Long-term spatially explicit information on crop yield is essential for understanding food security in a changing climate. Here we present a study that combines twenty-years of Landsat and MODIS data, climate and weather records, municipality-level crop yield statistics, random forests and linear regression models for mapping crop yield in a multi-temporal, multi-scale modeling framework. The study was conducted for soybean in Brazil. Using a recently developed 30 m resolution, annual (2001–2019) soybean classification map product, we aggregated multi-temporal phenological metrics derived from Landsat and MODIS data over soybean pixels to the municipality scale. We combined phenological metrics with topographic features, long-term climate data, in-season weather data and soil variables as inputs to machine learning models. We trained a multi-year random forests model using yield statistics as reference and subsequently applied linear regression to adjust the biases in the direct output of the random forests model. This model combination achieved the best performance with a root-mean-square-error (RMSE) of 344 kg/ha (12% relative to long-term mean yield) and an r² of 0.69, on the basis of 20% withheld test data. The RMSE of the leave-one-year-out model assessment ranged from 259 kg/ha to 816 kg/ha. To eliminate the artifacts caused by the coarse-resolution climate and weather data, we developed multiple models with different categories of input variables. Employing the per-pixel uncertainty estimates of different models, the final soybean yield maps were produced through per-pixel model composition. We applied the models trained on 2001–2019 data to 2020 data and produced a soybean yield map for 2020, demonstrating the predictive capability of trained machine learning models for operational yield mapping in future years. Our research showed that combining satellite, climate and weather data and machine learning could effectively map crop yield at high resolution, providing critical information to understand yield growth, anomaly and food security.

有关作物产量的长期空间明确信息对于了解气候变化中的粮食安全至关重要。在这里，我们提出了一项研究，该研究结合了 20 年的 Landsat 和 MODIS 数据、气候和天气记录、市级作物产量统计数据、随机森林和线性回归模型，用于在多时间、多尺度建模框架中绘制作物产量图。该研究是针对巴西的大豆进行的。使用最近开发的 30 m 分辨率的年度（2001-2019）大豆分类地图产品，我们将来自大豆像素的 Landsat 和 MODIS 数据的多时相物候指标汇总到市级规模。我们将物候指标与地形特征、长期气候数据、季节性天气数据和土壤变量相结合，作为机器学习模型的输入。我们使用产量统计作为参考训练了一个多年随机森林模型，随后应用线性回归来调整随机森林模型直接输出中的偏差。该模型组合实现了最佳性能，均方根误差 (RMSE) 为 344 公斤/公顷（相对于长期平均产量为 12%）和 r²0.69，基于 20% 保留的测试数据。留置一年模型评估的 RMSE 范围为 259 公斤/公顷至 816 公斤/公顷。为了消除由粗分辨率气候和天气数据引起的伪影，我们开发了具有不同类别输入变量的多个模型。采用不同模型的每像素不确定性估计，最终的大豆产量图是通过每像素模型组合生成的。我们将基于 2001-2019 年数据训练的模型应用于 2020 年数据，并制作了 2020 年的大豆产量图，展示了训练有素的机器学习模型对未来几年运营产量图的预测能力。我们的研究表明，结合卫星、气候和天气数据以及机器学习可以有效地绘制高分辨率的作物产量图，