APP下载

黑土养分含量的航空高光谱遥感预测

2019-12-19杨越超赵英俊赵宁博张东辉

农业工程学报 2019年20期
关键词:黑土全氮反演

杨越超,赵英俊,秦 凯,赵宁博,杨 晨,张东辉,崔 鑫

黑土养分含量的航空高光谱遥感预测

杨越超1,赵英俊1,秦 凯1,赵宁博1,杨 晨2,张东辉1,崔 鑫1

(1. 核工业北京地质研究院遥感信息与图像分析技术国家级重点实验室,北京 100029;2. 武汉大学城市设计学院,武汉 430072)

为监测黑龙江省黑土典型区土壤的养分元素含量,综合利用统计理论与光谱分析方法,研究建三江农场黑土土壤的3类养分含量与土壤光谱之间的关系,建立土壤全氮、有效磷、速效钾含量高光谱反演模型,实现土壤养分元素含量定量预测。对黑土土壤航空高光谱数据进行处理,应用偏最小二乘回归(PLSR)和BP神经网络方法分别建立土壤养分元素含量的高光谱定量反演模型,结果表明:全氮PLSR和BP神经网络预测模型的RPIQ值(样本观测值第三和第一四分位数之差与均方根误差的比值)分别为2.42和2.80;有效磷PLSR和BP神经网络模预测型的RPIQ值分别为0.83和1.67;速效钾PLSR和BP神经网络模型的RPIQ值分别为2.00和2.33。试验证明土壤全氮和速效钾的光谱定量预测模型具备较好的精度和预测能力。但有效磷的预测效果不是特别理想,仅可达到近似定量预测的要求;全氮、有效磷和速效钾的预测精度,BP神经网络建模相比偏最小二乘建模有更好的精度和预测能力,预测精度分别提高6.5%、10.1%和6.6%。

土壤;遥感;模型;偏最小二乘法;BP神经网络

0 引 言

土壤是植物生长养分的主要来源,尤其是土壤有机质、氮、磷、钾元素对植物生长具有重要的作用[1]。植物需要大量的氮素合成蛋白质;磷能促进植物根系的形成和生长,钾能够促进光合作用。土壤中主要养分(全氮、有效磷和速效钾)的含量是重要的农作物产量影响指标,是指导农业科学施肥的重要依据[2]。中国东北地区发育有全球非常重要的黑土地资源。黑土因土壤性状好、肥力高,非常适合粮食作物生长。快速准确获取黑土地土壤主要养分的含量,已然成为东北黑土区精准农业发展的必然需要[3]。

目前测量3类土壤养分主要采用实验室化学方法,利用某些试剂溶液提取土壤中养分相对值加以测定[4],传统方法工作量大、周期长,难以满足现代农业快速发展的需要。随着GIS及遥感技术的发展,多光谱影像解译也在农业信息监测中得到了一定程度的应用,在具备现势性强特点的同时,多光谱技术受制于谱段间隔较宽及环境干扰值的影响,一定程度上反演精度受限[5-6]。而通过高光谱技术反演土壤养分对于土壤信息快速测定具有重大意义[7]。国内外学者应用高光谱针对土壤矿物成分、水分及有机质等开展了一些定量研究,350~2 500 nm波段高光谱数据能映射一些土壤理化参数的微小差别,水分、有机质及铁氧化物的含量与土壤反射率存在一定明显的对应关系[8-10],可建立定量反演的预测模型[11]。综合来看,氮、磷、钾的高光谱分析预测研究相对较少,土壤中各类养分元素与光谱也存在较复杂的对应关系[10-11]。以往研究多数利用ASD FieldSpecPro地物光谱仪在室内或野外采集点状数据研究光谱养分对应关系并建立估测模型[12-14],对于大面积土地光谱数据测量效率低,同时模型建立有较大的随机性,不足以平衡局部和全局最优的问题,还需进一步挖掘土壤光谱信息[15-16]。

为提高黑土土壤养分信息定量预测的效率与精度,笔者将基于建三江地区航空高光谱遥感数据,在分析研究土壤光谱特征基础上,利用偏最小二乘回归和BP神经网络分别建立黑土地土壤3类养分(全氮、有效磷和速效钾)含量高光谱反演模型,探索快速测定黑土土壤养分的方法。

