从田间采集了150个田间土壤样本, 在分析了所有样本的土壤参数统计特征之后, 对原始近红外光谱数据进行了聚类分析, 分别得到了50个土壤全氮和50个土壤有机质的等价样本及其对应光谱。 对样本光谱曲线进行8层Biorthogonal小波包分解, 分解得到的最低低频［80］结点对应着土壤水分以及土壤质地的光谱变化趋势, 最高高频［8 255］结点对应着土壤粒度、 光谱仪精度等引起的高频震荡。 对以上两个结点进行重构并从样本光谱曲线中剔除以上影响成分, 得到了对应的土壤参数特征光谱。 基于特征光谱建立了土壤参数偏最小二乘回归模型： 全氮偏最小二乘预测模型的预测系数rc达到了0.960, 验证系数rv达到了0.920； 有机质偏最小二乘预测模型的预测系数rc达到0.922, 验证系数rv达到0.883。 模型精度明显提高, 满足实际生产的需要。

Using the method of wavelet analysis, the NIR spectra of soil samples were decomposed and reconstructed, and higher precision PLS models were established to estimate soil parameter (TN, SOM). One hundred fifty soil samples were collected from a winter wheat field and the NIR spectra of all samples were measured. Firstly, experiment statistic features were analyzed aiming at all soil samples, and the system clustering was carried out for TN and SOM respectively. Then 50 new TN samples and their corresponding spectra, and 50 new SOM samples and their corresponding spectra were obtained. Secondly, the PLS models were established with these new samples based on their corresponding spectra. The models showed a certain amount of accuracy, but it was still not practical. Therefore, wavelet analysis of NIR spectra was tried. The wavelet packet decomposing by eight-level biorthogonal algorithm was carried out, and 256 nodes were gotten. The lowest approximation signal is corresponding to soil moisture and soil texture spectrum trend. The maximal detail signal is corresponding to the high-frequency turbulence caused by the soil particle size, precision of spectrometer, and other uncertainties. After reconstructing these two nodes and then removed from the original spectra, the characteristic spectra corresponding to each soil parameter were acquired. Finally, the PLS models were established for TN and SOM content respectively: for TN content, the calibration coefficient of the PLS model is 0.960, the validation coefficient is 0.920; and for SOM content, the calibration coefficient of the PLS model is 0.922, and the validation coefficient is 0.883. It was showed that the accuracy of each model was highly improved and the models were able to meet the needs of actual production. The research results conclude that wavelet analysis can eliminate or substantially reduce the factors outside the parameters. It can also remove the obstacles in establishing linear models of soil parameters, and it is feasible and potential method for the real-time estimation of soil parameters.