We comprehensively examine methodologies tailored for subnational crop yield and production forecasting by integrating Earth Observation (EO) datasets and advanced machine learning approaches. We scrutinized diverse input data types, cross-validation methods, and training durations, focusing on maize production and yield predictions in Burkina Faso and Somalia. Central to our analysis is the comparative assessment of using time-invariant features within a panel data (PD) model versus a time-series data (TD) model. The TD model performed well in predicting both production and yield, while the PD model offered comparable yield predictions. Time-invariant features such as livelihood zones, soil properties, and cropland extents enriched the spatial understanding of crop data, enhancing the R-squared by 0.09 (0.21) for production and 0.11 (0.03) for yield, with corresponding reductions in the Mean Absolute Percentage Error by 90 % (238 %) for production and 5 % (4 %) for yield in Burkina Faso (Somalia). While Burkina Faso's consistent crop data allowed for effective modeling with brief training, Somalia benefited from the adaptability of the PD model to crop statistics outliers, particularly with extended training in high-producing regions. The PD approach showed promise in addressing data gaps, although predicting crop productions for unobserved districts remained a challenge. Our findings highlight the harmonious integration of EO data and machine learning in the field of agricultural forecasting and emphasize the importance of region-specific methodologies, especially in the rapidly changing landscape of EO data convergence.

中文翻译：

---

撒哈拉以南非洲地方作物预测面板和时间序列数据模型的对比


我们通过整合地球观测 （EO） 数据集和先进的机器学习方法，全面研究了为地方作物产量和产量预测量身定制的方法。我们仔细研究了不同的输入数据类型、交叉验证方法和训练持续时间，重点关注布基纳法索和索马里的玉米生产和产量预测。我们分析的核心是在面板数据 （PD） 模型与时间序列数据 （TD） 模型中使用时不变特征的比较评估。TD 模型在预测产量和产量方面都表现良好，而 PD 模型提供了相当的产量预测。生计区、土壤特性和农田范围等时不变特征丰富了对作物数据的空间理解，将产量的 R 平方提高了 0.09 （0.21） 和 0.11 （0.03），在布基纳法索（索马里）的生产平均绝对百分比误差分别降低了 90 % （238 %） 和 5 % （4 %）。虽然布基纳法索一致的作物数据允许通过简短的训练进行有效的建模，但索马里受益于 PD 模型对作物统计异常值的适应性，尤其是在高产地区进行扩展训练时。PD 方法在解决数据差距方面显示出希望，尽管预测未观察到的地区的作物产量仍然是一个挑战。我们的研究结果强调了 EO 数据和机器学习在农业预测领域的和谐整合，并强调了特定区域方法的重要性，尤其是在 EO 数据融合的快速变化环境中。