基于高光谱和卷积自编码网络的茄子叶片氮含量估测  

Estimation of Eggplant Leaf Nitrogen Content Based on Hyperspectral Imaging and Convolutional Auto-Encoders Networks

作　　者：王浩宇 魏子渊 杨永霞 侯军英 孙章彤 胡瑾[1,2,3] WANG Hao-yu;WEI Zi-yuan;YANG Yong-xia;HOU Jun-ying;SUN Zhang-tong;HU Jin(College of Mechanical and Electronic Engineering,Northwest A&F University,Yangling 712100,China;Key Laboratory of Agricultural Internet of Things,Ministry of Agric ulture and Rural Affairs,Yangling 712100,China;Key Laboratory of Agricultural Information Awareness and Intelligent Services,Yangling 712100,China)

机构地区：[1]西北农林科技大学机械与电子工程学院,陕西杨凌712100 [2]农业农村部农业物联网重点研究实验室,陕西杨凌712100 [3]陕西省农业信息感知与智能服务重点实验室,陕西杨凌712100

出　　处：《光谱学与光谱分析》2024年第8期2208-2215,共8页Spectroscopy and Spectral Analysis

基　　金：国家重点研发计划项目(2020YFD1100602);陕西省重点研发计划项目(2021ZDLNY03-02)资助。

摘　　要：叶片氮含量是评估植物生长状态与光合能力的重要依据之一,准确获取叶片氮含量有助于合理调控氮肥施用,对实现农业高效生产具有重要意义。化学分析方法虽能精准检测氮含量,但需破坏性采样,且步骤繁琐,难以实时大量测量。利用光谱技术可以实现叶片氮含量的无损检测,但受光谱数据固有的高维度、高噪声等特点影响,使得氮含量估测精度难以满足精准农业需求。为实现对叶片氮含量的准确预测,研究提出一种基于高光谱成像(HSI)技术和一维卷积自编码网络(CAE)结合的光谱数据特征提取方法,采用像素级光谱数据训练CAE网络,充分利用叶片HSI数据,保留由于氮含量在叶面上差异性分布产生的局部光谱特征信息,实现对光谱数据的深度特征提取,降低数据维度并滤除噪声,提高建立氮含量预测模型的精度与稳定性。以茄子为研究对象,设置四个施氮梯度,培养获取不同氮含量叶片样本并测定其HSI数据。采用多元散射校正算法进行数据预处理,分别使用HSI-CAE方法、竞争性自适应重加权(CARS)算法和随机蛙跳(RF)算法提取光谱数据深度特征和特征波长,并建立偏最小二乘回归(PLSR)模型。对比深度特征和不同特征波长组合对预测模型精度的影响,从而确定最优特征提取方法。结果表明,利用不同深度CAE编码器提取深度特征建立的预测模型,测试集决定系数均大于0.85,且当输出特征为28维时,测试集决定系数为0.9102,均方根误差为3.1189 mg·g^(-1),对比基于CARS与RF特征波长提取算法建立的预测模型,发现CAE-PLSR模型预测性能最优,验证了HSI-CAE特征提取方法用于茄子叶片氮含量估测的可行性和优越性。综上,HSI-CAE特征提取方法能够高效分析HSI数据,提取其深度特征。这些特征中含有与氮含量高度相关的信息,使用深度特征建模极大降低了模型的复杂度,有效提高了氮含量预测模�Leaf nitrogen content(LNC)is crucial for assessing plant growth status and photosynthetic capacity.Accurate LNC can aid in the rational control of nitrogen fertilizer application,which is critical for achieving efficient agricultural production.Chemical analysis methods can accurately detect nitrogen content.However,it often requires destructive sampling and cumbersome steps,which are difficult to measure in real-time.Spectral technology can enable nondestructive detection of LNC,but the high dimensionality and noise inherent in spectral data make accurate estimation challenging for precision agriculture.To achieve accurate prediction of nitrogen content in eggplant leaves,this paper proposed a feature extraction method of spectral data based on hyperspectral imaging(HSI)technology and a one-dimensional convolutional autoencoder network(CAE).The proposed method utilized pixel-level spectral data to train the CAE,fully utilizing the HSI data of leaves.This can extract deep features that retain local spectral features related to the nitrogen content distribution on the leaf surface,reducing data dimension,filtering out noise,and enhancing the accuracy and stability of the nitrogen content prediction model.In this paper,we set up four nitrogen application gradients for eggplant,obtained leaf samples with varying nitrogen content using culture,and measured their HIS data.Multiple scattering correction algorithm was used for data preprocessing.The HSI-CAE method,competitive adaptive reweighting(CARS)algorithm,and random frog(RF)algorithm were used to extract spectral data's deep features and characteristic wavelengths,respectively.The partial least squares regression(PLSR)models were built based on these features.The influence of deep features and different feature wavelength combinations on the accuracy of the prediction model was compared to determine the optimal feature extraction method.The results were as follows:the test set determination coefficient of the prediction model,established by using deep features fro

关 键 词：高光谱成像 卷积自编码 特征提取 氮含量 

分 类 号：O657.3[理学—分析化学]