基于高光谱图像的生菜叶片氮素含量预测模型研究
孙俊等
摘要为了便于更经济合理地为作物施肥，建立一种无损检测作物氮营养元素的高光谱图像模型。本实验以生菜为研究对象，无土栽培各氮素水平的生菜叶样本，在莲座期，采集生菜叶片样本的高光谱图像（390～1050 nm），同时采用凯氏定氮法测定对应生菜叶片样本的全氮含量。通过ENVI软件提取出生菜叶片中感兴趣区域的平均光谱作为该样本原始光谱信息，分别使用平滑处理（Smoothing）、多元散射矫正（MSC）、标准正态变量变换结合去趋势（SNV detrending）、一阶导数法（First derivative）、二阶导数法（Second derivative）、正交信号矫正（OSC）等预处理方法对样本原始光谱进行处理，然后利用偏最小二乘回归法（Partial least squares regression，PLSR）分别建立样本全波段光谱信息与氮含量的关系模型，研究各预处理方法对氮含量模型的影响，结果表明， 使用OSC预处理的模型效果最好。为了简化模型，根据OSC预处理光谱后的模型的PLSR回归系数优选出敏感波长，利用训练集中样本的敏感波长光谱信息与氮含量数据重新构建PLSR回归模型，并利用测试集样本进行测试试验。结果表明，该模型得到校正集和预测集的决定系数（R2p）分别为0.89， 0.81； 均方根误差RMSEC， RMSEP分别为0.33， 0.45。该回归模型大大降低了自变量个数， 简化了模型，并且取得了较优的效果，这为生菜氮素含量预测提供了一种新的快速有效方法。关键词高光谱图像；生菜叶片；氮素；敏感波长；偏最小二乘回归法
引言
生菜富含蛋白质、碳水化合物、维生素等营养物质，且种植面积广。氮素是影响生菜生长的最主要元素之一，建立快速、高效、无损的生菜的氮素含量检测方法，便于对生菜的合理施肥进行指导。
目前，对作物营养元素的无损检测研究方法主要集中在计算机视觉技术、光谱技术，如毛罕平等利用计算机视觉技术结合神经网络实现了番茄中各营养元素亏缺情况的识别
References
1MAO HanPing， WU XueMei， LI PingPing. Transactions of the CSAE.， 2005， 21（8）： 106-109
毛罕平， 吴雪梅， 李萍萍. 农业工程学报.， 2005， 21（8）： 106-109
2SHI JiYong， ZOU XiaoBo， ZHAO JieWen， MAO HanPing， WANG KaiLiang， CHEN ZhengWei， HUANG XiaoWei. Transactions of the CSAE.， 2011， 27（8）： 283-287
石吉勇， 邹小波， 赵杰文， 毛罕平， 王开亮， 陈正伟， 黄晓玮. 农业工程学报， 2011， 27（8）： 283-287
3Ulissi V， Antonucci F， Benincasa P， Farneselli M， Tosti G， Guiducci M， Tei F， Costa C， Pallottino F， Pari L， Menesatti P. Sensors， 2011， 11（6）： 6411-6424
4Barbin D F， Eimasry G， Sun D W， Allen P. Anal. Chim. Acta， 2012， 719： 30-42
5Baranowski P， Mazurek W， Wozniak J， Majewska U. J. Food Enginering， 2012， 110（3）： 345-355
6WANG ZhiHui， DING LiXia. Spectroscopy and Spectral Analysis， 2010， 30（7）： 1825-1829
王志辉， 丁丽霞. 光谱学与光谱分析， 2010， 30（7）： 1825-1829
7Huang M， Wan X M， Zhang M， Zhu Q B. J. Food Enginering， 2013， 116（1）： 45-49
8Zhang X L， Liu F， He Y， Gong X Y. Biosystems Enginering， 2013， 115（1）： 56-65
9Vigneau N， Ecarnot M， Rabatel G， Rounmet P. Field Crops Research， 2011， 122（1）： 25-31
10SUN Jun， WU XiaoHong， ZHANG XiaoDong， GAO HongYan. Spectroscopy and Spectral Analysis， 2013， 33（2）： 522-526
孙 俊， 武小红， 张晓东， 高洪燕. 光谱学与光谱分析.， 2013， 33（2）： 522-526
11PAN Bei， ZHAO GengXing， ZHU XiCun， LIU HaiTeng， LIANG Shuang， TIAN DaDe. Spectroscopy and Spectral Analysis， 2013， 33（8）： 2203-2206
潘 蓓， 赵庚星， 朱西存， 刘海腾， 梁 爽， 田大德. 光谱学与光谱分析.， 2013， 33（8）： 2203-2206
12Shetty N， Gislum R. Field Crops Research， 2011， 120（1）： 31-37
A Model for Predicting Nitrogen of Lettuce Leaves
Based on Hyperspectral Imaging
SUN Jun1， JIN XiaMing1， MAO HanPing2， WU XiaoHong1， ZHANG XiaoDong2， GAO HongYan2
1（School of Electrical and Information Engineering of Jiangsu university， Zhenjiang 212013， China）
2（Laboratory Venlo of Modern Agricultural Equipment of Jiangsu University， Zhenjiang 212013， China）
AbstractThe goal of this study was to study on the model of hyperspectral imaging for detecting the nitrogen of crops nondestructively. Lettuce， as the research object， was cultivated in different levels of nitrogen in soilless cultivation method. In rosette stage， the hyperspectral images of lettuce leaves （390-1050 nm） were collected， and nitrogen content of the corresponding lettuce leaves were determined by Kjeldahl method. Then the average spectral data of region of interest of lettuce leaves were extracted by ENVI software， and Smoothing， Multiplicative scatter correction （Multiplication scatter correction， MSC）， Standard normal variate transformation combined detrending （Standard normalized variable+detrending， SNV+detrending）， First derivative， Second derivative， Orthogonal signal correction （Orthogonal signal correction， OSC） were used