1 数据的获取与处理

1.1 研究区概况

研究区位于黑龙江省佳木斯市建三江管理局七星农场(见图1)。地处47°01′~47°10′ N,132°43′~133°02′ E,面积约380 km²;位于黑龙江、松花江和乌苏里江交汇河间地带,水资源丰富。区内分布着黑钙土、黑土、沼泽土、草甸土和水稻土等。土壤成土母质主要为黄土状粉质黏土、淤泥质粉质黏土[17]。隶属中温带大陆性季风气候。平均海拔50 m,耕地集中成片,地势平坦,适宜现代农业规模化经营。

图1 研究区地理位置及采样点示意图

1.2 航空高光谱数据采集及处理

野外航空高光谱数据采集使用CASI-1500和SASI-600线阵推扫型成像光谱仪器,空间分辨率分别为1.5和3.75 m,总视场角40°,每行像元数1470,绝对辐射精度小于<2%。波段范围分别为380~1 058 nm和950~2 450 nm,波段数分别为72和100,光谱分辨率分别为9.3和15 nm[18]。地面铺设黑白布,采用ASD FieldSpecPro光谱仪获取定标光谱,其光谱范围为350~2 500 nm,光谱分辨率为1 nm。

将航空高光谱原始辐射数据进行定标、大气辐射校正,利用POS 510系统进行几何校正。经过光谱去噪、重采样、归一化和包络线去除等预处理,获得地表反射率数据。进一步对光谱应用Savitzky-Golay方法选取3个像元为窗口进行平滑滤波,并进行一阶微分、对数变换和去连续统处理,突出分离光谱变化趋势和光谱吸收谷。

1.3 土壤样品采集

野外土壤样品采样深度5~15 cm,选取耕地地块中心,土壤裸露区域,清除表层杂草、砾石等杂质。为增加样本代表性,采样时以采样点为中心原点,周围15 m范围内多点采集3~5个子样进行组合,混合后留取1.5 kg,共采集96组。经过风干、拌匀、研磨后,过200目筛后用于实验室测试。元素含量采用NaOH扩散法(N)、NaHCO3浸提-钼蓝比色法(P2O5)和NH4OAC浸提-火焰光度法(K2O)测定,参考Kennard-Stone法选取72组代表性样品作为养分元素预测的建模样品,24组为模型预测样品[19]。其各元素统计特征描述见表1。

表1 土壤样品3类养分含量信息

1.4 不同含量黑土养分光谱特征分析

将96组样本按养分含量大小排序,对比在可见光-近红外波段范围内光谱变化规律[20]。

1)每个养分含量区间取2条光谱进行分析,得出全氮变化规律是随含量增高,反射率逐渐降低(图2a)。其中3号样品全氮质量分数为4.56g/kg,反射率显著低于其他样品。而22号和68号样品全氮含量在0.60g/kg左右,其反射率相对高于总体光谱均值。变化规律与有机质光谱曲线类似[21]。但当全氮含量较低时受土壤含水量及混合像元干扰,此规律会逐渐减弱至不显著。(2)有效磷含量在此波段范围内无显著规律(图2b)。黑土中有效磷含量相对较低,在光谱曲线上特征不明显。(3)速效钾在此波段范围内无显著规律(图2c)。黑土中速效钾含量相对较低,在光谱曲线上特征不明显。

图2 不同养分含量黑土光谱特征

1.5 相关性分析

针对3类养分元素进行相关性分析(表2),各光谱变换的相关性不同,其中显著相关性出现在一阶微分光谱变换中[22-23]。挑选其中5个较为代表性波段列出,如580 nm一阶微分光谱与TN含量呈显著相关,相关系数为−0.43;与P2O5含量相关系数为−0.36。1 730~2 200 nm一阶微分光谱与K2O呈显著相关,相关系数最大为−0.31。以TN为例,对比原始光谱波形,一阶微分与三种养分含量间的相关系数波动变化、正负交差相对剧烈,峰值系数点较多[24](图3)。对数一阶微分变换与包络线去除变换光谱与养分元素含量相关性相对不高。因此选取一阶微分变换光谱中于养分相关性较高的波段(N:456~600,809~856,1 025~1 190,1 355~1 415,1 685~1 805,2 195~2 285 nm;P2O5:447~495,562~580,819~886,1 085~1 145,1 715~1 790,1 910~1 955,2 195~2 300 nm;K2O:467~485,542~571,886~933,1 250~1 295,1 355~1 430,1 685~1 805,1 920~2 360 nm)应用于研究,波段数共计为86个。

表2 土壤TN、P2O5、K2O含量与部分波段的相关系数

注:*在0.05水平(双侧)上显著相关。

Note: Significant correlation at *0.05 level (bilateral).

