基于高光谱的水稻叶片氮含量估计的深度森林模型研究

doi:10.3724/SP.J.1006.2021.02060

摘要/Abstract

摘要：

高光谱遥感已经成为快速诊断作物水氮状态的一种有效手段。然而, 传统的回归方法和机器学习往往难以挖掘高光谱的全部信息, 深度神经网络又通常需要大量的训练数据, 因此本研究试图探索在少量数据条件下构建深度学习模型并实现叶片氮含量的精准估计。通过在湖北省监利县开展了连续2年不同氮素胁迫水平的水稻试验, 测量了作物全生育期内的216组冠层光谱和叶片氮含量。基于一阶导数光谱, 本文构建了一种新的深度学习模型(深度森林DF)来进行叶片氮含量的反演, 并与2种经典机器学习模型(随机森林RF和支持向量机SVM)和一种深度神经网络模型(多层感知器MLP)进行比较。结果表明, 在基于少量高光谱数据的情况下, DF对水稻叶片氮含量的估算精度要高于MLP, 其中预测精度最高的模型为全波段光谱反演的DF模型(R²=0.919, RMSE=0.327)。在2种经典机器学习模型中, RF的估计效果优于SVM, 但2种模型结果都不够稳定。研究表明, 深度森林可以提升高光谱反演叶片氮含量的精度和稳定性, 并且可以通过多粒度扫描相对减轻过拟合程度。该研究结果可为少量数据条件下快速监测作物叶片氮含量提供参考。

关键词: 叶片氮含量, 深度学习, 机器学习, 高光谱遥感, 水稻

Abstract:

Rapid and nondestructive detection of crop nitrogen status is crucial to precision agriculture management. Hyperspectral remote sensing has been proposed to be a powerful tool for expediently monitoring crop nitrogen status. However, conventional regression methods and machine learning (ML) are difficult to utilize the full information of hyperspectral data, and deep neural networks (DNN) usually require a huge number of training data. Therefore, we attempt to construct deep learning models with a small amount of data and achieve accurate estimation of leaf nitrogen concentration (LNC). A two-year field experiment of paddy rice with four nitrogen levels were conducted at Jianli, Hubei province, China. A total of 216 samples containing canopy hyperspectral data and rice LNC were measured during two growing seasons. Based on the first derivative of hyperspectral data, a new deep learning model (deep forest, DF) was constructed for LNC estimation and compared with two traditional machine learning models (random forest, RF and support vector machine, SVM) and one deep neural network model (multi-layer perceptron, MLP). The results showed that, based on a small number of hyperspectral data, deep forest acquired higher accuracy than MLP. And the optimal estimation (R²= 0.919, RMSE = 0.327) was obtained by the deep forest model based on full-wave band spectrum (350-2500 nm). Between two classical machine learning models, random forest achieved better results than SVM, but both methods were unstable. In conclusion, deep forest improved the prediction accuracy and model robustness for all band situations, and alleviated the degree of overfitting by multi-grained scanning. These results can provide a deep insight to detect crop nitrogen status rapidly when confronted with limited data.

Key words: leaf nitrogen concentration, deep learning, machine learning, hyperspectral remote sensing, paddy rice

李金敏, 陈秀青, 杨琦, 史良胜. 基于高光谱的水稻叶片氮含量估计的深度森林模型研究[J]. 作物学报, 2021, 47(7): 1342-1350.

LI Jin-Min, CHEN Xiu-Qing, YANG Qi, SHI Liang-Sheng. Deep learning models for estimation of paddy rice leaf nitrogen concentration based on canopy hyperspectral data[J]. Acta Agronomica Sinica, 2021, 47(7): 1342-1350.

