Inversion of pepper SPAD values from UAV hyperspectral data

doi:10.16178/j.issn.0528-9017.20230980

Abstract

Abstract:

In order to establish a more stable and predictive inversion model of chlorophyll content in pepper, this study analyzed the correlation between the original spectrum and other transformation spectra and the chlorophyll relative content (SPAD value) based on the UAV hyperspectral data and SPAD value. The maximum correlation coefficient method (MCC) was used to select the feature bands with good correlation to generate the feature band dataset, and then the genetic algorithm-partial least squares method (GAPLS) was used to reduce the dimension to obtain the optimal feature band combination. Five machine learning algorithms, namely partial least squares (PLSR), backpropagation neural network (BPNN), random forest (RF), least squares support vector machine (LSSVM) and genetic algorithm optimized least squares support vector machine (GA-LSSVM), were used to construct a chlorophyll content inversion model of pepper. The results showed that the SPAD value of pepper leaves was inversely proportional to the hyperspectral reflectance. The sensitive band of chlorophyll in pepper was mainly concentrated in 400~700 nm. The first-order differential spectra had the best correlation with the SPAD value, and the second-order differential spectra at 671 nm wavelength had the largest negative correlation with chlorophyll content, with a correlation coefficient of -0.69. The models based on reciprocal logarithmic spectra generally had high accuracy. The best performance of the model was the GA-LSSVM model based on differential spectroscopy, with the coefficient of determination (R²), root mean square error (RMSE) and relative analysis error (RPD) values of 0.84, 1.41 and 2.24, respectively, followed by the RF model based on reciprocal logarithmic spectroscopy, with R², RMSE and RPD values of 0.83, 1.57 and 2.13, respectively.

Key words: pepper, chlorophyll, canopy leaf, UAV hyperspectral, combined algorithm

CLC Number:

S641.3

WANG Yu, YANG Kun, RAO Weidong, FENG Difei, WANG Hong, XIAO Jiujun, ZHANG Shengguo. Inversion of pepper SPAD values from UAV hyperspectral data[J]. Journal of Zhejiang Agricultural Sciences, 2024, 65(10): 2323-2337.

Figures/Tables 31

Fig.1 Spectral reflectance corresponding to different SPAD values

Fig.2 Correlation coefficient between original spectrum and SPAD value

Fig.3 Correlation coefficient between reciprocal logarithmic spectrum and SPAD value

Fig.4 Correlation coefficient between MSC spectra and SPAD values

Fig.5 Correlation coefficient between continuum removal spectrum and SPAD value

Fig.6 Correlation coefficient between first order differential spectrum and SPAD value

Table 1 Contribution rate and RMSECV table based on the original spectrum

选择变量数	主成分数	贡献率/%	RMSECV
1	1	94.75	0.172 9
2	2	97.82	0.213 4
3	2	99.75	0.018 7
4	4	99.82	0.001 3
5	4	99.81	0.001 4
…	…	…	…

Table 2 The frequency sequence of selected bands based on the original spectrum

排序	波段	被选用频次
1	20	100
2	4	51
3	1	46
4	2	45
…	…	…

Fig.7 The frequency of the reciprocal logarithmic spectral band

Table 3 Contribution rate and RMSECV table based on reciprocal logarithmic spectra

选择变量数	主成分数	贡献率/%	RMSECV
1	1	91.75	0.256 2
2	2	93.82	0.167 1
4	3	99.77	0.001 6
6	5	99.85	0.001 2
7	5	99.73	0.002 1
…	…	…	…

Table 4 Screening bands based on reciprocal logarithmic spectra and their corresponding selection frequency sequences

排序	波段	被选用频次
1	19	82
2	3	52
3	5	41
4	4	40
5	6	39
6	8	39
…	…	…

Fig.8 The frequency of reciprocal logarithmic spectral bands

Table 5 Contribution rate and RMSECV based on multivariate scattering correction spectra

选择变量数	主成分数	贡献率/%	RMSECV
1	1	93.35	0.372 7
2	2	95.74	0.187 2
3	2	99.29	0.073 1
4	4	99.93	0.006 8
5	4	99.82	0.007 1
…	…	…	…

Table 6 Screening bands based on multivariate scattering correction spectra and their corresponding selection frequency sequence

排序	波段	被选用频次
1	20	79
2	4	52
3	3	48
4	6	41
…	…	…

Fig.9 The frequency of multivariate scattering correction spectral bands

Table 7 Contribution rate and RMSECV table based on remove envelope spectrum

选择变量数	主成分数	贡献率/%	RMSECV
1	1	84.45	0.342 3
2	2	91.45	0.135 6
3	3	98.54	0.021 2
5	3	98.34	0.024 5
…	…	…	…

Table 8 The sequence of bands screened based on remove envelope spectrum and their corresponding selection frequencies

排序	波段	被选用频次
1	20	77
2	15	58
3	19	58
…	…	…

Fig.10 Frequency diagram of the selection of spectral bands for remove envelope spectrum

Table 9 Contribution rate and RMSECV table based on first-order differential spectroscopy

