基于支持向量机的皖北旱作区表层土壤有机质含量估测

Estimation of soil organic matter content in surface soil of dryfarmland in northern Anhui Province based on support vector machine

文中旨在探究支持向量机(Support Vector Machine,SVM)的旱作区表层土壤有机质含量快速估测方法。以皖北旱作区为研究对象,利用两期Landsat8 OLI影像获取采样点光谱反射率,采用不同的波段变换方法构建光谱参量,并结合实测值分析不同建模方式与核函数下模型的估算效果。在实现模型优选后进行全区范围内的有机质含量空间特征分析。实验表明,对原始波段进行波段组合形式的变换后可有效提升光谱数据对土壤有机质含量的解释能力。在4种建模方式中,基于多元逐步回归模型选取的波段组合参量(即优化参量)所建立的估测模型整体具有较好的估测能力,其中,在三次多项式核函数下估测模型对研究区3种土类的整体估测效果较好,其平均相对误差(Mean Relative Error,MRE)达到17.73。上述现象表明,利用两期影像波段组合变换后得到的优化参量建立估测模型具备一定的优势,且在三次多项式下SVM模型估测效果较好。

This paper aims to explore the rapid estimation method of surface soil organic matter content in dry farmland based on support vector machine.Taking the dry farmland area in northern Anhui as the research object,by using two phases of Landsat8 OLI image to obtain the spectral reflectance of sampling points,it uses different band transform methods to construct the spectral parameters,and the estimated results of different modeling modes and kernel functions are analyzed with the measured values.After the optimization realization of the model,the spatial characteristics of the organic matter content in the whole area are analyzed.The experiment shows that the combined transformation of the original bands can effectively improve the interpretation ability of the spectral data on the content of soil organic matter.In the four modeling modes,the estimation model based on the band combination parameters(optimization parameters)selected by the multivariate stepwise regression model has the best estimation ability,and the estimation model under the cubic polynomial kernel function has a good overall estimation effect on the three soil types in the study area,and its mean relative error(MRE)reaches 17.73.The above phenomenon shows that using the optimization parameters obtained after the combined transformation of the two phases image bands to establish an estimation model has certain advantages,and the SVM model estimation effect is better under the cubic polynomial.This research can provide some technical support for regional soil composition monitoring.