for pretreating the extracted raw spectral data respectively. Partial least squares regression （PLSR） was used to correlate the reflectance spectral of whole wavelengths with nitrogen content of lettuce leaves respectively， and the pretreatment methods above were studied on the influence of the models. The results showed that the model using OSC pretreatment method was the best. In addition， according to the regression coefficient of OSC+PLSR model， sensitive wavelengths were selected to simplify the model. The spectral data at sensitive wavelengths in calibration set were reconstructed for the model，and prediction set was used to test the model. The results showed that determination coefficients （R2c， R2p） from calibration set and prediction set were 0.89， 0.81， and the root meansquare error of calibration （RMSEC） and root meansquare error of prediction （RMSEP） were 0.33 and 0.45， respectively. An easier model with good performance was developed in this study， and it could provide an effective modeling method for predicting the nitrogen content in lettuce leaves.
KeywordsHyperspectral imaging； Lettuce leaves； Nitrogen； Sensitive wavelengths； Partial least squares regression
（Received 19 November 2013； accepted 22 January 2014）
This work was supported by the National Natural Science
Based on Hyperspectral Imaging
SUN Jun1， JIN XiaMing1， MAO HanPing2， WU XiaoHong1， ZHANG XiaoDong2， GAO HongYan2
1（School of Electrical and Information Engineering of Jiangsu university， Zhenjiang 212013， China）
2（Laboratory Venlo of Modern Agricultural Equipment of Jiangsu University， Zhenjiang 212013， China）
AbstractThe goal of this study was to study on the model of hyperspectral imaging for detecting the nitrogen of crops nondestructively. Lettuce， as the research object， was cultivated in different levels of nitrogen in soilless cultivation method. In rosette stage， the hyperspectral images of lettuce leaves （390-1050 nm） were collected， and nitrogen content of the corresponding lettuce leaves were determined by Kjeldahl method. Then the average spectral data of region of interest of lettuce leaves were extracted by ENVI software， and Smoothing， Multiplicative scatter correction （Multiplication scatter correction， MSC）， Standard normal variate transformation combined detrending （Standard normalized variable+detrending， SNV+detrending）， First derivative， Second derivative， Orthogonal signal correction （Orthogonal signal correction， OSC） were used for pretreating the extracted raw spectral data respectively. Partial least squares regression （PLSR） was used to correlate the reflectance spectral of whole wavelengths with nitrogen content of lettuce leaves respectively， and the pretreatment methods above were studied on the influence of the models. The results showed that the model using OSC pretreatment method was the best. In addition， according to the regression coefficient of OSC+PLSR model， sensitive wavelengths were selected to simplify the model. The spectral data at sensitive wavelengths in calibration set were reconstructed for the model，and prediction set was used to test the model. The results showed that determination coefficients （R2c， R2p） from calibration set and prediction set were 0.89， 0.81， and the root meansquare error of calibration （RMSEC） and root meansquare error of prediction （RMSEP） were 0.33 and 0.45， respectively. An easier model with good performance was developed in this study， and it could provide an effective modeling method for predicting the nitrogen content in lettuce leaves.