图3 TN含量与不同变换形式的光谱相关系数图

2 反演与验证方法

2.1 偏最小二乘回归

偏最小二乘(PLSR)是一种多对多回归建模的算法[24]。建模流程中融合了主成分分析、典型相关性分析和线性回归的方法优点,同时克服主成分分析对自变量解释较强,因变量解释不够的缺点。本次研究应用偏最小二乘回归模型,以土壤养分含量为因变量针对光谱特征波段多自变量进行回归。

2.2 BP神经网络

BP神经网络较为适用于预测、分类及评价等方面。由输入层、隐含层、输出层构成,采用误差反向传播算法进行学习,逐层传播数据,连接权值逐层向前修正,层层之间全部互相连接,同层单元之间不存在相互连接,每一层神经元只针对下一层神经元有影响。若输出层未能达到期望输出,便转入误差逆向传播阶段,依据误差信号修改每个单元权值。学习过程将持续到误差减小到可接受范围或预定训练次数为止。为防止学习速度过快或过拟合造成的模型误差,BP神经网络建模的过程分为训练建模和测试校正两个步骤,达到一定测试精度即可确定为模型[25-29]。

2.3 模型验证

反演模型精度验证由模型稳定性和预测能力决定[30-31]。决定系数(2)、均方根误差(RMSE)和RPIQ值分别衡量模型的稳定性和精度。建模集决定系数2 c越大,均方根RMSEC误差越小,说明模型越稳定,精度越好。预测集决定系数2 p越大,均方根误差RMSEP越小,说明预测效果越好。RPIQ(样本观测值第三四分位数Q3和第一四分位数Q1的差IQ与RMSE的比值)对于非正态分布土壤数据的光谱预测模型精度评价更为客观,其值越大,说明预测效果越好。

3 结果与分析

3.1 偏最小二乘回归模型

应用Unsramble 9.7建立最小二乘回归模型,将建模集样品进行土壤TN、P2O5和K2O含量预测建模。建模中变量投影重要性指标VIPj值所指示变量集合与相关性较高的波段对应,证明其对应波段在解释因变量集合即养分元素时具有重要作用[32-35]。建模集TN和K2O的模型决定系数2 c分别为0.891和0.816,RMSEC为0.23 g/kg和0.06 g/kg均小于样本平均值的10%,预测集决定系数2 p对比建模集也较为稳定,分别为0.851 2和0.808 6,RMSEP分别为0.29 g/kg和0.07 g/kg,RPIQ值分别为2.42和2.00,模型具备较好的精度和预测能力。P2O5的模型决定系数2 c=0.693,RMSEC为0.03 g/kg,预测集决定系数2 p=0.707 5,RMSEP为0.06 g/kg,RPIQ值为0.83,表明P2O5的预测效果不是特别理想,仅可达到近似定量预测的精度要求。三类养分的回归系数与回归方程均能通过显著性检验(<0.01),回归方程如下:

(TN)=31.57723−46.55943+15.43950+11.2011730−

34.5602 105+25.6302 120−48.0702 180+

86.822 195−40.672 210+2.879 1 (1)

(P2O5)=31.55950−47.59965−6.613980+19.6995−

37.611 295+43.441 310+45.282 090−

10.272 105+63.752 120−25.162 135+

65.022 195−50.072 210+5.4512 225−

3.5242 435+1.6812 450+0.949 6 (2)

(K2O)=0.764933+0.865943+0.898950−1.0051100−

1.0481 115−1.0131 130+0.5231 355+0.6921 430+

0.6821 445+2.0861 760+0.9912 015−2.3592 210−

2.522 375+1.7242 435−0.4072 450+2.49 (3)

运用PLSR模型对黑土土壤样本进行养分含量预测,3类养分的实测与预测值散点拟合对比结果见图4。TN预测值范围为1.35~3.45 g/kg,平均值为2.37 g/kg,标准差为0.03 g/kg。P2O5预测值范围为0.13~0.27 g/kg,平均值为0.18 g/kg,标准差为0.04 g/kg;K2O预测值范围为2.34~2.56 g/kg,平均值为2.45 g/kg,标准差为0.05 g/kg。

