摘要
为快速无损地获取托克托县土壤全氮含量,满足当今精准农业的要求,文章以研究区内120个采样点的土壤全氮含量与高光谱数据为数据源,利用分数阶微分(1~2阶)间隔0.1对光谱数据进行处理,筛选敏感波段,利用支持向量机(support vector machine,SVM)与BP神经网络模型共建立24个土壤全氮反演模型,结果表明:①经过分数阶微分处理后,光谱的波峰、波谷处信息被放大,随分解尺度的增加,其余波段的反射率逐渐趋于0;②原始光谱与土壤全氮的皮尔森相关系数r=0.61,经分数阶微分处理后,在1.1阶处达到最大值r=-0.67,绝对值较之前提升了0.06;③BP神经网络预测模型结果优于SVM预测模型结果,本研究最佳土壤全氮预测模型为1.1阶微分处理后建立的BP神经网络模型,建模集R 2为0.75,均方根误差(root mean square error,RMSE)为0.16,验证集R 2为0.71,RMSE为0.16,相对分析误差(relative percent deviation,RPD)为2.06,可有效反演当地土壤全氮含量,相对于原始光谱建立的BP神经网络模型精度有较高提升。因此,利用1.1阶微分处理后的高光谱数据建立BP神经网络模型可实现对研究区土壤全氮含量的反演预测,可为当地精准农业的发展提供理论参考与技术支撑。
This study aims to determine the total nitrogen content(TNC)in soils in Tuoketuo County quickly and nondestructively,thus meeting the requirements of precision agriculture.With the soil TNC and hyperspectral data of 120 sampling sites in the study area as the data source,this study processed the hyperspectral data using the 1~2 orders fractional order differential(FOD)interval of 0.1 to screen the sensitive wavebands.Then,this study built 24 inversion models for soil TNC using the support vector machine(SVM)and the back propagation neural network(BPNN).The results are as follows:①After FOD processing,the information at the wave crests and troughs of the spectra was amplified,and the reflectance of the remaining wavebands approached zero gradually with an increase in the decomposition scale.②The Pearson correlation coefficient between original spectra and soil TNC was r=0.61.This correlation coefficient was up to a maximum of 0.67 at 1.1-order after FOD processing,increasing by 0.06.③The BPNN prediction models outperformed the SVM prediction models.The optimal soil TNC prediction model was the BPNN model built after 1.1-order differential processing.This model yielded an R 2 of 0.75 and a root mean square error(RMSE)of 0.16 for the modeling set and an R 2 of 0.71 and an RMSE of 0.16 for the verification set,with a relative percent deviation(RPD)of 2.06.This model produced effective inversion results of the soil TNC in the study area,with a much higher accuracy than the BPNN model built using original spectra.Therefore,the BPNN model built using hyperspectral data through 1.1-order differential processing allows for the inversion-based prediction of soil TNC in the study area,providing a theoretical reference and technical support for local precision agriculture.
作者
陈昊宇
项磊
高贺
牟金燚
索晓晶
滑博伟
CHEN Haoyu;XIANG Lei;GAO He;MU Jinyi;SUO Xiaojing;HUA Bowei(Hohhot General Survey of Natural Resources Center,China Geological Survey,Hohhot 010010,China)
出处
《自然资源遥感》
CSCD
北大核心
2023年第3期170-178,共9页
Remote Sensing for Natural Resources