KeywordsHyperspectral imaging； Lettuce leaves； Nitrogen； Sensitive wavelengths； Partial least squares regression
（Received 19 November 2013； accepted 22 January 2014）
This work was supported by the National Natural Science
Based on Hyperspectral Imaging
SUN Jun1， JIN XiaMing1， MAO HanPing2， WU XiaoHong1， ZHANG XiaoDong2， GAO HongYan2
1（School of Electrical and Information Engineering of Jiangsu university， Zhenjiang 212013， China）
2（Laboratory Venlo of Modern Agricultural Equipment of Jiangsu University， Zhenjiang 212013， China）
AbstractThe goal of this study was to study on the model of hyperspectral imaging for detecting the nitrogen of crops nondestructively. Lettuce， as the research object， was cultivated in different levels of nitrogen in soilless cultivation method. In rosette stage， the hyperspectral images of lettuce leaves （390-1050 nm） were collected， and nitrogen content of the corresponding lettuce leaves were determined by Kjeldahl method. Then the average spectral data of region of interest of lettuce leaves were extracted by ENVI software， and Smoothing， Multiplicative scatter correction （Multiplication scatter correction， MSC）， Standard normal variate transformation combined detrending （Standard normalized variable+detrending， SNV+detrending）， First derivative， Second derivative， Orthogonal signal correction （Orthogonal signal correction， OSC） were used for pretreating the extracted raw spectral data respectively. Partial least squares regression （PLSR） was used to correlate the reflectance spectral of whole wavelengths with nitrogen content of lettuce leaves respectively， and the pretreatment methods above were studied on the influence of the models. The results showed that the model using OSC pretreatment method was the best. In addition， according to the regression coefficient of OSC+PLSR model， sensitive wavelengths were selected to simplify the model. The spectral data at sensitive wavelengths in calibration set were reconstructed for the model，and prediction set was used to test the model. The results showed that determination coefficients （R2c， R2p） from calibration set and prediction set were 0.89， 0.81， and the root meansquare error of calibration （RMSEC） and root meansquare error of prediction （RMSEP） were 0.33 and 0.45， respectively. An easier model with good performance was developed in this study， and it could provide an effective modeling method for predicting the nitrogen content in lettuce leaves.
KeywordsHyperspectral imaging； Lettuce leaves； Nitrogen； Sensitive wavelengths； Partial least squares regression
（Received 19 November 2013； accepted 22 January 2014）
This work was supported by the National Natural Science
摘要为了便于更经济合理地为作物施肥，建立一种无损检测作物氮营养元素的高光谱图像模型。本实验以生菜为研究对象，无土栽培各氮素水平的生菜叶样本，在莲座期，采集生菜叶片样本的高光谱图像（390～1050 nm），同时采用凯氏定氮法测定对应生菜叶片样本的全氮含量。通过ENVI软件提取出生菜叶片中感兴趣区域的平均光谱作为该样本原始光谱信息，分别使用平滑处理（Smoothing）、多元散射矫正（MSC）、标准正态变量变换结合去趋势（SNV detrending）、一阶导数法（First derivative）、二阶导数法（Second derivative）、正交信号矫正（OSC）等预处理方法对样本原始光谱进行处理，然后利用偏最小二乘回归法（Partial least squares regression，PLSR）分别建立样本全波段光谱信息与氮含量的关系模型，研究各预处理方法对氮含量模型的影响，结果表明， 使用OSC预处理的模型效果最好。为了简化模型，根据OSC预处理光谱后的模型的PLSR回归系数优选出敏感波长，利用训练集中样本的敏感波长光谱信息与氮含量数据重新构建PLSR回归模型，并利用测试集样本进行测试试验。结果表明，该模型得到校正集和预测集的决定系数（R2p）分别为0.89， 0.81； 均方根误差RMSEC， RMSEP分别为0.33， 0.45。该回归模型大大降低了自变量个数， 简化了模型，并且取得了较优的效果，这为生菜氮素含量预测提供了一种新的快速有效方法。关键词高光谱图像；生菜叶片；氮素；敏感波长；偏最小二乘回归法
引言
生菜富含蛋白质、碳水化合物、维生素等营养物质，且种植面积广。氮素是影响生菜生长的最主要元素之一，建立快速、高效、无损的生菜的氮素含量检测方法，便于对生菜的合理施肥进行指导。
目前，对作物营养元素的无损检测研究方法主要集中在计算机视觉技术、光谱技术，如毛罕平等利用计算机视觉技术结合神经网络实现了番茄中各营养元素亏缺情况的识别
References
1MAO HanPing， WU XueMei， LI PingPing. Transactions of the CSAE.， 2005， 21（8）： 106-109
