The model for grain wheat yield prediction at high spatial resolution based on physical-geographical properties and satellite vegetation indices

Precision agriculture is promising approach for improving agricultural production, especially nowadays when the population is rapidly increasing. For that, crop yield estimation provides valuable information. The main research focus was to predict within-field grain yield and detect its drivers. The Random Forest regression model on data from diverse sources at the 10-meter spatial resolution was developed. The study was conducted in the Vojvodina region (Serbia) for eight wheat-planted fields, having precise grain yield data. Open-source data including 15 vegetation indices (VIs) was calculated from Sentinel-2 satellite bands, physical-geographical features obtained from the digital elevation model and soil properties. The model succeeded in predicting the wheat grain yield with the RMSE of 0.66 t/ha (average yield of 0.09 t/ha) and the best predictors were VIs considering chlorophyll and moisture content in plants, while physical-geographical properties managed to explain within-field variability. This methodology can be applied to other crops (maize, soybean).

精准农业（Precision Agriculture）是提升农业生产水平的极具潜力的途径，尤其在当前全球人口快速增长的背景下。为此，作物产量估算可提供极具价值的决策参考信息。本研究的核心研究方向为预测田块内谷物产量并识别其影响驱动因子。

本研究构建了基于10米空间分辨率多源数据的随机森林回归模型。实验在塞尔维亚伏伊伏丁那地区的8个小麦种植田块开展，所有田块均配有精准的谷物产量实测数据。开源数据集包含15种植被指数（Vegetation Indices，VIs），该类指数由哨兵-2号（Sentinel-2）卫星波段计算得到；数据集同时涵盖了从数字高程模型（Digital Elevation Model，DEM）提取的自然地理特征以及土壤属性数据。

实验结果显示，该模型对小麦谷物产量的预测均方根误差（Root Mean Square Error，RMSE）为0.66 t/ha（对应平均产量基准为0.09 t/ha）；其中表现最优的预测因子为考量植物叶绿素与含水量的植被指数，而自然地理特征则能够有效解释田块内部的产量变异。本研究提出的方法可推广应用至其他作物（玉米、大豆）。