基于深度学习的甘蓝叶片超氧化物歧化酶活性模型构建  

Construction of Superoxide Dismutase Activity Model of Cabbage Leaves Based on Deep Learning

作　　者：马思艳 马玲[1] 马燕[1] 王静[1] 张祎洋 吴龙国[1,2] MA Si-yan;MA Ling;MA Yan(College of Wine and Horticulture,Ningxia University,Yinchuan,Ningxia 750021)

机构地区：[1]宁夏大学葡萄酒与园艺学院,宁夏银川750021 [2]宁夏现代设施园艺工程技术研究中心,宁夏银川750021

出　　处：《安徽农业科学》2025年第1期229-235,共7页Journal of Anhui Agricultural Sciences

基　　金：国家重点研发计划项目子课题专项“集约化育苗系列机械优化”(2021YFD1600302-3)。

摘　　要：超氧化物歧化酶(SOD)是判断作物受胁迫程度的关键指标,可以用来衡量植物生长状况,同时也在植物逆境胁迫研究中具有重要意义。为了实现甘蓝叶片SOD活性的快速无损检测,提出了一种利用高光谱成像技术结合深度学习的方法来对甘蓝叶片进行检测。试验共采集了200片甘蓝叶片在不同生长发育时期的光谱信息,通过7种预处理方法对原始光谱进行优化,最后选用高斯滤波方法为SOD活性的预处理方法。采用连续投影算法、无信息变量消除算法、遗传偏最小二乘算法、竞争自适应重加权采样算法和区间变量迭代空间收缩分析算法提取特征波长,建立偏最小二乘回归模型。基于优选的特征波长建立PLSR、主成分回归、多元线性回归、最小二乘支持向量机和深度学习模型。结果表明,CARS算法提取的17个最佳波长效果较好,最优预测模型CNN的相关系数R c和R p值分别为0.9098和0.8235,均方根误差RMSEC和RMSEP分别为2.0382和3.6492。该研究为今后盐胁迫下植株长势在线无损监测提供技术支撑,具有良好的发展前景。Superoxide dismutase(SOD)is a key index to judge the degree of crop stress,which can be used to measure plant growth status,and also has important significance in the study of plant stress.In order to achieve rapid nondestructive detection of SOD activity in cabbage leaves,a method of hyperspectral imaging(HSI)combined with deep learning was proposed to detect cabbage leaves.In the experiment,spectral information of 200 cabbage leaves at different growth and development stages was collected,and the original spectra were optimized by 7 pretreatment methods through sample set division.Finally,the Gaussian Filter(GF)method was selected as the pretreatment method for SOD activity.Successive projection algorithm(SPA),uninformative variable elimination algorithm(UVE),genetic algorithm-partial least squares algorithm(GAPLS),competitive adaptive reweighted sampling(CARS)and interval variable iterative space shrinking analysis(IVISSA)were used to extract feature wavelengths and partial least squares regression(PLSR)model was established.PLSR,principal component regression(PCR),multiple linear regression(MLR),least square SVM(LSSVM)and convolutional neural network(CNN)models were established based on the preferred characteristic wavelength.The results showed that the 17 optimal wavelengths extracted by CARS algorithm had the best effect,and the correlation coefficients R c and R p values of the optimal prediction model CNN were 0.9098 and 0.8235,respectively.And the root-mean-square error RMSEC and RMSEP were 2.0382 and 3.6492,respectively.This study provided technical support for non-destructive on-line monitoring of plant growth under salt stress in the future,and had a good development prospect.

关 键 词：高光谱成像 深度学习 SOD活性 无损检测 

分 类 号：S-058[农业科学]