选择变量数	主成分数	贡献率/%	RMSECV
1	1	68.18	0.559 8
4	3	99.29	0.124 8
5	4	99.56	0.010 7
7	4	99.86	0.010 6
8	4	99.67	0.010 8
…	…	…	…

Table 10 The screening bands based on the first-order differential spectroscopy and the corresponding frequency sequence

排序	波段	被选用频次/次
1	2	94
2	12	53
3	13	41
4	14	37
5	3	36
6	4	35
7	15	26
…	…	…

Fig.11 Frequency diagram of the selection of the first-order differential spectral band

Table 11 SPAD value inversion model based on PLSR

参数	建模集			验证集
参数	R²	RMSE	RPD	R²	RMSE	RPD
R	0.44	6.37	1.35	0.48	2.28	1.39
lg(1/R)	0.63	5.20	1.65	0.66	1.80	1.76
M(R)	0.54	5.94	1.44	0.48	2.26	1.40
RE	0.56	6.03	1.42	0.54	2.49	1.27
R'	0.69	4.74	1.81	0.71	2.22	1.42

Fig.12 Fitting analysis of SPAD values for PLSR prediction model

Table 12 SPAD value inversion model based on RF

参数	建模集			验证集
参数	R²	RMSE	RPD	R²	RMSE	RPD
R	0.53	6.14	1.40	0.58	2.51	1.26
lg(1/R)	0.81	4.21	2.04	0.83	1.57	2.13
M(R)	0.60	6.05	1.42	0.59	2.22	1.43
RE	0.67	5.16	1.66	0.64	2.57	1.23
R'	0.68	4.97	1.73	0.70	2.55	1.24

Fig.13 Fitting analysis of SPAD values for RF prediction model

Table 13 SPAD value inversion model based on BPNN

参数	建模集			验证集
参数	R²	RMSE	RPD	R²	RMSE2	RPD
R	0.41	6.52	1.32	0.46	2.35	1.36
lg(1/R)	0.71	4.78	1.78	0.77	1.67	1.90
M(R)	0.64	5.27	1.63	0.46	2.63	1.20
RE	0.70	4.71	1.82	0.79	1.54	2.05
R'	0.74	4.34	1.98	0.57	2.12	1.49

Fig.14 Fitting analysis of SPAD values for BPNN prediction model

Table 14 SPAD value inversion model based on LSSVM

参数	建模集			验证集
参数	R²	RMSE	RPD	R²	RMSE	RPD
R	0.51	5.98	1.43	0.36	2.77	1.14
lg(1/R)	0.74	4.63	1.85	0.68	1.85	1.71
M(R)	0.70	5.27	1.63	0.61	2.02	1.57
RE	0.60	5.48	1.57	0.49	2.40	1.32
R'	0.72	4.63	1.85	0.71	2.58	1.23

Fig.15 SPAD value fitting analysis of LSSVM prediction model

Table 15 SPAD value inversion model based on GA-LSSVM

参数	建模集			验证集
参数	R²	RMSE	RPD	R²	RMSE	RPD
R	0.59	5.82	1.47	0.52	2.36	1.34
lg(1/R)	0.70	4.73	1.81	0.77	1.67	1.89
M(R)	0.58	5.90	1.45	0.66	1.90	1.66
RE	0.58	5.72	1.50	0.56	2.47	1.28
R'	0.81	3.57	2.40	0.84	1.41	2.24