毛罕平， 吴雪梅， 李萍萍. 农业工程学报.， 2005， 21（8）： 106-109
2SHI JiYong， ZOU XiaoBo， ZHAO JieWen， MAO HanPing， WANG KaiLiang， CHEN ZhengWei， HUANG XiaoWei. Transactions of the CSAE.， 2011， 27（8）： 283-287
石吉勇， 邹小波， 赵杰文， 毛罕平， 王开亮， 陈正伟， 黄晓玮. 农业工程学报， 2011， 27（8）： 283-287
3Ulissi V， Antonucci F， Benincasa P， Farneselli M， Tosti G， Guiducci M， Tei F， Costa C， Pallottino F， Pari L， Menesatti P. Sensors， 2011， 11（6）： 6411-6424
4Barbin D F， Eimasry G， Sun D W， Allen P. Anal. Chim. Acta， 2012， 719： 30-42
5Baranowski P， Mazurek W， Wozniak J， Majewska U. J. Food Enginering， 2012， 110（3）： 345-355
6WANG ZhiHui， DING LiXia. Spectroscopy and Spectral Analysis， 2010， 30（7）： 1825-1829
王志辉， 丁丽霞. 光谱学与光谱分析， 2010， 30（7）： 1825-1829
7Huang M， Wan X M， Zhang M， Zhu Q B. J. Food Enginering， 2013， 116（1）： 45-49
8Zhang X L， Liu F， He Y， Gong X Y. Biosystems Enginering， 2013， 115（1）： 56-65
9Vigneau N， Ecarnot M， Rabatel G， Rounmet P. Field Crops Research， 2011， 122（1）： 25-31
10SUN Jun， WU XiaoHong， ZHANG XiaoDong， GAO HongYan. Spectroscopy and Spectral Analysis， 2013， 33（2）： 522-526
孙 俊， 武小红， 张晓东， 高洪燕. 光谱学与光谱分析.， 2013， 33（2）： 522-526
11PAN Bei， ZHAO GengXing， ZHU XiCun， LIU HaiTeng， LIANG Shuang， TIAN DaDe. Spectroscopy and Spectral Analysis， 2013， 33（8）： 2203-2206
潘 蓓， 赵庚星， 朱西存， 刘海腾， 梁 爽， 田大德. 光谱学与光谱分析.， 2013， 33（8）： 2203-2206
12Shetty N， Gislum R. Field Crops Research， 2011， 120（1）： 31-37
A Model for Predicting Nitrogen of Lettuce Leaves
Based on Hyperspectral Imaging
SUN Jun1， JIN XiaMing1， MAO HanPing2， WU XiaoHong1， ZHANG XiaoDong2， GAO HongYan2
1（School of Electrical and Information Engineering of Jiangsu university， Zhenjiang 212013， China）
2（Laboratory Venlo of Modern Agricultural Equipment of Jiangsu University， Zhenjiang 212013， China）
AbstractThe goal of this study was to study on the model of hyperspectral imaging for detecting the nitrogen of crops nondestructively. Lettuce， as the research object， was cultivated in different levels of nitrogen in soilless cultivation method. In rosette stage， the hyperspectral images of lettuce leaves （390-1050 nm） were collected， and nitrogen content of the corresponding lettuce leaves were determined by Kjeldahl method. Then the average spectral data of region of interest of lettuce leaves were extracted by ENVI software， and Smoothing， Multiplicative scatter correction （Multiplication scatter correction， MSC）， Standard normal variate transformation combined detrending （Standard normalized variable+detrending， SNV+detrending）， First derivative， Second derivative， Orthogonal signal correction （Orthogonal signal correction， OSC） were used for pretreating the extracted raw spectral data respectively. Partial least squares regression （PLSR） was used to correlate the reflectance spectral of whole wavelengths with nitrogen content of lettuce leaves respectively， and the pretreatment methods above were studied on the influence of the models. The results showed that the model using OSC pretreatment method was the best. In addition， according to the regression coefficient of OSC+PLSR model， sensitive wavelengths were selected to simplify the model. The spectral data at sensitive wavelengths in calibration set were reconstructed for the model，and prediction set was used to test the model. The results showed that determination coefficients （R2c， R2p） from calibration set and prediction set were 0.89， 0.81， and the root meansquare error of calibration （RMSEC） and root meansquare error of prediction （RMSEP） were 0.33 and 0.45， respectively. An easier model with good performance was developed in this study， and it could provide an effective modeling method for predicting the nitrogen content in lettuce leaves.
