| 引用本文: | 李宗鹏,陈震,程千,等.基于Stacking法的无人机光谱遥测冬小麦产量[J].灌溉排水学报,0,():-. |
| LI Zongpeng,chenzhen,chengqian,et al.基于Stacking法的无人机光谱遥测冬小麦产量[J].灌溉排水学报,0,():-. |
|
| |
|
|
| 本文已被：浏览 2141次 下载 0次 |
|
|
|
| 基于Stacking法的无人机光谱遥测冬小麦产量 |
|
李宗鹏1, 陈震1, 程千2, 徐洪刚2, 庞超凡3, 李连豪3
|
|
1.河南农业大学机电工程学院;2.中国农业科学院农田灌溉研究所;3.河南农业大学
|
|
| 摘要: |
| 为了精确、高效的预测作物产量，本文以冬小麦为研究对象，利用无人机搭载多光谱相机，获取抽穗期、开花期和灌浆期的多光谱图像数据。根据多光谱波段选取对产量敏感的14种植被指数，并优选出与产量极显著相关的13种指数；利用十折交叉验证的方法分别建立各生长阶段的MLR（多元线性回归）、PLSR（偏最小二乘回归）、SVM（支持向量机）和Cubist产量估算初级模型进行对比分析，并利用Stacking方法集成初级模型分别建立各个时期MLR和Cubist次级产量估测模型。结果表明初级学习器中Cubist模型在抽穗期（R2 = 0.41，RMSE = 1.21t ha-1，NRMSE = 18.59%）、开花期（R2 = 0.45，RMSE = 1.19t ha-1，NRMSE = 18.23%）和灌浆期（R2 = 0.57，RMSE = 1.07t ha-1，NRMSE = 16.36%）的估产精度均为最高；MLR作为次级学习器时R2比初级学习器各生育期的最高值分别提升了0.11、0.10和0.04，RMSE比最低值分别降低了0.10t ha-1、0.12t ha-1和0.02t ha-1，Cubist作为次级学习器时R2比初级学习器各生育期最大值分别提升了0.12、0.13和0.04，RMSE比最低值分别降低了0.11t ha-1、0.14t ha-1和0.02t ha-1，并且Cubist作为次级学习器时在抽穗期（R2 = 0.54，RMSE = 1.12t ha-1）、开花期（R2 = 0.58，RMSE = 1.06t ha-1）和灌浆期（R2 = 0.61，RMSE = 1.05t ha-1）的产量估算精度仍是最高。使用Stacking方法对不同初级学习器模型集成能够有效提高产量估算精度，可为今后的冬小麦产量估测研究提供参考。 |
| 关键词: 多光谱；植被指数；十折交叉验证；Stacking；学习器 |
| DOI： |
| 分类号:S275 |
| 基金项目:中国农业科学院科技创新工程重大产出培育项目“天空地农田精准灌溉信息智能感知技术与装备研发”，河南省科技攻关计划“无人机光谱感知喷灌机变量喷洒水氮空间变异性研究”作物抗逆育种与减灾国家地方联合工程实验室开放基金资助项目（NELCOF20190104），河南省科技开放合作项目（172106000015），河南农业大学“百名教授、千名学生、服务万村”基金项目（3080163） |
|
| UAV spectrum remote measurement of winter wheat yield based on Stacking method |
|
LI Zongpeng,chenzhen,chengqian,et al
|
|
1.College of mechanical and electrical engineering, Henan Agricultural University;2.Institute of irrigation, Chinese Academy of Agricultural Sciences
|
| Abstract: |
| In order to accurately and efficiently predict crop yields, this paper takes winter wheat as the research object and uses drones equipped with multispectral cameras to obtain multispectral image data of the heading, flowering, and filling periods.Select 14 vegetation indices that are sensitive to yield according to multispectral bands, and select 13 indices that are extremely significantly related to yield;The 10-fold cross-validation method was used to establish MLR (Multiple Linear Regression), PLSR (Partial Least Squares Regression), SVM (Support Vector Machine) and Cubist yield estimation primary models for each growth stage for comparative analysis, and stacking method integration The primary model establishes the MLR and Cubist secondary production estimation models for each period.The results show that the Cubist model in the primary learner is in the heading stage (R2 = 0.41, RMSE = 1.21t ha-1, NRMSE = 18.59%), flowering stage (R2 = 0.45, RMSE = 1.19t ha-1, NRMSE = 18.23%) and During the grouting period (R2 = 0.57, RMSE = 1.07t ha-1, NRMSE = 16.36%), the accuracy of production estimation is the highest;When MLR is used as a secondary learner, R2 is increased by 0.11, 0.10, and 0.04, respectively, than the highest value of each growth period of the primary learner, and the RMSE is reduced by 0.10t ha-1, 0.12t ha-1, and 0.02t ha-1 respectively from the lowest value. When Cubist is used as a secondary learner, R2 is increased by 0.12, 0.13, and 0.04, respectively, than the maximum value of each growth period of the primary learner, and the RMSE is reduced by 0.11 t ha-1, 0.14 t ha-1, and 0.02 t ha-1, respectively, from the lowest value. And Cubist is used as a secondary learner in the heading stage (R2 = 0.54, RMSE = 1.12 t ha-1), flowering stage (R2 = 0.58, RMSE = 1.06 t ha-1) and filling stage (R2 = 0.61, RMSE = 1.05 t ha-1) The yield estimation accuracy is still the highest.Using the Stacking method to integrate different primary learner models can effectively improve the accuracy of yield estimation, which can provide a reference for future winter wheat yield estimation research. |
| Key words: Multispectral; vegetation index; ten-fold cross-validation; Stacking; learner |
|
|
|
|