图4 黑土养分样本实测值与PLSR预测值对比图

3.2 BP神经网络拟合

利用MATLAB编程实现神经网络的设计、训练及仿真函数实现BP神经网络建立模型,采用三层BP网络,将相关性较高的特征波段提取的8个主成分分量作为神经网络的训练输入节点,其主成分累计方差贡献率达99.96%。隐含层为tansig传递函数,节点数经测试为5。输出层采用purelin传递函数,输出节点分别为三类土壤养分含量。训练函数为trainlm,训练次数为1 000次,期望误差为0.000 1。以全氮为例,其BP神经网络训练的误差性能变化及数据训练回归情况如图5,经过训练的网络误差为0.001 279 3,相关系数达到0.998,模型拟合程度较高。

BP神经网络拟合的TN预测模型决定系数2 p= 0.906 5,P2O5预测模型决定系数2 p=0.7786,K2O预测模型决定系数2 p=0.862 2。RMSEP分别为0.25、0.03和0.06 g/kg,RPD值分别为2.39、1.34和2.49。模型具备较好的精度和预测能力。三类黑土土壤养分的实测与预测值散点拟合对比结果见图6。TN预测值范围为1.33~3.65 g/kg,平均值为2.29 g/kg,标准差为0.53 g/kg;P2O5预测值范围为0.12~0.28 g/kg,平均值为0.18 g/kg,标准差为0.04 g/kg;K2O预测值范围为2.37~2.67 g/kg,平均值为2.47 g/kg;标准差为0.07 g/kg。

图5 全氮BP神经网络训练情况

图6 黑土养分样本实测值与BP神经网络预测值对比图

3.3 结果对比

针对黑土土壤的可见光-近红外航空高光谱数据,将全氮、有效磷和速效钾3类土壤养分分别应用偏最小二乘和BP神经网络建模预测,模型精度对比见表3。结果表明,在全氮定量预测方面,偏最小二乘法与BP神经网络均展现了较高的拟合精度,BP神经网络有较高的2 p和较小的相对误差值,两种方法均可用于全氮定量预测,但BP神经网络有着更高的精度,2值提高了0.053 3,预测平均相对误差提高了1.76%,RPIQ提高至2.80。在有效磷定量预测方面,偏最小二乘法拟合精度较低,BP神经网络相比偏最小二乘法2 p提高了0.071 1,预测平均相对误差提高了1.61%,RPIQ提高至1.67。速效钾的定量预测中BP神经网络相比偏最小二乘法2 p提高了 0.053 6,预测平均相对误差提高了0.26%,RPIQ提高至2.33。在实测与预测值对比情况中,全氮、有效磷和速效钾的定量预测中BP神经网络相比偏最小二乘法具备更高的精度,2 p分别提高6.5%、10.1%和6.6%。将其应用到3类养分的定量预测,得到黑土养分含量的空间预测分布情况(图7)。

表3 预测模型精度对比

图7 黑土3类养分含量航空高光谱定量提取图

4 结 论

航空高光谱遥感为土壤养分元素含量预测提供了一种高效的数据获取手段,面状全区光谱测量相对点状测量在养分元素含量预测上避免了插值方法带来的二次误差,反演效果得到提高。将偏最小二乘法及BP神经网络模型应用于航空高光谱黑土养分信息提取,结果表明:1)全氮含量的光谱特征较为明显,因此两种方法模型预测精度均较高。2)BP神经网络比偏最小二乘法建模的预测效果更佳,黑土土壤光谱反射率与土壤养分含量之间,受其他物质因素影响存在一定的非线性关系,采用BP神经网络回归建模能较好的处理这种关系,可以更好地实现对土壤全氮和速效钾的含量预测,预测精度分别提高6.5%和6.6%。3)两种方法的有效磷的预测效果不是特别理想,其含量与光谱特征走势规律不明显,含量标准差也较低仅为0.04 g/kg,导致较难得到较高精度的回归模型,仅可达到近似定量预测的要求。

[1] 史舟. 土壤地面高光谱遥感原理与方法[M]. 北京:科学出版社,2014.

[2] Bendor E, Banin A. Near-infrared analysis as a rapid method to simultaneously evaluate several soil properties[J]. Soil Science Society of America Journal, 1995, 59(2): 364-372.

[3] Bendor E, Chabrillat S, Demattê J A M, et al. Using imaging spectroscopy to study soil properties[J]. Remote Sensing of Environment, 2009, 113(1): S38-S55.

