Abstract Soybean yield predictions in Brazil are of great interest for market behavior, to drive governmental policies and to increase global food security. In Brazil soybean yield data generally demand various revisions through the following months after harvest suggesting that there is space for improving the accuracy and the time of yield predictions. This study presents a novel model to perform in-season (“near real-time”) soybean yield forecasts in southern Brazil using Long-Short Term Memory (LSTM), Neural Networks, satellite imagery and weather data. The objectives of this study were to: (i) compare the performance of three different algorithms (multivariate OLS linear regression, random forest and LSTM neural networks) for forecasting soybean yield using NDVI, EVI, land surface temperature and precipitation as independent variables, and (ii) evaluate how early (during the soybean growing season) this method is able to forecast yield with reasonable accuracy. Satellite and weather data were masked using a non-crop-specific layer with field boundaries obtained from the Rural Environment Registry that is mandatory for all farmers in Brazil. Main outcomes from this study were: (i) soybean yield forecasts at municipality-scale with a mean absolute error (MAE) of 0.24 Mg ha−1 at DOY 64 (march 5) (ii) a superior performance of the LSTM neural networks relative to the other algorithms for all the forecast dates except DOY 16 where multivariate OLS linear regression provided the best performance, and (iii) model performance (e.g., MAE) for yield forecast decreased when predictions were performed earlier in the season, with MAE increasing from 0.24 Mg ha−1 to 0.42 Mg ha−1 (last values from OLS regression) when forecast timing changed from DOY 64 (March 5) to DOY 16 (January 6). This research portrays the benefits of integrating statistical techniques, remote sensing, weather to field survey data in order to perform more reliable in-season soybean yield forecasts.

中文翻译：

---

基于卫星的大豆产量预测：整合机器学习和天气数据以改进巴西南部的作物产量预测

摘要 巴西大豆产量预测对市场行为、推动政府政策和提高全球粮食安全具有重要意义。在巴西，大豆产量数据通常需要在收获后的接下来几个月进行各种修正，这表明产量预测的准确性和时间还有提高的空间。本研究提出了一种使用长短期记忆 (LSTM)、神经网络、卫星图像和天气数据进行巴西南部季节性（“近实时”）大豆产量预测的新模型。本研究的目的是：(i) 比较三种不同算法（多元 OLS 线性回归、随机森林和 LSTM 神经网络）使用 NDVI、EVI、地表温度和降水作为自变量预测大豆产量的性能，(ii) 评估该方法在多早（在大豆生长季节）能够以合理的准确度预测产量。卫星和天气数据使用非作物特定层和从农村环境登记处获得的田间边界进行掩蔽，这对巴西所有农民都是强制性的。这项研究的主要结果是：（i）在 DOY 64（3 月 5 日）时以 0.24 Mg ha−1 的平均绝对误差 (MAE) 进行市政规模的大豆产量预测 (ii) LSTM 神经网络相对于到除 DOY 16 以外的所有预测日期的其他算法，其中多元 OLS 线性回归提供了最佳性能，并且 (iii) 当预测在本季节早些时候进行时，用于产量预测的模型性能（例如，MAE）下降，MAE 从0.24 Mg ha-1 至 0。当预测时间从 DOY 64（3 月 5 日）更改为 DOY 16（1 月 6 日）时，为 42 Mg ha−1（来自 OLS 回归的最后值）。这项研究描绘了将统计技术、遥感、天气与实地调查数据相结合的好处，以便进行更可靠的当季大豆产量预测。