Fig.16 SPAD value fitting analysis of GA-LSSVM prediction model

References 32

[1]	张沛健, 尚秀华, 吴志华. 基于图像处理技术的5种红树林叶片形态特征及叶绿素相对含量的估测[J]. 热带作物学报, 2020, 41(3): 496-503.
[2]	JANG S, HWANG Y, LEE H J, et al. Selecting significant wavelengths to predict chlorophyll content of grafted cucumber seedlings using hyperspectral images[J]. Journal of Remote Sensing, 2018, 34: 681-692.
[3]	陈澜, 常庆瑞, 高一帆, 等. 猕猴桃叶片叶绿素含量高光谱估算模型研究[J]. 西北农林科技大学学报(自然科学版), 2020, 48(6): 79-89, 98.
[4]	马春艳, 王艺琳, 翟丽婷, 等. 冬小麦不同叶位叶片的叶绿素含量高光谱估算模型[J]. 农业机械学报, 2022, 53(6): 217-225, 358.
[5]	徐若涵, 杨再强, 申梦吟, 等. 苗期低温胁迫对 “红颜” 草莓叶绿素含量及冠层高光谱的影响[J]. 中国农业气象, 2022, 43(2): 148-158.
[6]	冯海宽, 陶惠林, 赵钰, 等. 利用无人机高光谱估算冬小麦叶绿素含量[J]. 光谱学与光谱分析, 2022, 42(11): 3575-3580.
[7]	李宗飞, 苏继霞, 费聪, 等. 基于高光谱数据的滴灌甜菜叶片全氮含量估算[J]. 作物学报, 2020, 46(4): 557-570.
[8]	GITELSON A A, GRITZ Y, MERZLYAK M N. Relationships between leaf chlorophyll content and spectral reflectance and algorithms for non-destructive chlorophyll assessment in higher plant leaves[J]. Journal of Plant Physiology, 2003, 160(3): 271-282.
[9]	吕晓, 战莘晔, 于海秋, 等. 花生冠层APAR和FAPAR高光谱遥感估测模型研究[J]. 中国油料作物学报, 2020, 42(5): 878-887.
[10]	童庆禧, 张兵, 郑兰芬. 高光谱遥感: 原理、技术与应用[M]. 北京: 高等教育出版社, 2006.
[11]	翁海勇. 基于光学成像技术的柑橘黄龙病快速检测方法研究[D]. 杭州: 浙江大学, 2019.
[12]	刘燕德, 张光伟, 蔡丽君. 基于高光谱的GA和SPA算法对赣南脐橙叶绿素定量分析[J]. 光谱学与光谱分析, 2012, 32(12): 3377.
[13]	尼格拉·吐尔逊, 苏磊·乃比, 高健, 等. GWLS-SVR模型的红枣树叶片叶绿素含量估算[J]. 光谱学与光谱分析, 2021, 41(6): 1730-1736.
[14]	DARVISHZADEH R, SKIDMORE A, SCHLERF M, et al. LAI and chlorophyll estimation for a heterogeneous grassland using hyperspectral measurements[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2008, 63(4): 409-426.
[15]	刘涛, 张寰, 王志业, 等. 利用无人机多光谱估算小麦叶面积指数和叶绿素含量[J]. 农业工程学报, 2021, 37(19): 65-72.
[16]	NAUŠ J, PROKOPOVÁ J, REBÍČEK J, et al. SPAD chlorophyll meter reading can be pronouncedly affected by chloroplast movement[J]. Photosynthesis Research, 2010, 105(3): 265-271.
[17]	付波霖, 邓良超, 张丽, 等. 联合星载高光谱影像和堆栈集成学习回归算法的红树林冠层叶绿素含量遥感反演[J]. 遥感学报, 2022, 26(6): 1182-1205.
[18]	李胜, 张培林, 李兵, 等. 量子GA-PLS特征选择算法及其应用[J]. 量子电子学报, 2014, 31(2): 194.
[19]	JIA W K, ZHAO D A, DING L. An optimized RBF neural network algorithm based on partial least squares and genetic algorithm for classification of small sample[J]. Applied Soft Computing, 2016, 48: 373-384.
[20]	SONG K S, LI L, LI S, et al. Hyperspectral retrieval of phycocyanin in potable water sources using genetic algorithm-partial least squares (GA-PLS) modeling[J]. International Journal of Applied Earth Observation and Geoinformation, 2012, 18: 368-385.
[21]	纪文君. 基于野外vis-NIR高光谱的土壤属性预测及田间水分影响去除研究[D]. 杭州: 浙江大学, 2014.
[22]	彭晓伟, 张爱军, 杨晓楠, 等. 谷子叶绿素含量高光谱特征分析及其反演模型构建[J]. 干旱地区农业研究, 2022, 40(2): 69-77.
[23]	ZHOU W, YANG H, XIE L J, et al. Hyperspectral inversion of soil heavy metals in Three-River Source Region based on random forest model[J]. CATENA, 2021, 202: 105222.
[24]	郭云开, 许敏, 张晓炯, 等. 结合PRO-4SAIL和BP神经网络的叶绿素含量高光谱反演[J]. 测绘通报, 2020(3): 21-24.
[25]	SHAO Y H, LI C N, HUANG L W, et al. Joint sample and feature selection via sparse primal and dual LSSVM[J]. Knowledge-Based Systems, 2019, 185: 104915.
[26]	王国辉, 张宝栋, 李向荣. 基于遗传算法的BP神经网络火炮身管烧蚀磨损量预测[J]. 火炮发射与控制学报, 2019, 40(4): 6-10, 15.
[27]	CHANG C W, LAIRD D A, MAUSBACH M J, et al. Near-infrared reflectance spectroscopy-principal components regression analyses of soil properties[J]. Soil Science Society of America Journal, 2001, 65(2): 480-490.
[28]	罗靖, 杨再强, 杨立, 等. 基于高光谱参数建立苗期高温条件下草莓叶片叶绿素含量估算模型[J]. 中国农业气象, 2022, 43(10): 832-845.
[29]	方慧, 宋海燕, 曹芳, 等. 油菜叶片的光谱特征与叶绿素含量之间的关系研究[J]. 光谱学与光谱分析, 2007, 27(9): 1731-1734.
[30]	蔡建楠, 刘海龙, 姜波, 等. 基于GA-PLS算法的河网水体化学需氧量高光谱反演[J]. 灌溉排水学报, 2020, 39(9): 126-131.
[31]	WANG C, FENG M C, YANG W D, et al. Extraction of sensitive bands for monitoring the winter wheat (Triticum aestivum) growth status and yields based on the spectral reflectance[J]. PLoS One, 2017, 12(1): e0167679.
[32]	张严, 赵忠盖, 刘飞. 黄酒总酚含量检测: 一种基于GA-LSSVM的近红外光谱波段选择方法[J]. 红外技术, 2015, 37(7): 613-617.