[4] 陈颂超,彭杰,纪文君,等. 水稻土可见-近红外-中红外光谱特性与有机质预测研究[J].光谱学与光谱分析,2016,36(6):1712-1716. Chen Songchao, Peng Jie, Ji Wenjun, et al. Study on the characteristics and organic matter prediction of rice soil visible-near infrared - mid-Infrared spectroscopy[J]. Spectroscopy and Spectral Analysis, 2016, 36(6): 1712-1716. (in Chinese with English abstract)

[5] 王锐,蔡朕. 基于多光谱遥感的耕地土壤有机质定量反演[J]. 农业工程,2018,8(11):85-89. Wang Rui, Cai zhen. Quantitative inversion of cultivated soil organic matter based on multispectral remote sensing[J]. Agricultural Engineering, 2018, 8(11): 85-89. (in Chinese with English abstract)

[6] 夏楠,塔西甫拉提·特依拜,丁建丽,等. 基于多光谱数据的荒漠矿区土壤有机质估算模型[J]. 农业工程学报,2016,32(6):263-267. Xia Nan, Taxipulati Teyibai, Ding Jianli, et al. Estimation model of soil organic matter in desert mining area based on multi-spectral data [J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2016, 32(6): 263-267. (in Chinese with English abstract)

[7] Daniel Žížala, Tereza Zádorová, Jiří Kapička. Assessment of soil degradation by erosion based on analysis of soil properties using aerial hyperspectral images and ancillary data[J]. Remote Sense, 2017, 9(1): 28-40.

[8] 何挺,王静,林宗坚,等. 土壤有机质光谱特征研究[J]. 武汉大学学报:信息科学版,2006,31(11):975-979. He Ting, Wang Jing, Lin Zongjian, et al. Spectral features of soil organic matter[J]. Geomatics and Information Science of Wuhan University, 2006, 31(11): 975-979. (in Chinese with English abstract)

[9] 刘焕军,潘越,窦欣,等. 黑土区田块尺度土壤有机质含量遥感反演模型[J]. 农业工程学报,2018,34(1):127-133. Liu Huanjun, Pan Yue, Dou Xin, et al. Soil organic matter content inversion model with remote sensing image in field scale of black soil area[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2018, 34(1): 127-133. (in Chinese with English abstract)

[10] 李焱,王让会,管延龙,等. 基于高光谱反射特性的土壤全氮含量预测分析[J]. 遥感技术与应用,2017,32(1):173-179. Li Yan, Wang Ranghui, Guan Yanlong, et al. Prediction of total nitrogen content in soil based on high spectral reflectance[J]. Remote Sensing Technology and Application, 2017, 32(1): 173-179. (in Chinese with English abstract)

[11] 祁亚琴,吕新,邵玉林,等. 基于高光谱数据提取土壤养分信息的研究进展[J]. 中国农学通报,2014,30(12):28-31. Qi Yaqin, Lü Xin, Shao Yulin, et al. Research progress of soil nutrient information extraction based on hyperspectral data[J]. Chinese Agricultural Science Bulletin, 2014, 30(12): 28-31. (in Chinese with English abstract)

[12] 邱壑,陈瀚阅,邢世和,等.基于Hyperion数据的耕地土壤有机质含量遥感反演[J]. 福建农林大学学报:自然版,2017,46(4):460-467. Qiu He, Chen Hanyue, Xing Shihe, et al. Soil organic matter estimation models based on hyperion data[J]. Journal of Fujian Agriculture and Forestry University: Natural Science, 2017, 46(4): 460-467. (in Chinese with English abstract)

[13] 李雪莹,范萍萍,侯广利,等,可见-近红外光谱的土壤养分快速检测[J]. 光谱学与光谱分析,2017,37(11):3562-3566. Li Xueying, Fan Pingping, Hou Guangli, et al. Visible–near infrared spectrum of soil nutrient rapid detection[J]. Spectroscopy and Spectral Analysis, 2017, 37(11): 3562-3566. (in Chinese with English abstract)

[14] 史舟,梁宗正,杨媛媛,等,农业遥感研究现状与展望[J]. 农业机械学报,2015,46(2):247-260. Shi Zhou, Liang Zongzheng, Yang Yuanyuan, et al. Current situation and prospect of agricultural remotesensing research[J]. Transactions of The Chinese Society of Agricultural Machinery, 2015, 46(2): 247-260. (in Chinese with English abstract)