图/表 10

表1

图1

表2

图2

表3

图3

表4

图4

表5

图5

参考文献 36

[1]	Zhang X, Davidson E A, Mauzerall D L, Searchinger T D, Dumas P, Shen Y.Managing nitrogen for sustainable development.Nature, 2015, 528: 51-59. doi: 10.1038/nature15743 pmid: 26595273
[2]	Yu C, Huang X, Chen H, Godfray H C J, Wright J S, Hall J W, Gong P, Ni S, Qiao S, Huang G. Managing nitrogen to restore water quality in China.Nature, 2019, 567: 516-520. pmid: 30818324
[3]	Vigneau N, Ecarnot M, Rabatel G, Roumet P.Potential of field hyperspectral imaging as a non destructive method to assess leaf nitrogen content in wheat.Field Crops Res, 2011, 122: 25-31.
[4]	高林, 杨贵军, 李长春, 冯海宽, 徐波, 王磊, 董锦绘, 付奎. 基于光谱特征与PLSR结合的叶面积指数拟合方法的无人机画幅高光谱遥感应用. 作物学报, 2017, 43: 549-557.
	Gao L, Yang G J, Li C C, Feng H K, Xu B, Wang L, Dong J H, Fu K.Application of an improved method in retrieving leaf area index combined spectral index with PLSR in hyperspectral data generated by unmanned aerial vehicle snapshot camera.Acta Agron Sin, 2017, 43: 549-557 (in Chinese with English abstract).
[5]	吾木提·艾山江, 买买提·沙吾提, 陈水森, 李丹. 基于GF-1/2卫星数据的冬小麦叶面积指数反演. 作物学报, 2020, 46: 787-797.
	Umut H, Mamat S, Chen S S, Li D.Inversion of leaf area index of winter wheat based on GF-1/2 image.Acta Agron Sin, 2020, 46: 787-797 (in Chinese with English abstract).
[6]	Curran P J.Remote sensing of foliar chemistry.Remote Sens Environ, 1989, 30: 271-278.
[7]	Feng W, Yao X, Zhu Y, Tian Y C, Cao W X.Monitoring leaf nitrogen status with hyperspectral reflectance in wheat.Eur J Agron, 2008, 28: 394-404.
[8]	田永超, 杨杰, 姚霞, 曹卫星, 朱艳. 利用叶片高光谱指数预测水稻群体叶层全氮含量. 作物学报, 2010, 36: 1529-1537.
	Tian Y C, Yang J, Yao X, Cao W X, Zhu Y.Monitoring canopy leaf nitrogen concentration based on leaf hyperspectral indices in rice.Acta Agron Sin, 2010, 36: 1529-1537 (in Chinese with English abstract).
[9]	陈兵, 韩焕勇, 王方永, 刘政, 邓福军, 林海, 余渝, 李少昆, 王克如, 肖春华. 利用光谱红边参数监测黄萎病棉叶叶绿素和氮素含量. 作物学报, 2013, 39: 319-329.
	Chen B, Han H Y, Wang F Y, Liu Z, Deng F J, Lin H, Yu Y, Li S K, Wang K R, Xiao C H.Monitoring chlorophyll and nitrogen contents in cotton leaf infected by Verticillium wilt with spectra red edge parameters.Acta Agron Sin, 2013, 39: 319-329 (in Chinese with English abstract).
[10]	吴亚鹏, 贺利, 王洋洋, 刘北城, 王永华, 郭天财, 冯伟. 冬小麦生物量及氮积累量的植被指数动态模型研究. 作物学报, 2019, 45: 1238-1249.
	Wu Y P, He L, Wang Y Y, Liu B C, Wang Y H, Guo T C, Feng W.Dynamic model of vegetation indices for biomass and nitrogen accumulation in winter wheat. Acta Agron Sin, 2019, 45: 1238-1249 (in Chinese with English abstract).
[11]	李宗飞, 苏继霞, 费聪, 李阳阳, 刘宁宁, 戴宇祥, 张开祥, 王开勇, 樊华, 陈兵. 基于高光谱数据的滴灌甜菜叶片全氮含量估算. 作物学报, 2020, 46: 557-570.
	Li Z F, Su J X, Fei C, Li Y Y, Liu N N, Dai Y X, Zhang K X, Wang K Y, Fan H, Chen B.Estimation of total nitrogen content in sugarbeet leaves under drip irrigation based on hyperspectral characteristic parameters and vegetation index.Acta Agron Sin, 2020, 46: 557-570 (in Chinese with English abstract).
[12]	陈秀青, 杨琦, 韩景晔, 林琳, 史良胜. 基于叶冠尺度高光谱的冬小麦叶片含水量估算. 光谱学与光谱分析, 2020, 40: 891-897.
	Chen X Q, Yang Q, Han J Y, Lin L, Shi L S.Estimation of winter wheat leaf water content based on leaf and canopy hyperspectral data.Spectrosc Spect Anal, 2020, 40: 891-897 (in Chinese with English abstract).
[13]	张筱蕾, 刘飞, 聂鹏程, 何勇, 鲍一丹. 高光谱成像技术的油菜叶片氮含量及分布快速检测. 光谱学与光谱分析, 2014, 34: 2513-2518.
	Zhang X L, Liu F, Nie P C, He Y, Bao Y D.Rapid detection of nitrogen content and distribution in oilseed rape leaves based on hyperspectral imaging.Spectrosc Spect Anal, 2014, 34: 2513-2518 (in Chinese with English abstract).
[14]	李旭青, 刘湘南, 刘美玲, 吴伶. 水稻冠层氮素含量光谱反演的随机森林算法及区域应用. 遥感学报, 2014, 18: 923-945.
	Li X Q, Liu X N, Liu M L, Wu L.Random forest algorithm and regional applications of spectral inversion model for estimating canopy nitrogen concentration in rice.J Remote Sens, 2014, 18: 923-945 (in Chinese with English abstract).
