A Phenology-Aligned Temporal Framework Improves Satellite-Based Field-Level Wheat Grain Protein Prediction

Abstract

Abstract Satellite-based prediction of grain protein concentration (GPC) in wheat typically relies on spectral observations composited over fixed calendar windows, implicitly assuming phenological synchrony across fields. This study tests whether aligning multi-source remote sensing time series to field-specific phenology-based windows improves field-level GPC prediction. We integrated Sentinel-2 multispectral imagery (32 vegetation indices, 10 spectral bands), ERA5-Land meteorological reanalysis, gSSURGO soil properties, and USGS 3DEP topographic data, and systematically compared six temporal strategies, the factorial combination of two normalization approaches (peak-relative vs.\calendar) and three resolutions (monthly, biweekly, growth stages), across 228 commercial winter wheat fields in western Kansas (2024--2025). Three ensemble tree models (Random Forest, XGBoost, LightGBM) were trained under nested cross-validation with Boruta feature selection. Peak-relative monthly normalization achieved the highest accuracy (\((R^2 = 0.304 \pm 0.051)\), RMSE \((= 1.11)\)%), explaining an additional 5.1% of variance compared with the best calendar strategy (\((R^2 = 0.253)\)). A single 30-day post-peak window (M\((+)\)1, \((\sim)\)15--45 days after maximum canopy greenness) carried more predictive information than any broader aggregation. SHAP analysis identified topsoil organic matter, SWIR-based senescence indices (NBR2, MIRBI), and grain-filling temperature as the most influential predictors. Three-class quality classification reached 47--49% accuracy (versus 33.3% by chance), indicating practical utility for early grain segregation. While demonstrated for wheat GPC, the framework is transferable to other crop traits with temporally concentrated satellite signals, particularly those tied to specific developmental stages. The results highlight phenological alignment as a generalizable strategy for trait prediction from Earth observation data.

Plant phenotyping relevance

衛星リモートセンシング時系列を用いた小麦粒タンパク質濃度予測のため、フェノロジー整列と複数の時間集約戦略を体系的に比較・検証しており、植物形質推定手法が研究の中心である。

Code and data availability

The paper's data availability statement releases a de-identified field-level GPC dataset alongside a public authors' code repository (Ciampitti-Lab WheatGPCPipeline) implementing the data-acquisition, feature-engineering, and modeling pipeline. Both are paper-specific, public, and actionable.