Phenology-adaptive machine learning for early mapping of field-scale corn crop yield using fusion of Sentinel-2 satellite spectral imagery, and weather-based accumulated heat units

Abstract

Abstract Purpose Timely, accurate, and field-scale crop yield mapping is essential for precision crop management, yet most existing studies rely on late- or full-season data, limiting in-season decision-making. This study aims to develop a stage-aware earliest possible corn yield mapping framework that balances early data availability, predictive accuracy, and spatial fidelity by integrating Sentinel-2 imagery, vegetation indices (VIs), and accumulated growing degree days (AGDD). Methods Corn yield data were collected from a commercial farm over three growing seasons (2018–2020). The final modeling dataset included 51,794 Sentinel-2 10 m aggregated yield samples across three seasons: 2018 ( n = 16400), 2019 ( n = 18534), and 2020 ( n = 16860). Sentinel-2 raw spectral bands and derived VIs were organized for V4, V6, R1, R5, and R6 growth stages based on AGDD- and DAP-defined corn growth-stage windows, also confirmed visually on-ground, allowing observations from different planting and harvest dates across the three growing seasons to be aligned by crop developmental stage rather than calendar date. A stage-wise Pearson correlation and frequency-based selection identified informative and non-redundant VI subsets across crop development. Four machine learning models: Random Forest (RF), XGBoost (XGB), k-Nearest Neighbors (kNN), and a Neural Network (NNET; multi-layer perceptron) were trained using 13 input configurations, including individual growth stages and multi-stage combinations capturing phenological progression. Models were tuned via randomized search, trained on 2018–2019 data, and independently validated on the 2020 season. Yield predictions were mapped directly at 10 m Sentinel-2-pixel resolution without spatial interpolation to preserve fine-scale variability. Results Model performance was strongly influenced by phenological stage selection. Among single-stage inputs, R1 was the earliest stage where reliable yield mapping could be availed (RF: R² = 0.56, RMSE = 29.50%). While combining V6 with R1 stage inputs substantially improved predictive performance (RF: R² = 0.70, RMSE = 24.13%) for the yield mapping at the R1 stage, where V6-stage signals provided complementary yield-related information. The full-season combination (V4 + V6 + R1 + R5 + R6) produced the highest accuracy (RF: R² = 0.72, RMSE = 23.28%) but would be less suitable for early in-season decision-making. Early- or late-stage-only inputs (V4, R5, R6) showed weaker and less stable cross-year performance. Among algorithms, RF consistently generalized best to the independent 2020 dataset and is recommended for operational use. XGB showed strong training performance but reduced cross-year stability, kNN yielded moderate accuracy, and NNET achieved accuracy comparable to RF while closely reproducing observed spatial patterns such as center-pivot effects and edge gradients. Direct 10 m mapping preserved yield heterogeneity and avoided smoothing artifacts common in interpolation-based approaches. Conclusion Stage-aware feature selection and phenology-informed input combinations are critical for balancing yield prediction timeliness and accuracy. For operational in-season yield mapping, the RF model using the V6 + R1 stage combination provides a practical and reliable solution, enabling early, accurate, and spatially detailed yield estimates with robust cross-year performance. This study presents a deployable framework for integrating satellite time series and weather data into conventional (non-sequential) machine learning models to support proactive, within-season decision-making in precision agriculture.

Plant phenotyping relevance

衛星画像・気象データと機械学習を統合し、作物の収量という明示的な植物形質を圃場内10 m解像度で推定する手法を開発・独立年で検証しており、フェノタイピング手法が中心である。

Code and data availability

The paper's corn yield prediction data (Sentinel-2 features, AGDD, yield measurements) and trained models are not publicly deposited; the authors state data are available only upon request. No authors' public code or data URL is provided. Other URLs in the reference list are cited prior work or generic data portals, or

No evidence-backed public reproduction asset is currently recorded.