多源遥感与气象数据融合的油菜单产和总产预测

Prediction of Canola Yield and Production by Fusing Multisource Remote Sensing and Meteorological Data

摘要:

油菜是全球重要的油料作物之一,其单产与总产的准确预测对粮油安全评估、农业生产管理及生物能源潜力分析具有重要意义。针对油菜单产与总产预测中样本规模受限、总产建模和多源信息协同效应缺乏系统评估的问题,本文收集县域油菜单产和总产数据,综合利用多源遥感与气象大数据,构建油菜单产与总产的协同预测框架。 在此基础上,引入6种树结构机器学习模型,系统评估不同模型及不同数据源组合在油菜单产与总产预测中的性能差异。结果表明,遥感与气象数据联合输入在单产和总产预测中均优于单一遥感或气象数据源,且这一优势在不同树模型结构下表现出良好的一致性与稳定性。不同数据源组合在所有模型中的预测性能遵循一致排序:即联合输入表现最佳,其次为仅遥感输入,最低为仅气象输入;说明遥感与气象数据在生长状态与环境驱动信息上的互补性,是提高油菜单产和总产预测精度的重要基础。单产预测中,CatBoost 表现最佳( R2 为0. 679、RMSE 为 0. 269 t/ hm2、MAE 为0. 194 t/ hm2);总产预测中,Cubist 表现最佳(R2为0. 952、RMSE为9. 42 kt、MAE为5. 00 kt)。 油菜单产与总产预测结果在空间分布上与观测值总体一致。本研究提出的多源遥感与气象信息融合结合树结构机器学习方法,能够有效提升区域尺度油菜单产与总产预测精度,为油菜产量监测、精准农业管理及农机收获调度决策提供了可靠的技术支撑。

Abstract:

Canola is one of the most important oilseed crops worldwide, and accurate prediction of its yield and total production is of great significance for food and edible oil security assessment, agricultural production management, and bioenergy potential analysis. To address the limited sample size in canola yield and production prediction, as well as the lack of systematic evaluation of production modeling and multi-source information synergy, county-level canola yield and total production data were collected, and multi-source remote sensing and meteorological big data were integrated to develop a collaborative prediction framework for canola yield and total production. On this basis, six tree-based machine learning models were employed to systematically evaluate the performance differences among models and data- source combinations in predicting canola yield and production. The results demonstrated that the joint use of remote sensing and meteorological data consistently outperformed single-source inputs for both yield and production prediction, with this advantage showing strong consistency and robustness across different tree- based model structures. For yield prediction, CatBoost achieved the best performance ( R2 = 0. 679, RMSE = 0. 269 t/ hm2, MAE = 0. 194 t/ hm2), whereas for production prediction, Cubist performed best (R2 = 0. 952, RMSE = 9. 42 × 103 t, MAE = 5. 00 × 103 t). The spatial patterns of the predicted yield and production were generally consistent with the observations. Overall, the results indicated that the integration of multi-source remote sensing and meteorological information with tree-based machine learning methods can effectively improve the accuracy of regional-scale canola yield and production prediction. The research result can provide reliable technical support for canola yield monitoring, precision agricultural management, and harvest scheduling decisions for agricultural machinery.