[15] 周鼎浩,薛利红,李颖,等. 基于可见–近红外光谱的水稻土全磷反演研究[J]. 土壤,2014,46(1):47-52. Zhou Dinghao, Xue Lihong, Li Ying, et al. Visible–near infrared reflectance spectroscopy for prediction of total phosphorus content in paddy soil[J]. Soil, 2014, 46(1): 47-52. (in Chinese with English abstract)

[16] 王人潮,苏海萍,王深法. 浙江省主要土壤光谱反射特性及其模糊分类在土壤分类中的应用研究[J]. 浙江大学学报:农业与生命科学版,1986,12(4):464-471. Wang Renchao, Su Haiping, Wang Shenfa. Spectral reflectance characteristics of main soils in zhejiang province and Its fuzzy classification applied to soil classification[J]. Journal of Zhejiang University (Agriculture and Life Sciences), 1986, 12(4): 464-471. (in Chinese with English abstract)

[17] 吴嵩. 典型黑土区土壤有机质含量反演研究[D]. 长春:吉林大学,2016. Wu Song. Research of Soil Organic Matter Content Inversion in Typical Black Soil Area[D]. Changchun:Jilin university, 2016.

[18] 叶发旺,刘德长,赵英俊. CASI/SASI航空高光谱遥感测量系统及其在铀矿勘查中的初步应用[J]. 世界核地质科学,2011,28(4):231-236. Ye Fawang, Liu Dechang, Zhao Yingjun. Airborne hyper-spectral survey system CASI/SASI and its preliminary application in uranium exploration[J]. World Nuclear Geoscience, 2011, 28(4): 231-236. (in Chinese with English abstract)

[19] 张瑶,李民赞,郑立华,等. 基于近红外光谱分析的土壤分层氮素含量预测[J].农业工程学报,2015,31(9):121-126. Zhang Yao, Li Minzan, Zheng Lihua, et al. Prediction of nitrogen content in soil based on near infrared spectrum analysis[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2015, 31(9): 121-126. (in Chinese with English abstract)

[20] 张俊华,马天成,贾科利. 典型龟裂碱土土壤光谱特征影响因素研究[J]. 农业工程学报,2014,30(23):158-165. Zhang Junhua, Ma Tiancheng, Jia Keli. Factors affecting spectral characteristics of typical takyr solonetzs[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2014, 30(23): 158-165. (in Chinese with English abstract)

[21] 刘焕军,张柏,赵军,等. 黑土有机质含量高光谱模型研究[J]. 土壤学报,2007,44(1):27-32. Liu Huanjun, Zhang Bai, Zhao Jun, et al. Spectral models for prediction of organic matter in black soil[J]. Acta Pedologica Sinica, 2007, 44(1): 27-32. (in Chinese with English abstract)

[22] Nour-Omid B, Parlett B N, Ericsson T, et al. How to implement the spectral transformation[J]. Mathematics of Computation, 1987, 48(178): 663-663.

[23] Du P J, Chen Y H, Fang T, et al. Study on the extraction and applications of spectral features in hyperspectral remote sensing[J]. Journal of China University of Mining & Technology, 2003, 32(5): 500-504.

[24] 王昶,黄驰超,徐光辉,等. 近红外光谱结合偏最小二乘法快速评估土壤质量[J]. 土壤学报,2013,50(5):36-45. Wang Chuang, Huang Chichao, Xu Guanghui, et al. Rapid evaluation of soil quality through a near infrared-partial least squares (NIR-PLS) method[J]. Acta Pedologica Sinica, 2013, 50(5): 36-45. (in Chinese with English abstract)

[25] 李硕,汪善勤,张美琴,等. 基于可见长丘红外光谱比较主成分回归、偏最小二乘回归和反向传播神经网络对土壤氮的预测研究[J]. 光学学报,2012,32(8):0830001-0830005. Li Shuo, Wang Shanqin, Zhang Meiqin, et al. Comparison among principal component regression,partial least squares regression and back propagation neural network for prediction of soil nitrogen with visible-near infrared spectroscopy[J]. Acta Optica Sinica, 2012, 32(8): 0830001-0830005. (in Chinese with English abstract)

[26] Zhang P, Li Y. Study on the comparisons of the establishment of two mathematical modeling methods for soil organic matter content based on spectral reflectance[J]. Spectroscopy and Spectral Analysis, 2016, 36(3): 903-910.

[27] Doustfatemeh I, Baleghi Y. Comprehensive urban area extraction from multispectral medium spatial resolution remote-sensing imagery based on a novel structural feature[J]. International Journal of Remote Sensing, 2016, 37(18): 4225-4242.

