A Systematic Evaluation of Spectral-Peak-Relative Temporal Alignment for Satellite-Based Field-Level Wheat Grain Protein Prediction

Abstract

Abstract Satellite-based prediction of grain protein concentration (GPC) in wheat typically composites spectral observations over fixed calendar windows, implicitly assuming phenological synchrony across fields. We present a systematic evaluation of whether aligning multi-source remote sensing time series to field-specific, spectral-peak-relative windows improves field-level GPC prediction, for a quality trait whose physiology, senescence-linked nitrogen remobilization, contrasts with the season-integrating behavior of yield. Integrating Sentinel-2 imagery (31 vegetation indices, 10 spectral bands), ERA5-Land reanalysis, gSSURGO soil properties, and USGS 3DEP topography across 228 commercial winter wheat fields in western Kansas (2024–2025), we compared six temporal strategies (peakrelative vs. calendar × monthly, biweekly, growth-stage) using three ensemble tree models under nested cross-validation with Boruta feature selection. A single 30-day post-peak window (peak + [16,45] days) was the top-performing and most consistently selected window, chosen in 4 of 5 outer folds, reproducing prior accuracy under random cross-validation (R2 ≈ 0.28); though its advantage over the best calendar window was not statistically significant (paired bootstrap p = 0.08). Under leave-county spatial cross-validation, however, this skill did not transfer across counties (Sentinel-2–only R2 ≈ 0.01; per-county median R 2 = −0.23), indicating the satellite signal supports within-region interpolation but not spatial extrapolation to unseen counties; ablation shows that neither the spectral nor the static features transfer across counties on their own, and the residual crosscounty skill emerges only from their combination. A near-real-time application at ∼3 weeks before harvest retains most within-region skill at a modest accuracy cost. The results delineate where spectral-peak-relative alignment helps, concentrating a senescence-linked signal within region, and where it does not, providing an honest operational baseline for satellite-based grain-quality monitoring.

Plant phenotyping relevance

小麦の穀粒タンパク質濃度という植物形質を対象に、衛星時系列のスペクトルピーク相対アラインメントを開発・比較評価し、交差検証で性能と空間移 transfer 性を検証しているため、方法が中心的である。

Code and data availability

The preprint explicitly releases the authors' analysis code (data-acquisition pipeline, feature engineering, cross-validation/modeling, figure scripts) at a public GitHub repository, and a de-identified field-level GPC dataset released alongside the code repository. Both are paper-specific, public, and actionable. The