Abstract An accurate model for predicting sugarcane yield will benefit many aspects of managing growth and harvest of sugarcane crops. In this study, Sentinel-1 and Sentinel-2 satellite imagery were used in combination with climate, soil and elevation data to predict field-level sugarcane yield across the multiple sugar mill areas in the Wet Tropics of Australia at different time steps over four consecutive growing seasons (2016–2019). A total of ≈1400 field-level measurements were used to train predictive machine learning models of cane yield (t/ha), commercial cane sugar (CCS, %), sugar yield (t/ha), crop varieties and ratoon numbers. We compared the predictive performance of models based on both satellite imagery only and a fusion of satellite imagery with climate, soil and topographical information. Randomized search on hyperparameters was the method used to optimize and identify the most accurate decision tree-based machine model. Overall, gradient boosting was the most accurate method for predicting sugarcane attributes. The analysis resulted in cane yield, CCS and sugar yield predicted at the field level with R2 of up to 0.51 (RMSE = 16 t/ha), 0.63 (RMSE = 1 %) and 0.62 (RMSE = 2 t/ha) as soon as four months before the harvest season. It was also found that sugarcane varieties could be mapped with an accuracy of up to 73.4 %, while the differentiation of planted and ratoon crops exhibited the lowest accuracy of 45.4 %. Using a novel SHapley Additive exPlanations (SHAP) approach to explain the output of our machine learning models we found that Sentinel-2 derived spectral indices were the most important in predicting cane yield as well as differentiating sugarcane varieties and ratoon numbers. In contrast, climate and elevation derived predictors were the most important in predicting CCS and sugar yield. At the whole sugar mill area level, spatially averaged field-level results predicted mill area cane yield, CCS and sugar yield with R2 of 0.75 (RMSE = 4.6 t/ha), 0.80 (RMSE = 0.6 %) and 0.77 (RMSE = 1 t/ha). Early season prediction of sugarcane yields at both field- and mill-area level could be valuable for informing fertilizer application, harvest scheduling and marketing decisions.

中文翻译：

---

整合卫星图像和环境数据，使用机器学习预测澳大利亚的田间甘蔗和糖产量

摘要 预测甘蔗产量的准确模型将有益于管理甘蔗作物的生长和收获的许多方面。在这项研究中，Sentinel-1 和 Sentinel-2 卫星图像与气候、土壤和海拔数据结合使用，以预测澳大利亚湿热带多个糖厂区域在连续四个连续时间步长的田间甘蔗产量生长季节（2016-2019）。总共 ≈1400 个田间测量用于训练甘蔗产量 (t/ha)、商业蔗糖 (CCS, %)、糖产量 (t/ha)、作物品种和宿根数量的预测机器学习模型。我们比较了仅基于卫星图像和卫星图像与气候、土壤和地形信息融合的模型的预测性能。超参数的随机搜索是用于优化和识别最准确的基于决策树的机器模型的方法。总的来说，梯度提升是预测甘蔗属性最准确的方法。分析结果在田间水平预测甘蔗产量、CCS 和糖产量，R2 最高可达 0.51 (RMSE = 16 t/ha)、0.63 (RMSE = 1 %) 和 0.62 (RMSE = 2 t/ha)就像收获季节前的四个月一样。还发现甘蔗品种的定位精度高达 73.4%，而种植和宿根作物的分化精度最低，为 45.4%。使用一种新颖的 SHapley Additive exPlanations (SHAP) 方法来解释我们机器学习模型的输出，我们发现 Sentinel-2 派生的光​​谱指数在预测甘蔗产量以及区分甘蔗品种和宿根数量方面最重要。相比之下，气候和海拔衍生的预测因子在预测 CCS 和糖产量方面最为重要。在整个糖厂面积水平，空间平均的田间水平结果预测糖厂面积甘蔗产量、CCS 和糖产量，R2 为 0.75 (RMSE = 4.6 t/ha)、0.80 (RMSE = 0.6 %) 和 0.77 (RMSE = 1)吨/公顷）。在田间和碾磨区层面对甘蔗产量的早期预测对于为施肥、收获计划和营销决策提供信息可能很有价值。气候和海拔衍生的预测因子在预测 CCS 和糖产量方面是最重要的。在整个糖厂区域水平，空间平均田间水平结果预测糖厂区域甘蔗产量、CCS 和糖产量，R2 为 0.75 (RMSE = 4.6 t/ha)、0.80 (RMSE = 0.6 %) 和 0.77 (RMSE = 1)吨/公顷）。在田间和碾磨区层面对甘蔗产量的早期预测对于为施肥、收获计划和营销决策提供信息可能很有价值。气候和海拔衍生的预测因子在预测 CCS 和糖产量方面是最重要的。在整个糖厂区域水平，空间平均田间水平结果预测糖厂区域甘蔗产量、CCS 和糖产量，R2 为 0.75 (RMSE = 4.6 t/ha)、0.80 (RMSE = 0.6 %) 和 0.77 (RMSE = 1)吨/公顷）。在田间和碾磨区层面对甘蔗产量的早期预测对于为施肥、收获计划和营销决策提供信息可能很有价值。