ValerianFourel/SOCrebuttal

    # SOCrebuttal — Soil Organic Carbon Mapping (Geoderma rebuttal artifacts)

    Architecture comparison, production maps, and supporting analyses
    for Bavaria-wide soil organic carbon mapping at LUCAS data scale
    (~16,360 samples). Published as the artifact bundle accompanying
    the manuscript revision.

    ## Headline finding

    A **lightweight (~215k-parameter) CNN + Transformer hybrid** is the
    recommended architecture at this data scale (Vanilla in the
    comparison table):

    - **+0.23 R² gain** from the CNN spatial encoder over a
      parameter-matched transformer-alone baseline (Lightweight).
    - **4× lower cross-fold variance** than transformer-alone.
    - **~3× faster convergence** (median best-epoch 7-8 vs 20).
    - **No measurable benefit** from the gated-residual mechanism on
      top of CNN+Transformer (SGT and Vanilla tie at matched
      hyperparameters; SGT is 1.7× larger for no R² gain).
    - 30× larger pure-transformer (SimpleTransformerV2, 11M params)
      matches Vanilla in mean R² but offers no other advantages.

    See `final_models/maps_comparison_2023.png` for the
    cross-architecture production-map figure, and
    `final_models/maps_comparison_2023_rebal.png` for the
    KDE-rebalanced training variant.

    ## Repository structure

    - **`sweep/`** — Spatial-CV k-fold sweep RESULTS (per-config kfold_results_summary.json, per-fold metrics, sweep_ranking, README). Excludes the heavy .pth files — request "sweep-checkpoints" for those.

• final_models/checkpoints/ — Production-mapping trained models: .pth weights + stats.json + config.json + train_log.txt per run.

• final_models/maps/ — Bavaria 2023 SOC predictions: parquet + summary.json + map.png per architecture × per band variant × per sampling mode (rebalanced and non-rebalanced).

• final_models/ — Cross-architecture comparison figures: maps_comparison_{.png,.md,.json} and the rebalanced variant.

• code/architectures/SpatiotemporalGatedTransformer/ — Spatiotemporal Gated Transformer family source: SimpleSGT, EnhancedSGT, VanillaSpatiotemporalTransformer, LightweightTransformer (the four core architectures from the rebuttal's 3-way ablation).

• code/rebuttal/ — Rebuttal scripts: spatial-CV orchestration, final-model training/inference, inspector tools, sweep orchestration, this publish script.

• docs/ — REVISION_LOG.md plus any per-folder READMEs.

    ## Methodology summary
  
    - **Spatial cross-validation**: 10-fold latitude-decile splits with
      a 1.2 km train/test buffer (Roberts 2017, Ploton 2020). The R²
      values reported in `sweep/sweep_ranking.md` are the honest
      generalization estimates.
    - **Production maps**: full-data training (95% train + 5% random
      monitor holdout — non-spatial, used only for best-epoch weight
      selection). All architectures inferred on the 1mil-point Bavaria
      reference grid for target year 2023 over a 5-year covariate
      window {2019, …, 2023}.
    - **Rebalanced production maps** (`*_rebal`): same architectures,
      retrained with KDE-inverse-density sample weighting on log(SOC)
      at α=0.5 (Yang et al. ICML 2021) to ensure organic-rich
      regions (Alpine peat, fen/bog) are not under-predicted.
  
    ## Reproducing
  
    All training, inference, and analysis scripts are in `code/`.
    Architecture source under `code/architectures/`; rebuttal pipeline
    scripts (run_kfold, train_full, infer_bavaria, sweep_submit,
    submit_finals, inspect_run, compare_maps) under `code/rebuttal/`.
  
    ## Citation
  
    Will be populated once the manuscript is accepted.
  
    ---
  
    Last updated: fourel1@jpbl-s03-02