[28] Andreas Steinberg, Sabine Chabrillat, Antoine Stevens, et al. Prediction of common surface soil properties based on Vis-NIR airborne and simulated EnMAP imaging spectroscopy data: Prediction Accuracy and Influence of Spatial Resolution[J]. Remote Sense, 2016, 8(7): 613-627.

[29] 郑立华,李民赞,潘娈,等. 基于近红外光谱技术的土壤参数BP神经网络预测[J]. 光谱学与光谱分析,2008(5):1160-1164. Zheng Lihua, Li Minzan, Pan Luan, et al. Prediction of soil parameters BP neural network based on near-infrared spectroscopy[J]. Spectroscopy and Spectral Analysis, 2008(5): 1160-1164 (in Chinese with English abstract)

[30] 薛利红,周鼎号,李颖,等. 不同利用方式下土壤有机质和全磷的可见近红外高光谱反演[J]. 土壤学报,2014,51(5):993-1001. Xue Lihong, Zhou Dinghao, Li Ying, et al. Prediction of soil organic matter and total phosphorus with vis-nir hyperspectral inversion relative to land use[J]. Acta Pedologica Sinica, 2014, 51(5): 993-1001. (in Chinese with English abstract)

[31] 徐永明,蔺启忠,王璐,等. 基于高分辨率反射光谱的土壤营养元素估算模型[J]. 土壤学报,2006,43(5):709-716. Xu Yongming, Lin Qizhong, Wang Lu, et al. Model for estimating soil nutrient elements based on high resolution reflectance spectra[J]. Acta Pedologica Sinica, 2006, 43(5): 709-716. (in Chinese with English abstract)

[32] 高灯州,曾从盛,章文龙,等. 闽江口湿地土壤全氮含量的高光谱遥感估算[J]. 生态学杂志,2016,35(4):952-959. Gao Dengzhou, Zeng Congsheng, Zhang Wenlong, et al. Estimating of soil total nitrogen concentration based on hyperspectral remote sensing data in Minjiang River estuarine wetland[J]. Chinese Journal of Ecology, 2016, 35(4): 952-959. (in Chinese with English abstract)

[33] 张东辉,赵英俊,秦凯. 一种新的光谱参量预测黑土养分含量模型[J]. 光谱学与光谱分析,2018,38(9):1-5. Zhang Donghui, Zhao Yingjun, Qin Kai. A new model for predicting black soil nutrient content by spectral parameters[J]. Spectroscopy and Spectral Analysis, 2018, 38(9): 1-5. (in Chinese with English abstract)

[34] 李晨,张国伟,周治国,等. 滨海盐土土壤水分的高光谱参数及估测模型[J]. 应用生态学报,2016,27(2):525-531. Li Chen,Zhang Guowei,Zhou Zhiguo,et al. Hyperspectral parameters and prediction model of soil moisture in coastal saline[J]. Chinese Journal of Applied Ecology 2016, 27(2): 525-531. (in Chinese with English abstract)

[35] 程先锋,宋婷婷,陈玉,等. 滇西兰坪铅锌矿区土壤重金属含量的高光谱反演分析[J]. 岩石矿物学杂志,2017,36(1):60-69. Cheng Xianfeng, Song Tingting, Chen Yu, et al. Retrieval and analysis of heavy metal content in soil based on measured spectra in the Lanping Zn-Pb mining area, western Yunnan Province[J]. Acta Petrologica ET Mineralogica, 2017, 36(1): 60-69. (in Chinese with English abstract)

Prediction of black soil nutrient content based on airborne hyperspectral remote sensing

Yang Yuechao1, Zhao Yingjun1, Qin Kai1, Zhao Ningbo1, Yang Chen2, Zhang Donghui1, Cui Xin1

(1.,,100029,; 2.,,430072,)

