基于多源遥感数据融合与机器学习的多尺度冬油菜生物量监测和产量预测研究

Multi-source Remote Sensing Data Fusion and Machine Learning in Multi-scale Winter Rapeseed Applications in Biomass Monitoring and Yield Forecasting

摘要:

油菜是我国重要油料作物之一,准确的生物量估算可以提高油菜产量的预测精度,对制定合理的种植策略、 提升产量有重要作用。 针对传统人工监测油菜地上生物量(Aboveground biomass, AGB)与产量方法成本高、时效差、空间尺度受限等问题,本文提出遥感数据、气象数据和机器学习的多模态特征融合框架以实现低成本、大范围、 高精度监测生物量与产量。 通过融合无人机高分辨率影像、599 组田间生物量实测数据、327 组产量数据和逐日气象数据,构建包含植被指数与纹理特征的多维特征空间用于构建田间尺度生物量预测模型。 随后以生物量模型的输出和田间尺度的特征数据为输入构建产量预测模型,并将优化后的机器学习模型用于卫星遥感数据与县级气象数据融合生成的多维特征空间,实现县级尺度范围的产量预测。 结果表明,经过超参数寻优后,基于 XGB 的生物量预测模型 R2 (0. 788 8)最佳,在将生物量预测值作为关键特征数据之一的基于随机森林的产量预测模型均方误差最佳为 0. 135 3 t2 / hm4 。 最后将模型外推至湖北省油菜种植第一大市———荆州市的 8 个区县,在不依赖生物量真值并且县域异质性较高的复杂条件下实现 72% (最高可达 91. 97% )的平均精度。 最终在多模态数据融合与尺度迁移的双重挑战下实现了高精高效低成本的田间尺度和县级尺度的油菜生物量监测和产量预测,为区域作物监测提供了有效参考。

Abstract:

Rapeseed is one of the important oil crops in China, and accurate biomass estimation can improve the prediction accuracy of rapeseed yield, which plays an important role in formulating reasonable planting strategies and increasing yield. In view of the high cost, poor timeliness and limited spatial scale of traditional manual monitoring of rapeseed biomass and yield, a multimodal feature fusion framework of remote sensing data, meteorological data and machine learning was proposed to achieve low- cost, large-scale and high-precision monitoring of biomass and yield. By integrating UAV high-resolution images, totally 599 sets of field biomass measured data, 327 sets of yield data and daily meteorological data, a multi-dimensional feature space containing vegetation index and texture features was constructed to construct a field-scale biomass prediction model. Then, the yield prediction model was constructed by using the output of the biomass model and field-scale feature data as inputs, and the optimized machine learning model was used in the multi-dimensional feature space generated by the fusion of satellite remote sensing data and county-level meteorological data to achieve yield prediction at the county level. The results showed that the biomass prediction model based on XGB got the best R2 (0. 788 8), and the mean square error of the yield prediction model based on random forest, which took the biomass prediction value as one of the key feature data, was 0. 135 3 t2 / hm4 . Finally, the model was extrapolated to eight districts and counties in Jingzhou City, the largest rapeseed growing city in Hubei Province, and achieved an average accuracy of 72% (up to 91. 97% )under complex conditions that did not rely on biomass truth value and high county heterogeneity. Finally, under the dual challenges of multimodal data fusion and scale migration, high-precision, high efficiency and low cost field-scale and county-level rapeseed biomass monitoring and yield prediction were realized, which provided an effective reference for regional crop monitoring.