[15]	Liang L, Di L, Huang T, Wang J, Lin L, Wang L, Yang M.Estimation of leaf nitrogen content in wheat using new hyperspectral indices and a random forest regression algorithm.Remote Sens-Basel, 2018, 10: 1940.
[16]	依尔夏提·阿不来提, 买买提·沙吾提, 白灯莎·买买提艾力, 安申群, 马春玥. 基于随机森林法的棉花叶片叶绿素含量估算. 作物学报, 2019, 45: 81-90.
	Ershat A, Mamat S, Baidengsha M, An S Q, Ma C Y.Estimation of leaf chlorophyll content in cotton based on the random forest approach.Acta Agron Sin, 2019, 45: 81-90 (in Chinese with English abstract).
[17]	LeCun Y, Bengio Y, Hinton G. Deep learning.Nature, 2015, 521: 436-444. doi: 10.1038/nature14539 pmid: 26017442
[18]	Mohanty S P, Hughes D P, Salathé M.Using deep learning for image-based plant disease detection.Front Plant Sci, 2016, 7: 1419. doi: 10.3389/fpls.2016.01419 pmid: 27713752
[19]	Yang Q, Shi L, Han J, Zha Y, Zhu P.Deep convolutional neural networks for rice grain yield estimation at the ripening stage using UAV-based remotely sensed images.Field Crops Res, 2019, 235: 142-153.
[20]	Zhong L, Hu L, Zhou H.Deep learning based multi-temporal crop classification.Remote Sens Environ, 2019, 221: 430-443.
[21]	Sun J, Yang J, Shi S, Chen B, Du L, Gong W, Song S.Estimating rice leaf nitrogen concentration: influence of regression algorithms based on passive and active leaf reflectance.Remote Sens-Basel, 2017, 9: 951.
[22]	Zhou Z, Feng J.Deep forest.Nat Sci Rev, 2019, 6: 74-86.
[23]	Liu X, Wang R, Cai Z, Cai Y, Yin X.Deep multigrained cascade forest for hyperspectral image classification. IEEE Trans Geosci Remote Sens, 2019, 57: 8169-8183.
[24]	Zhang J, Song H, Zhou B.SAR target classification based on deep forest model.Remote Sens-Basel, 2020, 12: 128.
[25]	Zhang L, Sun H, Rao Z, Ji H.Hyperspectral imaging technology combined with deep forest model to identify frost-damaged rice seeds.Spectrochimica Acta Part A: Mol Biomol Spect, 2020, 229: 117973.
[26]	Yu X, Lu H, Liu Q.Deep-learning-based regression model and hyperspectral imaging for rapid detection of nitrogen concentration in oilseed rape (Brassica napus L.) leaf. Chemometr Intell Lab, 2018, 172: 188-193.
[27]	Breiman L.Random forests.Mach Learn, 2001, 45: 5-32.
[28]	Berger K, Verrelst J, Féret J, Wang Z, Wocher M, Strathmann M, Danner M, Mauser W, Hank T.Crop nitrogen monitoring: recent progress and principal developments in the context of imaging spectroscopy missions.Remote Sens Environ, 2020, 242: 111758.
[29]	Hu G, Peng X, Yang Y, Hospedales T M, Verbeek J.Frankenstein: learning deep face representations using small data.IEEE Trans Image Proc, 2017, 27: 293-303.
[30]	Koppe G, Meyer-Lindenberg A, Durstewitz D.Deep learning for small and big data in psychiatry. Neuropsychopharmacol, 2020, 46: 176-190.
[31]	Yi Q, Huang J, Wang F, Wang X, Liu Z.Monitoring rice nitrogen status using hyperspectral reflectance and artificial neural network.Environ Sci Technol, 2007, 41: 6770-6775.
[32]	Tan K, Wang S, Song Y, Liu Y, Gong Z.Estimating nitrogen status of rice canopy using hyperspectral reflectance combined with BPSO-SVR in cold region. Chemometr Intell Lab, 2018, 172: 68-79.
[33]	Fu Y, Yang G, Li Z, Li H, Li Z, Xu X, Song X, Zhang Y, Duan D, Zhao C.Progress of hyperspectral data processing and modelling for cereal crop nitrogen monitoring.Comput Electron Agric, 2020, 172: 105321.
[34]	Yao X, Huang Y, Shang G, Zhou C, Cheng T, Tian Y, Cao W, Zhu Y.Evaluation of six algorithms to monitor wheat leaf nitrogen concentration.Remote Sens-Basel, 2015, 7: 14939-14966.
[35]	Kokaly R F.Investigating a physical basis for spectroscopic estimates of leaf nitrogen concentration. Remote Sens Environ, 2001, 75: 153-161.
[36]	Wang J, Chen Y, Chen F, Shi T, Wu G.Wavelet-based coupling of leaf and canopy reflectance spectra to improve the estimation accuracy of foliar nitrogen concentration. Agric For Meteorol, 2018, 248: 306-315.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

