基于深度学习的梨树养分含量高光谱监测

doi:10.3964/j.issn.1000-0593(2024)12-3543-10

摘要
参考文献
相关文章 (15)

全文: PDF (6657 KB)
输出: BibTeX | EndNote (RIS)

摘要：为了实现梨树生长和结果时期对于梨树养分的精准和及时地监测，进一步为梨树施肥管理及梨果品质提升提供保障和调整策略，养分监测十分重要。使用ASD FieldSpec3高光谱仪获取梨树果实膨大期和成熟期的叶片高光谱数据，采集叶片氮磷钾含量信息。对比分析原始反射率、一阶导数变换（FD）、卷积平滑算法（SG）、标准正态变量变换法（SNV）等光谱预处理方法对于高光谱反射率监测模型拟合效果的影响。利用主成分分析（PCA）、竞争自适应重加权抽样（CARS）和连续投影算法（SPA）选择高光谱数据的特征波段。之后，使用偏最小二乘法回归（PLSR）、支持向量回归（SVR）、随机森林（RF）、梯度提升树（GBDT）、卷积神经网络（CNN）、深度森林（DF）算法建立基于特征波段的监测模型，筛选梨树氮磷钾3种营养元素的最优高光谱监测模型。结果表明，深度学习算法DF具备较好的处理高光谱数据的能力。氮元素的最佳建模组合为SG-SNV预处理的PCA特征波段的DF回归模型（R²=0.928 3，RMSE=0.238 1 g·kg^-1）；磷元素的最佳建模组合为SG-SNV预处理的SPA特征波段的GBDT回归模型（R²=0.936 7，RMSE=0.043 1 g·kg^-1）；钾元素的最佳建模组合为SG-SNV预处理的PCA特征波段的DF回归模型（R²=0.954 4，RMSE=0.276 7 g·kg^-1）。基于高光谱特征波段的监测模型拟合效果良好（R²>0.9），可以实现梨树果实膨大期和成熟期氮磷钾含量的准确监测。

关键词：梨树；高光谱；机器学习；深度学习；养分含量

Abstract：To achieve wide-scale, accurate, and timely monitoring of pear nutrients during the growth and fruiting periods of pear trees and to further provide protection and adjustment strategies for pear fertilization management and pear fruit quality improvement, we used the ASD FieldSpec3 hyperspectral spectroscopy system to monitor the nutrient content of pear trees. The ASD FieldSpec3 hyperspectrometer was used to obtain leaf hyperspectral data during the fruit expansion and ripening periods of pear trees and to collect information on the leaf nitrogen, phosphorus, and potassium content. The effects of spectral preprocessing methods such as raw reflectance, first-order derivative transformation (FD), convolutional smoothing algorithm (SG), and standard normal variable transformation (SNV) on the fitting effect of the hyperspectral reflectance monitoring model were comparatively analyzed through the raw spectral curves. Principal component analysis (PCA), competitive adaptive reweighted sampling (CARS), and successive projection algorithm (SPA) are then used to select the characteristic bands of hyperspectral data. After that, partial least squares regression (PLSR), support vector regression (SVR), random forest (RF), gradient boosted tree (GBDT), convolutional neural network (CNN), and deep forest (DF) algorithms were used to establish monitoring models based on the edge bands to screen the optimal hyperspectral monitoring models for the three nutrient elements, nitrogen, phosphorus, and potassium, in pear trees. The optimal modeling combination of nitrogen was the DF regression model of PCA eigenbands pre-processed by SG-SNV (R²=0.928 3, RMSE=0.238 1 g·kg^-1). The optimal modeling combination of phosphorus was the GBDT regression model of SPA eigenbands pre-processed by SG-SNV (R²=0.936 7, RMSE=0.043 1 g·kg^-1). The best modeling combination for potassium was the DF regression model (R²=0.954 4, RMSE=0.276 7 g·kg^-1) for the SG-SNV preprocessed PCA feature band. The monitoring model based on hyperspectral edge bands was well fitted (R²>0.9), which can realize the accurate monitoring of nitrogen, phosphorus, and potassium content in pear fruit during expansion and ripening.

Key words：Pear tree; Hyperspectral; Machine learning; Deep learning; Nutrient content

收稿日期: 2024-03-05 修订日期: 2024-06-18

中图分类号:

TP79

基金资助: 江苏省现代农机装备与技术示范推广项目（NJ2022-12），江苏省科技计划专项资金（重点研发计划现代农业）项目（BE2022374），国家自然科学基金项目（62305166）资助

通讯作者: 蒋雪松 E-mail: xsjiang@njfu.edu.cn

作者简介: 黄林峰，1999年生，南京林业大学机械电子工程学院研究生 e-mail: huanglf@njfu.edu.cn

引用本文:

黄林峰，蒋雪松，贾志成，周宏平，周磊，戎子凡. 基于深度学习的梨树养分含量高光谱监测[J]. 光谱学与光谱分析, 2024, 44(12): 3543-3552.
HUANG Lin-feng, JIANG Xue-song, JIA Zhi-cheng, ZHOU Hong-ping, ZHOU Lei, RONG Zi-fan. Deep Learning-Based Monitoring of Nutrient Content in Pear Trees. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2024, 44(12): 3543-3552.

链接本文:

https://www.gpxygpfx.com/CN/10.3964/j.issn.1000-0593(2024)12-3543-10 或 https://www.gpxygpfx.com/CN/Y2024/V44/I12/3543