Abstract Forecasts of crop yield are an important tool for a variety of stakeholders, but most studies produce large-scale, late-season yield forecasts that are not appropriate for farmers. Farmers require forecasts of crop yield mid-season and at fine spatial resolutions to guide site-specific adaptive crop management practices. This study created empirical random forest models to forecast cotton yield mid-season at three different spatial resolutions: 30 m, field, and farm aggregation. Large yield mapping datasets from 14 different seasons were utilised to build the model across 68 different fields from six large cotton farms in eastern Australia. A generic conceptual framework is proposed to empirically model crop yield. This topsaw framework represents the spatial and temporal factors that drive yield, including topography (t), organisms (o), plant measurements (p), soil (s), agronomy/management (a), and weather (w). In this study, the specific predictor variables to represent these driving factors of yield included elevation (t), Normalised Difference Vegetation Index, Enhanced Vegetation Index, MODIS Evapotranspiration (p), gamma radiometrics grid, clay content digital soil map (s), rainfall data, and growing day degrees (w). Forecasts of cotton yield increased in accuracy as the spatial resolution of predictions became coarser. Using a leave-one-year-out cross-validation, cotton yield could be forecasted to an accuracy of 0.42 Lin's Concordance Correlation Coefficient (LCCC) and Root Mean Square Error (RMSE) of 2.11 bales ha−1 at 30 m resolution. This improved to a 0.63 LCCC and 1.72 bales ha−1 RMSE at the field resolution, and a 0.65 LCCC and 0.77 bales ha−1 at the aggregation-scale. As the developed approach solely relied on publicly available predictor variables, there is an opportunity to apply this to any cotton field in Australia, and over a much larger area. Overall, this study is an important step in building an operational approach to forecast mid-season cotton yield at fine spatial resolutions, and has the potential to improve production, input use efficiency, and profitability for individual growers and the Australian Cotton Industry as a whole.

中文翻译：

---

使用大型产量绘图数据集和不同的空间协变量以精细分辨率进行中期经验棉花产量预测

摘要 作物产量预测是各种利益相关者的重要工具，但大多数研究产生不适合农民的大规模、晚季产量预测。农民需要在季节中期以精细的空间分辨率预测作物产量，以指导特定地点的适应性作物管理实践。本研究创建了经验随机森林模型，以在三种不同的空间分辨率下预测季中棉花产量：30 m、田地和农场聚集。来自 14 个不同季节的大产量绘图数据集被用来在澳大利亚东部六个大型棉花农场的 68 个不同领域建立模型。提出了一个通用的概念框架来对作物产量进行经验建模。这个 topsaw 框架代表了驱动产量的空间和时间因素，包括地形 (t)、生物 (o)、植物测量 (p)、土壤 (s)、农学/管理 (a) 和天气 (w)。在本研究中，代表这些产量驱动因素的特定预测变量包括海拔 (t)、归一化差异植被指数、增强型植被指数、MODIS 蒸散量 (p)、伽马辐射测量网格、粘土含量数字土壤图 (s)、降雨量数据和生长日度 (w)。随着预测的空间分辨率变得更粗糙，棉花产量的预测准确性增加。使用留出一年的交叉验证，在 30 m 分辨率下，棉花产量可以预测为 0.42 Lin 相关系数 (LCCC) 和 2.11 包 ha−1 的均方根误差 (RMSE)。在现场分辨率下，这改进为 0.63 LCCC 和 1.72 包 ha-1 RMSE，以及 0.65 LCCC 和 0。77 包 ha−1 在聚集规模。由于开发的方法完全依赖于公开可用的预测变量，因此有机会将其应用于澳大利亚的任何棉田，以及更大的区域。总体而言，这项研究是构建以精细空间分辨率预测中季棉花产量的操作方法的重要一步，并且有可能提高个体种植者和整个澳大利亚棉花行业的产量、投入使用效率和盈利能力.