中国农业科学 ›› 2019, Vol. 52 ›› Issue (17): 2939-2950.doi: 10.3864/j.issn.0578-1752.2019.17.003

• 耕作栽培·生理生化·农业信息技术 • 上一篇    下一篇

基于机器学习的滴灌玉米光合响应特征

刘慧芳,贺正,贾彪(),刘志,李振洲,付江鹏,慕瑞瑞,康建宏   

  1. 宁夏大学农学院,银川 750021
  • 收稿日期:2019-04-02 接受日期:2019-07-03 出版日期:2019-09-01 发布日期:2019-09-10
  • 通讯作者: 贾彪
  • 作者简介:刘慧芳,E-mail:18747555998@163.com。
  • 基金资助:
    国家自然科学基金(31560339);宁夏自然基金项目(2019AAC03068);宁夏高等学校科研项目(NGY2017025);宁夏回族自治区科技重大专项(2018BBF0200404);宁夏区重点研发计划项目(2018BBF02018);国家科技支撑计划项目(2015BAD22B01)

Photosynthetic Response Characteristics of Maize Under Drip Irrigation Based on Machine Learning

LIU HuiFang,HE Zheng,JIA Biao(),LIU Zhi,LI ZhenZhou,FU JiangPeng,MU RuiRui,KANG JianHong   

  1. School of Agriculture, Ningxia University, Yinchuan 750021
  • Received:2019-04-02 Accepted:2019-07-03 Online:2019-09-01 Published:2019-09-10
  • Contact: Biao JIA

摘要:

【目的】提出一种优化模型精度的机器学习网格搜索方法,解决滴灌玉米光合响应曲线模型参数确定难、精度低等问题,为滴灌玉米光合生理机制及光合响应特征提供新思路。【方法】2017年和2018年以宁夏玉米主栽品种(TC19)为试验材料,设置6个施钾水平(0(K0)、90 kg·hm -2(K1)、180 kg·hm -2(K2)、270 kg·hm -2(K3)、360 kg·hm -2(K4)、450 kg·hm -2(K5)),使用Li-6400XT光合仪测定不同钾肥水平下玉米吐丝期光响应曲线。运用机器学习网格搜索法和非线性回归分析法对基于直角双曲线修正模型的光响应曲线进行拟合。选取决定系数(R 2)、均方根误差(RMSE)及平均绝对误差(MAE)对模型精度进行评价。【结果】在玉米吐丝期,叶片光合参数Pn、Tr和Gs随施钾量的增加呈先增大后减小的趋势。拟合评价结果表明,在K0和K1处理下机器学习方法计算效果优于传统方法,R 2均大于0.991,RMSE均小于1.487,MAE均小于1.350。在K2—K5处理下,2种方法拟合效果相当,R 2均大于0.993,RMSE均小于0.952、MAE均小于0.860。最优拟合方法(网格搜索法)对光响应特征参数计算结果表明,α、Pnmax、Rd、LSP和LCP的变化趋势与其光合参数相似。在施钾量为360 kg·hm -2(K4)时,各光响应特征参数均达到最大,在450 kg·hm -2(K5)时出现光抑制现象。【结论】基于机器学习的网格搜索法可准确地拟合宁夏滴灌玉米光响应特征,且施钾量为360 kg·hm -2时玉米光合性能达到最佳。

关键词: 玉米, 光合参数, 光响应曲线, 机器学习, 模型优化

Abstract:

【Objective】 This study proposed an optimized grid search method based on machine learning to solve the problem of model parameters of the photosynthetic light-response curve for drip-irrigated maize, which was often hard to determine and possesses low precision, so as to provide new ideas for photosynthetic characteristics and mechanisms of drip-irrigated maize in Ningxia. 【Method】 The experiment was conducted in 2017 and 2018 with the maize cultivar TC19, which was widely cultivated in Ningxia. Six levels of potassium application (0 (K0), 90 kg·hm -2 (K1), 180 kg·hm -2 (K2), 270 kg·hm -2 (K3), 360 kg·hm -2 (K4), 450 kg·hm -2 (K5)) were set, and the portable gas exchange system (Li-6400XT) was used to measure the light-response curves of maize under different potassium levels at silking stage. The grid search method based on machine learning and nonlinear regression analysis was used to revise the light response curve based on the right angle and hyperbolic correction model. The correlation coefficient (R 2), root-mean-square error (RMSE) and mean absolute error (MAE) were used to evaluate the accuracy of the model. 【Result】 The results showed that the photosynthetic parameters (Pn), transpiration rate (Tr) and stomatal conductance (Gs) of maize leaves increased first and then decreased with the increase of potassium application rate. The results of fitting evaluation indicated that the calculation results of machine learning method under K0 and K1 were better than the traditional method, in which R 2 was greater than 0.991, RMSE and MAE were less than 1.487 and 1.350, respectively. The two methods have similar fitting effect under K2-K5, while R 2was greater than 0.993, RMSE and MAE was less than 0.952 and 0.860, respectively. The result of optical response characteristic parameter calculation by using the optimum fitting method (grid search method) showed that the trends of α, Pnmax, Rd, LSP and LCP were similar to their photosynthetic parameters. When the potassium application rate was 360 kg·hm -2 (K4), the light response characteristic parameters reached the maximum value, however, the light suppression phenomenon occurred at 450 kg·hm -2 (K5). 【Conclusion】 The grid search method based on machine learning could accurately fit the photo-response characteristics of drip-irrigated maize in Ningxia, and the photosynthetic performance of maize was the best when the potassium application rate was 360 kg·hm -2.

Key words: maize, photosynthetic parameters, light response curve, machine learning, model optimization