In order to improve the efficiency and accuracy of the quantitative prediction of soil nutrient content in black soil of Heilongjiang province, in this paper, we utilized statistical theory and spectral analysis method, researched the relationship of three kinds of soil nutrient content and soil spectrum to established hyperspectral inversion model of soil total nitrogen, available phosphorus, available kalium content. We acquired the aerial hyperspectral data by using CASI-1500 and SASI-600 linear array push-broom imaging spectrometers. Preprocessing of calibration and atmospheric radiation correction of Airborne Hyperspectral raw radiation data was studied. 96 samples were evenly sampled. In order to increase the representativeness of samples, 96 groups of samples were collected from 3-5 samples collected from 15 meters around the sampling point, and 1.5 kg was retained after mixing. After air-drying, mixing and grindingetc, it is used for the contents of total nitrogen, available phosphorus and available kalium were obtained through laboratory tests. The content of total nitrogen, available phosphorus and available kalium was determined by NaOH diffusion method, NaHCO3extraction-molybdenum blue colorimetry and NH4OAC extraction-flame photometry. Referring to Kennard-Stone method, 72 groups of representative samples were selected as model samples for nutrient content prediction, and 24 groups were model prediction samples. 96 black soil samples were sorted according to nutrient content, and the spectral transformation in the visible near red range was analyzed. The change rule of total nitrogen is that the reflectance decreases with the increase of content. The first order differential spectra at 580 nm were significantly correlated with total nitrogen and available phosphorus content, with a correlation coefficient of -0.43 and -0.36, respectively. The first-order differential spectra at 1 730-2 200 nm were significantly correlated with K2O, and the maximum correlation coefficient was -0.31. Compared with the original spectral waveform, the correlation coefficient between the first derivative and three nutrient contents fluctuated sharply, and the positive and negative cross-sections were relatively sharp, with more peak coefficients .After spectral contrast analysis and correlation coefficient calculation, 86 bands with higher correlation coefficient were selected for the study under the first order differential variation. On black soil airborne hyperspectral data processing, the application of partial least squares regression (PLSR) and BP neural network method respectively establish soil nutrient content of high spectral quantitative inversion model. The results showed that RPIQ values (Difference between the third and the first quartile of sample observations ratio to RMSE) of total nitrogen PLSR and BP neural network prediction model were 2.42 and 2.80, respectively. The RPIQ values of effective phosphorus PLSR and BP neural network model were 0.83 and 1.67 respectively. The RPIQ values of the available kalium PLSR and BP neural network models were 2.00 and 2.33 respectively. Experiments showed that the spectral quantitative prediction model of soil total nitrogen and available kalium has good accuracy and prediction ability. Nitrogen, phosphorus and potassium, and the spatial distribution of nutrient content in black soil were obtained. However, the prediction effect of effective Phosphorus was not particularly ideal, which could only meet the requirements of approximate quantitative prediction. At the same time, the BP neural network modeling has better accuracy and prediction ability than the partial least square modeling, and the prediction accuracy increased by 6.5%, 10.1% and 6.6% respectively. Due to the limitation of soil samples and other conditions, more samples are needed to verify the universality of the model. More data mining methods are expected to establish more robust prediction models, which will provide more reliable information for the prediction and evaluation of black soil quality information.

soils; remote sensing; models; partial least squares method; BP neural network

杨越超,赵英俊,秦 凯,赵宁博,杨 晨,张东辉,崔 鑫. 黑土养分含量的航空高光谱遥感预测[J]. 农业工程学报,2019,35(20):94-101.doi:10.11975/j.issn.1002-6819.2019.20.012 http://www.tcsae.org

Yang Yuechao, Zhao Yingjun, Qin Kai, Zhao Ningbo, Yang Chen, Zhang Donghui, Cui Xin. Prediction of black soil nutrient content based on airborne hyperspectral remote sensing[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(20): 94-101. (in Chinese with English abstract) doi:10.11975/j.issn.1002-6819.2019.20.012 http://www.tcsae.org

2019-06-05

2019-10-07

国家自然科学基金项目(41602333);东北黑土地1:25万土地质量地球化学调查(DD20160316);遥感信息与图像分析技术国家级重点实验室基金项目(ZJ2019-1)

杨越超,工程师,主要从事高光谱遥感及GIS的科研工作。Email:ycyangcug@qq.com

10.11975/j.issn.1002-6819.2019.20.012

S15

A

1002-6819(2019)-20-0094-08

猜你喜欢

黑土全氮反演
自然资源部:加强黑土耕地保护
添加木本泥炭和膨润土对侵蚀退化黑土理化性质的影响*
反演对称变换在解决平面几何问题中的应用
基于Sentinel-2遥感影像的黑土区土壤有效磷反演
基于ADS-B的风场反演与异常值影响研究
利用锥模型反演CME三维参数
寒地黑土无公害水产品健康养殖发展思路
一类麦比乌斯反演问题及其应用
西藏主要农区土壤养分变化分析
三峡库区消落带紫色土颗粒分形的空间分异特征