土壤有机质高光谱特征与波长变量优选方法

发布时间：2018-04-27 00:26

本文选题：土壤有机质 + 高光谱　；参考：《中国农业科学》2017年22期

【摘要】：【目的】探究土壤有机质的高光谱特征及响应规律,优选土壤有机质的敏感波长,降低土壤有机质高光谱估测模型复杂度,提高模型稳健性,为利用高光谱技术对农田土壤肥力的定量监测提供理论支撑。【方法】采集江汉平原潮土土样130个,将其中40个样本作为训练集,测量其去有机质前、后的土壤有机质含量及光谱数据,计算差值及变化率,分析土壤有机质含量变化对光谱特征的影响,结合无信息变量消除(uninformative variables elimination,UVE)、竞争适应重加权采样(competitive adaptive reweighted sampling,CARS)变量优选方法确定土壤有机质敏感波长;采用45个建模集样本,基于偏最小二乘回归(partial Least Squares Regression,PLSR)和反向传播神经网络(back propagation neural network,BPNN)建立土壤有机质含量的估算模型;利用45个验证集样本检验敏感波长对同类土壤的适用性。【结果】通过有机质去除试验,供试土壤的平均光谱反射率在全波段均有所增加,在可见光波段变化率高于近红外波段;比较UVE、CARS、UVE-CARS、CARS-UVE这4种变量优选方法,得到最佳变量优选方法为UVE-CARS,该方法从2001个波长变量中优选得到84个变量作为土壤有机质的敏感波长,分布于561—721、1 920—2 280 nm波段覆盖范围;基于敏感波长的PLSR、BPNN模型性能均优于全波段模型,其中,基于敏感波长的BPNN模型的估测能力高于PLSR,模型验证集R~2、RMSE、RPD、MAE、MRE值分别为0.74、1.33 g·kg~(-1)、2.02、1.04 g·kg~(-1)、6.2%,可实现土壤有机质含量的有效估测。【结论】通过训练集获得的土壤有机质敏感波长,能够实现对该试验区同种土壤类型样本土壤有机质含量的有效估测;利用去有机质试验结合变量优选方法确定的敏感波长建模,不仅将输入波长压缩至全波段波长数目的 4.2%,而且提升了模型估测精度,降低了变量维度和模型复杂度,为快速准确评估农田土壤有机质含量提供了新途径。
[Abstract]:[objective] to study the hyperspectral characteristics and response law of soil organic matter, to optimize the sensitive wavelength of soil organic matter, to reduce the complexity of hyperspectral estimation model of soil organic matter, and to improve the robustness of the model. In order to provide theoretical support for quantitative monitoring of farmland soil fertility by using hyperspectral technique. [methods] 130 soil samples were collected from Jianghan Plain, 40 of which were used as training set to measure their organic matter removal. The soil organic matter content and spectral data were calculated, the difference and change rate were calculated, and the influence of the change of soil organic matter content on the spectral characteristics was analyzed. The sensitive wavelength of soil organic matter was determined by the method of competitive adaptive reweighted sampling with competitive selection method, and 45 modeling set samples were used to determine the sensitive wavelength of soil organic matter by combining the elimination of uninformative variables elimination method with non-information variable, and the competitive adaptive reweighted sampling method was used to determine the sensitive wavelength of soil organic matter. The estimation model of soil organic matter content was established based on partial Least Squares regression (partial Least Squares regression) and backpropagation neural network back propagation neural network (BPNN). The applicability of sensitive wavelength to the same soil was tested by 45 validation set samples. [results] through the organic matter removal test, the average spectral reflectance of the tested soil increased in the whole wave band. The change rate of visible light wave band is higher than that of near infrared wave band, comparing the four methods of selecting UVE-CARS-UVE variables, the best method is UVE-CARSs, 84 variables are obtained as sensitive wavelengths of soil organic matter from 2001 wavelength variables, and the results show that UVE-CARSs can be used as the sensitive wavelength of soil organic matter, and UVE-CARSs can be used as the sensitive wavelength of soil organic matter. The performance of PLSR-BPNN model based on sensitive wavelength is better than that of full-band model. The ability of BPNN model based on sensitive wavelength to estimate soil organic matter is higher than that of BPNN model, and the MRE values of model verification set RMS-E are 0.74 ~ 1.33 g 路kg ~ (-1) ~ (-1), respectively, which can realize the effective estimation of soil organic matter content. [conclusion] the sensitive wavelength of soil organic matter can be obtained by training set. It can realize the effective estimation of the soil organic matter content of the same soil type sample in the experimental area, and model the sensitive wavelength determined by the organic matter removal test combined with the variable optimal selection method. It not only compresses the input wavelength to 4.2 wavelength of the whole wavelength, but also improves the accuracy of the model estimation, reduces the variable dimension and the complexity of the model, and provides a new way for the rapid and accurate evaluation of soil organic matter content in farmland.
【作者单位】：华中师范大学地理过程分析与模拟湖北省重点实验室/华中师范大学城市与环境科学学院;
【基金】：国家自然科学基金(41401232) 中央高校基本科研业务费专项资金(CCNU15A05006) 华中师范大学研究生教育创新资助项目(2017CXZZ007)
【分类号】：S153.621

【相似文献】