基于可见-近红外光谱预处理建模的土壤速效氮含量预测

以皖南地区采集的188份黄红壤样本为研究对象,利用地物非成像光谱仪获取原始光谱数据。首先,分析样本在350~1 657 nm波段经过预处理变换的平均光谱反射率曲线特征,再基于原始光谱,以及经29种预处理变换后的光谱,分别结合偏最小二乘回归(PLSR)和径向基核函数(RBF)-PLSR算法,建立60个针对土壤速效氮含量的预测模型,并进行模型优化;然后,以模型的决定系数(R²)和相对分析误差(RPD)来评价模型性能。结果显示,基于Savitaky-Golay卷积平滑和对数变换预处理的光谱,用PLSR建立的模型最适用于土壤速效氮含量的校正预测,其在建模集中R²=0.94、RPD=3.88,预测集中R²=0.91、RPD=3.38。该模型达到A类预测精度,可实现对土壤速效氮含量的定量估测。

Prediction of soil available nitrogen content based on visible and near infrared spectroscopy preprocess and modeling

In the present study, 188 yellow-red loam soil samples were collected in Southern Anhui, and the original spectrum was obtained by non-imaging spectrometer. Firstly, the characteristics of the average spectral reflectance curve at 350-1 657 nm were analyzed after preprocessing. Based on the original spectrum and spectra after 29 preprocesses, a total of 60 models were constructed either by partial least squares regression (PLSR) or radial basis function (RBF)-PLSR, and the constructed models were optimized and evaluated by the models' decision coefficient (R²) and relative percent deviation (RPD). It was shown that the PLSR model constructed on spectra after Savitaky-Golay filtering and log transformation (SG+LG/PLSR model) was most suitable for the prediction of soil available nitrogen content. Its R² and RPD were 0.94 and 3.88 in calibration set and were 0.91 and 3.38 in prediction set, which belonged to A level, and indicated that this model was feasible for soil available nitrogen content prediction.