QDANN (reimplementation)

A clean-room reimplementation of the Quantile-loss Domain Adversarial Neural Network from:

Ma, Y., Liang, S.-Z., Myers, D.B., Swatantran, A., Lobell, D.B. (2024). Subfield-level crop yield mapping without ground truth data: A scale transfer framework. Remote Sensing of Environment 315, 114427. https://doi.org/10.1016/j.rse.2024.114427

Not affiliated with or endorsed by the authors. The published model code was never released; this is written from the paper alone. The authors' yield maps are distributed separately under CC-BY-NC-SA 4.0 and are not used as an input here, this implementation trains only on public-domain data (USDA NASS Quick Stats, Landsat, gridMET, USDA CDL), so its outputs carry no non-commercial restriction.

Status

Model and pipeline implemented; the source domain is built and scored, the target domain is not finished.

• Source table: data/source_maize.parquet, 99 Iowa counties, 2008-2018, 27 features.
• County holdout: mean R2 0.640 over five seeds, +0.172 over a per-year mean. Numbers and their caveats in notes/county_holdout_scores.md.
• Adversarial branch: verified end to end on a synthetic domain shift with known target labels; on real pixels it awaits the target feature table (data/gridmet_target_maize.parquet is the weather leg only, the GCVI leg is an unrun Earth Engine job).
• Not yet: trained weights, subfield yield maps, crops other than maize.

Layout

Path	Contents
SPEC.md	Every equation and architectural claim in the paper, transcribed to pseudocode with section references
AMBIGUITIES.md	Details the paper leaves unspecified, what was assumed, and expected impact
config.toml	Hyperparameters, each tagged paper_specified / reasonable_assumption / implementation_choice
notes/	Measured results, with what each table does and does not support
src/qdann/nass.py	County-year yields from USDA NASS Quick Stats: the source labels
src/qdann/gee.py	County GCVI composites from Landsat Collection 2, via Earth Engine
src/qdann/harmonics.py	Harmonic regression over a GCVI series (Eq. 2)
src/qdann/weather.py	Monthly gridMET weather, and the joined source table
src/qdann/target.py	The unlabeled target domain: CDL-masked crop pixels
src/qdann/model.py	Feature extractor, yield predictor, domain discriminator, gradient reversal
src/qdann/losses.py	Quantile and domain losses, and the quantile reweighting (Eqs. 5-11, 16)
src/qdann/vae.py	The VAE data filter (Eqs. 12-16)
src/qdann/train.py	Training loop, county holdout, scoring, CLI
src/qdann/baselines.py	Ridge, random forest, and plain DNN comparisons (Section 4.2)
src/qdann/ablation.py	The component ablation of Fig. 16
src/qdann/synth.py	Synthetic domain shift with known target labels, to check QDANN end to end
tests/	Shapes, gradient signs, loss asymmetry, holdout leakage, and pipeline wiring

Development

uv sync
uv run pytest -v

Torch is pinned to the CPU build (torch==2.9.1+cpu) via an explicit index — the model is ~50k parameters and never needs a GPU.