KeywordsHyperspectral imaging； Lettuce leaves； Nitrogen； Sensitive wavelengths； Partial least squares regression
（Received 19 November 2013； accepted 22 January 2014）
This work was supported by the National Natural Science
Based on Hyperspectral Imaging
SUN Jun1， JIN XiaMing1， MAO HanPing2， WU XiaoHong1， ZHANG XiaoDong2， GAO HongYan2
1（School of Electrical and Information Engineering of Jiangsu university， Zhenjiang 212013， China）
2（Laboratory Venlo of Modern Agricultural Equipment of Jiangsu University， Zhenjiang 212013， China）
AbstractThe goal of this study was to study on the model of hyperspectral imaging for detecting the nitrogen of crops nondestructively. Lettuce， as the research object， was cultivated in different levels of nitrogen in soilless cultivation method. In rosette stage， the hyperspectral images of lettuce leaves （390-1050 nm） were collected， and nitrogen content of the corresponding lettuce leaves were determined by Kjeldahl method. Then the average spectral data of region of interest of lettuce leaves were extracted by ENVI software， and Smoothing， Multiplicative scatter correction （Multiplication scatter correction， MSC）， Standard normal variate transformation combined detrending （Standard normalized variable+detrending， SNV+detrending）， First derivative， Second derivative， Orthogonal signal correction （Orthogonal signal correction， OSC） were used for pretreating the extracted raw spectral data respectively. Partial least squares regression （PLSR） was used to correlate the reflectance spectral of whole wavelengths with nitrogen content of lettuce leaves respectively， and the pretreatment methods above were studied on the influence of the models. The results showed that the model using OSC pretreatment method was the best. In addition， according to the regression coefficient of OSC+PLSR model， sensitive wavelengths were selected to simplify the model. The spectral data at sensitive wavelengths in calibration set were reconstructed for the model，and prediction set was used to test the model. The results showed that determination coefficients （R2c， R2p） from calibration set and prediction set were 0.89， 0.81， and the root meansquare error of calibration （RMSEC） and root meansquare error of prediction （RMSEP） were 0.33 and 0.45， respectively. An easier model with good performance was developed in this study， and it could provide an effective modeling method for predicting the nitrogen content in lettuce leaves.
KeywordsHyperspectral imaging； Lettuce leaves； Nitrogen； Sensitive wavelengths； Partial least squares regression
（Received 19 November 2013； accepted 22 January 2014）
This work was supported by the National Natural Science
Based on Hyperspectral Imaging
SUN Jun1， JIN XiaMing1， MAO HanPing2， WU XiaoHong1， ZHANG XiaoDong2， GAO HongYan2
1（School of Electrical and Information Engineering of Jiangsu university， Zhenjiang 212013， China）
2（Laboratory Venlo of Modern Agricultural Equipment of Jiangsu University， Zhenjiang 212013， China）
AbstractThe goal of this study was to study on the model of hyperspectral imaging for detecting the nitrogen of crops nondestructively. Lettuce， as the research object， was cultivated in different levels of nitrogen in soilless cultivation method. In rosette stage， the hyperspectral images of lettuce leaves （390-1050 nm） were collected， and nitrogen content of the corresponding lettuce leaves were determined by Kjeldahl method. Then the average spectral data of region of interest of lettuce leaves were extracted by ENVI software， and Smoothing， Multiplicative scatter correction （Multiplication scatter correction， MSC）， Standard normal variate transformation combined detrending （Standard normalized variable+detrending， SNV+detrending）， First derivative， Second derivative， Orthogonal signal correction （Orthogonal signal correction， OSC） were used for pretreating the extracted raw spectral data respectively. Partial least squares regression （PLSR） was used to correlate the reflectance spectral of whole wavelengths with nitrogen content of lettuce leaves respectively， and the pretreatment methods above were studied on the influence of the models. The results showed that the model using OSC pretreatment method was the best. In addition， according to the regression coefficient of OSC+PLSR model， sensitive wavelengths were selected to simplify the model. The spectral data at sensitive wavelengths in calibration set were reconstructed for the model，and prediction set was used to test the model. The results showed that determination coefficients （R2c， R2p） from calibration set and prediction set were 0.89， 0.81， and the root meansquare error of calibration （RMSEC） and root meansquare error of prediction （RMSEP） were 0.33 and 0.45， respectively. An easier model with good performance was developed in this study， and it could provide an effective modeling method for predicting the nitrogen content in lettuce leaves.
KeywordsHyperspectral imaging； Lettuce leaves； Nitrogen； Sensitive wavelengths； Partial least squares regression
（Received 19 November 2013； accepted 22 January 2014）
This work was supported by the National Natural Science
