Precise and timely crop yield predictions at large scales are crucial for safeguarding global food security. A key factor in accurate yield forecasting is the integration of multi-source environmental data, where the choice of time windows for aggregating these variables plays a pivotal role. Segmenting time windows by phenological stages allows for more precise extraction of environmental variables, capturing the complex interactions between crop development and external factors. However, the effectiveness of this approach in improving yield prediction accuracy, especially when comparing ground-based and remote sensing-derived land surface phenology data, remains largely unexplored. In this study, we investigate how phenology-based time windows affect corn yield predictions, using machine learning algorithms and multi-source environmental data from the U.S. Corn Belt. We systematically analyzed and compared models incorporating either ground-based or land surface phenology. By segmenting the growing season into six crucial stages (BBCH 00-99) and formulating specific yield forecasting models for each stage, we determined the optimal lead times for predictions utilizing both sources of phenological data. Our findings suggested that the incorporation of phenology-derived crop growth windows significantly enhances the accuracy of yield prediction by approximately 10 % compared to the fixed-season method. Ground-based phenology data from the USDA crop progress report slightly outperformed MODIS-based land surface phenology data, achieved the highest accuracy with the XGBoost model (R2 = 0.668, RMSE = 1.09 t/ha, MAE = 0.84 t/ha). Furthermore, this study demonstrated that reliable corn yield predictions could be made as early as the second phenological stage (Emerged-Silking for ground phenology, MidGreenup-Maturity for MODIS phenology). In regions where ground observations are limited or unavailable, land surface phenology emerges as a promising alternative. This study presents a robust framework for precise and extensive crop yield modeling and early prediction, which is crucial for making informed agricultural decisions and ensuring food security.

中文翻译：

---

物候在作物产量预测中的作用：地基物候与遥感物候的比较


精确及时的大规模作物产量预测对于维护全球粮食安全至关重要。准确产量预测的一个关键因素是多源环境数据的整合，其中聚合这些变量的时间窗口的选择起着关键作用。按物候阶段分割时间窗口可以更精确地提取环境变量，捕捉作物发育与外部因素之间的复杂相互作用。然而，这种方法在提高产量预测准确性方面的有效性，特别是在比较地基和遥感衍生的地表物候数据时，在很大程度上仍未得到探索。在这项研究中，我们使用机器学习算法和来自美国玉米带的多源环境数据，调查了基于物候的时间窗口如何影响玉米产量预测。我们系统地分析和比较了结合地基或地表物候学的模型。通过将生长季节分为六个关键阶段 （BBCH 00-99） 并为每个阶段制定具体的产量预测模型，我们利用两种物候数据来源确定了预测的最佳提前期。我们的研究结果表明，与固定季节方法相比，物候衍生的作物生长窗口的结合显着提高了产量预测的准确性约 10%。美国农业部作物进展报告中的地面物候数据略优于基于 MODIS 的地表物候数据，XGBoost 模型的准确性最高（R2 = 0.668，RMSE = 1.09 吨/公顷，MAE = 0.84 吨/公顷）。 此外，这项研究表明，早在第二个物候阶段就可以做出可靠的玉米产量预测（地面物候的 Emerged-Silking，MODIS 物候的 MidGreenup-Maturity）。在地面观测有限或不可用的地区，地表物候学成为一种很有前途的替代方案。本研究为精确和广泛的作物产量建模和早期预测提供了一个强大的框架，这对于做出明智的农业决策和确保粮食安全至关重要。