2018年试验Experiment in 2018			2019年试验Experiment in 2019
小区编号 Plot number	氮素水平 Nitrogen level	小区面积 Area (m²)	小区编号 Plot number	氮素水平 Nitrogen level	小区面积 Area (m²)
B1	N3	353.8	B1	N3′	296.70
B2	N2	352.2	B2	N2′	293.09
B3	N0	347.2	B3	N1′	317.83
B4	N1	338.3	B4	N0	333.10
B5	N0	338.3	B5	N3′	329.80
B6	N3	337.5	B6	N2′	383.78
B7	N1	340.5	B7	N1′	306.47
B8	N2	342.6	B8	N0	295.22
B9	N3	341.3	B9	N3′	313.16
B10	N1	340.4	B10	N2′	291.99
B11	N2	327.6	B11	N1′	299.94
B12	N0	201.8	B12	N0	292.74

年份 Year	样本数量 Sample size	平均值 Mean (g g^-1)	最大值 Max. (g g^-1)	最小值 Min. (g g^-1)
2018	120	3.04	4.90	0.70
2019	96	3.38	5.37	1.63

随机森林 Random forest		支持向量机 SVM		多层感知器 Multi-layer perceptron		深度森林 Deep forest
树的数量 Number of trees	100	正则化参数 Penalty coefficient	50	激活函数 Activation function	“relu”	级联结构中森林数量 Forest in cascade	10
决策树最大深度 Maximum depth of decision tree	“None”	核函数 Kernel function	“RBF”	隐藏层神经元数量 Number of hidden units	[100, 50, 20]	滑动窗口大小 Sliding window size	{[d/16], [d/8], [d/4]}
		核函数系数 Kernel coefficient	100	优化器 Optimizer	“adam”	窗口滑动步长 Sliding step	25
				损失函数 Loss function	“mse”
				迭代次数 Iterations	500

回归模型 Regression model	光谱波段 Band	范围 Range (nm)	建模集 Calibration set		预测集 Prediction set
回归模型 Regression model	光谱波段 Band	范围 Range (nm)	R²_c	RMSE_c	R²_p	RMSE_p
随机森林 Random forest	全波段Full-wave band	350-2500	0.982	0.149	0.891	0.378
	可见光波段Visible waveband	400-760	0.976	0.171	0.804	0.507
	近红外波段Near-infrared waveband	760-2500	0.980	0.156	0.890	0.380
支持向量机 SVM	全波段Full-wave band	350-2500	0.960	0.223	0.725	0.601
	可见光波段Visible waveband	400-760	0.843	0.440	0.755	0.568
	近红外波段Near-infrared waveband	760-2500	0.936	0.281	0.697	0.631

回归模型 Regression model	光谱波段 Band	范围 Range (nm)	建模集 Calibration set		预测集 Prediction set
回归模型 Regression model	光谱波段 Band	范围 Range (nm)	R²_c	RMSE_c	R²_p	RMSE_p
多层感知器 Multi-layer perceptron	全波段Full-wave band	350-2500	0.937	0.282	0.872	0.392
	可见光波段Visible waveband	400-760	0.876	0.397	0.858	0.412
	近红外波段Near-infrared waveband	760-2500	0.932	0.293	0.860	0.408
深度森林 Deep forest	全波段Full-wave band	350-2500	0.981	0.151	0.919	0.327
	可见光波段Visible waveband	400-760	0.978	0.165	0.903	0.358
	近红外波段Near-infrared waveband	760-2500	0.979	0.162	0.917	0.330