欢迎访问《麦类作物学报》编辑部网站

作者	单位
马宇欣，胡笑涛，王亚昆，范晓懂，彭雪莲，孙骏，陈洪	（西北农林科技大学旱区农业水土工程教育部重点实验室，陕西杨陵 712100）

摘要点击次数: 475

全文下载次数: 213

中文摘要:

为进一步提升利用高光谱数据快速监测叶片含水量的能力，以关中地区冬小麦为研究对象，分别获取2022和2023年孕穗期、抽穗期及灌浆期冬小麦叶片含水量，并同步监测叶片高光谱信息。通过波段组合的形式构建植被指数，并利用相关性分析进行初步筛选，再以ReliefF算法优选得到输入特征变量，然后分别利用随机森林（random forest, RF）、长短期记忆（long short-term memory, LSTM）网络和基于粒子群（particle swarm optimization，PSO）优化的反向传播神经网络（back propagation neural network, BPNN）构建冬小麦叶片含水量估算模型，并进行精度评价。结果表明，通过相关性分析与ReliefF算法结合优选变量，能够较单独通过相关分析明显提升LSTM和PSO-BPNN模型的建模精度，但对RF模型则无法优化变量。相关性分析与ReliefF结合后PSO-BPNN模型在各生育时期均取得最佳预测结果，其中孕穗期、抽穗期和灌浆期验证集r²分别为0.816、0.736和0.806，RMSE分别为0.546%、0.899%和1.531%，NRMSE分别为0.681%、1.195%和2.185%。由此可见，在相关分析的基础上，通过ReliefF算法优选特征变量能够提升特定模型对冬小麦叶片含水量的估算精度，其中对PSO-BPNN模型的效果最好。

英文摘要:

Estimation of the leaf water content (LWC) plays an important role in field irrigation management. This study aimed to estimate the LWC of winter wheat based on hyperspectral data of leaf blades, especially focusing on the effect of different variable screening methods and growth stages on the estimation model. Research data were obtained from field trials in 2022 and 2023 at the booting, heading, and grain-filling stages. Vegetation indices were constructed for each growth stage by combining the two bands. Tthe input characteristic variables based on vegetation indices were screened by two methods:(I) the input characteristic variables were directly obtained by ranking the correlation coefficients; (II) based on the method I, the vegetation index was further screened by the ReliefF algorithm to obtain a second set of input characteristic variables. The LWC estimation models were constructed using random forest (RF), long short-term memory (LSTM) network and back propagation neural network (BPNN) based on particle swarm optimization (PSO). The best method for estimating LWC was derived by comparing the accuracy of the models. The results showed that comparing the two variable screening methods, the characteristic variables further screened by ReliefF could effectively improve the accuracy of the LSTM and PSO-BPNN models, while the effect of improving the RF model is not obvious. The best model for each growth stage was established by the ReliefF screening method combined with the PSO-BPNN, at the booting stage, heading stage and grain-filling stage. The r² of the validation set was 0.816, 0.736, and 0.806, respectively, and the RMSE was 0.546%, 0.899%, and 1.531%, respectively, and the NRMSE was 0.681%, 1.195%, and 2.185%, respectively. It was suggested that the screening method of feature variables through the ReliefF algorithm could improve its estimation accuracy in the particular model. Its combination with the PSO-BPNN model had the best application effect in the estimation of LWC in winter wheat at the growth stages.

查看全文查看/发表评论下载PDF阅读器

关闭