Accurate forecast of corn yields is important for decision making regarding food and energy management strategies. In this work, we developed an unprecedented optimized framework for the MODIS-based mid-season corn yield forecasting over five producing states of the United States: Illinois, Indiana, Iowa, Nebraska, and Ohio. We evaluated Enhanced Vegetation Index (EVI)-based forecasts under different schemes accounting for: different machine learning techniques with a mid-season composite or multi-temporal composites as inputs, four (county-, district-, state-, and global-based) training domains, and 16-day composites versus daily interpolated composites involving the day of pixels as predictors. Under the best identified scheme, we compared the EVI-based forecasts with those based on Normalized Difference Vegetation Index (NDVI), Leaf area Index (LAI) and Fraction of Absorbed Photosynthetic Active Radiation (FPAR). EVI and NDVI were transformed to LAI (named as LAI_EVI and LAI_NDVI) as predictors for producing EVI- and NDVI-based forecasts. We evaluated both county- and state-level forecasts using the percent error (PE), mean absolute PE (MAPE) and determination coefficient (R²). The linear regression models driven by the single latest composite in mid-season often outperformed elastic net and random forest models driven by multi-temporal composites. The forecast performance decreased with longer subsets of EVI composites being used. The performance under the different training domains varied by states and forecast level (county or state), although the changes within states were mostly non-significant except in Nebraska. The forecasts based on 16-day and daily composites performed similarly, indicating that the use of information about the day of pixel composite provides no additional benefit to the yield forecast. For the best EVI-based schemes, the medium annual MAPE (PE) at the county (state) level varied between 6.1% (2.4%) and 7.7% (5.3%) across states while the medium annual R² (interannual R²) varied between 0.54 (0.59) and 0.82 (0.86). Results suggested that, while EVI was, in general, the best predictor for the Corn Belt as a whole, the adequacy of the EVI- and NDVI-based forecasts varied by states and largely exceeded that of the LAI- and FPAR-based forecasts. Compared with the EVI-based forecasts, the NDVI-based forecasts performed better in Iowa (MAPE’s 0.9% and 1.43% lower at the county and state level), similar in Nebraska and worse in the other states. Overall, the best state-level forecasts consistently outperformed concurrent National Agricultural Statistical Service (NASS) forecasts.

玉米单产的准确预测对于食品和能源管理策略的决策至关重要。在这项工作中，我们为基于MODIS的季中玉米单产开发了前所未有的优化框架，该单产预测了美国五个产州：伊利诺伊州，印第安纳州，爱荷华州，内布拉斯加州和俄亥俄州。我们评估了在不同计划下基于增强植被指数（EVI）的预测，其原因包括：不同的机器学习技术，以中期复合或多时相复合作为输入，四种（基于县，地区，州和全球） ）训练域，以及以像素天为预测指标的16天合成视频与每日插值合成视频的比较。在最佳识别方案下，我们将基于EVI的预测与基于归一化植被指数（NDVI）的预测进行了比较，叶面积指数（LAI）和吸收的光合作用主动辐射的分数（FPAR）。EVI和NDVI被转换为LAI（命名为LAI_EVI和LAI _NDVI）作为产生基于_EVI和_NDVI的预测的预测器。我们使用百分比误差（PE），平均绝对PE（MAPE）和确定系数（R ²）。在季节中期，由单个最新合成函数驱动的线性回归模型通常优于由多时相合成函数驱动的弹性网和随机森林模型。随着使用更长的EVI复合材料子集，预测性能下降。尽管各州内部的变化除内布拉斯加州外大部分无意义，但不同培训领域的绩效因州和预测水平（县或州）而异。基于16天和每日合成的预测效果相似，这表明使用像素合成的当天信息不会给产量预测带来额外的好处。对于基于EVI的最佳计划，各州在县（州）的中等年度MAPE（PE）在6.1％（2.4％）和7.7％（5.3％）之间变化，而中型R ²（年度R ²）在0.54（0.59）和0.82（0.86）之间变化。结果表明，尽管总的来说，EVI是整个玉米带的最佳预测指标，但基于EVI和NDVI的预测的适当性因州而异，大大超过了基于LAI和FPAR的预测。与基于EVI的预测相比，在爱荷华州，基于NDVI的预测表现更好（在县和州，MAPE分别降低了0.9％和1.43％），与内布拉斯加州相似，在其他州则更差。总体而言，最佳的州级预报始终优于同期的国家农业统计局（NASS）预报。