基于高光谱的区域土壤颗粒组成及有机质预测模型尺度转换

Scale transformation of regional soil particle composition and organic matter prediction models based on hyper-spectrum

以内蒙古河套灌区3个尺度下1 024个样本的高光谱为模型输入,黏粒、粉粒、砂粒及有机质质量百分数为模型输出,通过多元回归、支持向量机及BP神经网络方法建立基于中尺度的反演模型,将其尺度上推至大尺度及下推至小尺度,并对其尺度转换的适用性进行评价.结果表明:基于中尺度建立的高光谱与土壤颗粒组成及有机质的反演模型均可以较好地应用于其他2个尺度,多元回归方法在其他2个尺度上的相关性为0.33~0.60,支持向量机方法为0.41~0.52,BP神经网络方法为0.52~0.72,其中BP神经网络方法建立的模型在其他2个尺度上具有更好的适用性;不同参数中,黏粒、粉粒、砂粒及有机质的相关系数分别为0.44~0.62,0.37~0.72,0.42~0.72及0.33~0.56,即颗粒组成的效果整体好于有机质质量百分数.

Mesoscale inversion models of mass percentage of clay, power, sand and organic matter were established by using multiple regression, support vector machine and BP neural network based on the hyper-spectrum of 1 024 samples with 3 scales in Hetao Irrigation District of Inner Mongolia. The mesoscale scale was shifted to large and small scales, and the applicability of scale transformation was assessed for these models. The results show that the mesoscale inversion models of soil particle composition and organic matter are well applicable at the other two scales with correlation coefficient of 0.33-0.60 in multiple regressions, 0.41-0.52 in support vector machine and 0.52-0.72 in BP neural network. Clearly, the models based on BP neural network method show even better applicability at the other two scales. The correlation coefficients of clay, power, sand and organic matter are 0.44-0.62, 0.37-0.72, 0.42-0.72 and 0.33-0.56, respectively, suggesting that the fitting effect of particle compositions is better than that that of organic matter as a whole.