ptand.finetune.json: the per-crop fine-tuned networks and the chosen settings.
The folders itransformer_v20Q/, tftnet_v20T/ and patchtst17_v20P/ hold version v20 (added stress inputs),
which was tested and not adopted. They are kept so that this result can be checked.
Inputs: 69 daily inputs over each crop's season (DWD station weather, ERA5-Land soil water and radiation, MODIS vegetation indices over arable land, DWD crop development stages) and 23 fixed inputs per district (SoilGrids soil, terrain, crop indicator). Target: the deviation of the yield from the district normal.
Sealed forecasts
forecasts/ holds the forecasts for 2026, 2027 and 2028, issued on 27 September, 9 October and 11 October
2026 before any official figure existed. FROZEN.sha256 lists their SHA-256 fingerprints, recorded at issue time
in the project's git history; check with sha256sum -c FROZEN.sha256.
Results
Rolling evaluation 2010 to 2025, 37,383 district-crop-years in 363 districts:
| Forecast | Accuracy (100 minus mean absolute percentage error) |
|---|---|
| YieldFormer-DE | 88.8 % |
| 5-year district average | 86.9 % |
| Long-term district normal | 86.4 % |
The gain over the 5-year average is +1.93 percentage points (90 % interval 1.04 to 3.12). It is ahead of the long-term normal in all 16 seasons and of the 5-year average in 13.
Limits
- The single Temporal Fusion Transformer with the error memory is about as accurate as the ensemble (88.84 %). The ensemble was chosen by a decision rule written before testing.
- With data cut at 30 June the model is no better than the 5-year average.
- It underestimates very large losses in single districts, and the evidence rests on 16 test seasons.
Use
The files are PyTorch checkpoints. They are loaded with the model classes of the project code (src/models/),
which also builds the input windows. They are not standalone transformers models.
Licence
The checkpoints are released under the MIT licence. The training data keep their own licences: DWD and SoilGrids CC BY 4.0, Destatis Data Licence Germany (dl-de/by-2-0), ERA5-Land Copernicus licence, MODIS data NASA LP DAAC.