以新疆奇台县为研究区域, 选取该县40个土壤样本, 采用多元线性逐步回归法和人工神经网络法两种方法分别建立了土壤有机质含量的反演模型, 并对模型进行了检验。 结果发现: 不同模型的精度值各异, 其拟合效果从高到低依次为人工神经网络（ANNs）集成模型＞单个人工神经网络（ANNs）模型＞多元逐步回归(MLSR)模型。 人工神经网络的线性和非线性逼近能力较强, 而其集成模型作为提高反演模型精度的重要手段, 相关系数高达0.938, 均方根误差和总均方根误差最小, 分别仅为2.13和1.404, 对土壤有机质含量的预测能力与实测光谱非常接近, 分析结果达到了较实用的预测精度, 为最优拟合模型。

The present paper, based on the Qitai county of Xinjiang, selected 40 soil samples, and used two methods respectively, i.e. multiple linear stepwise regression(MLSR) and artificial neural network (ANNs) , to establish the inversion and predieting model of soil organic matter (SOM) content and the model test from measured reflectance spectra and relative test were carried through to the models. Through quantitative analysis, the conclusions can be drawn as follows that the precision values of the different models vary from one to another, the model fitting effects order from high to low is that the integrated model for artificial neural networks (ANNs) is best, single artificial neural networks (ANNs) model is better, while stepwise multiple regression (MLSR) models are worse. Artificial neural networks (ANNs) has the strong abilities of linear and nonlinear approximation, while its integrated model for artificial neural networks (ANNs) is an important way to improve the inversion accuracy of soil organic matter (SOM) content, with the correlation coefficient up to 0.938, root mean square error and total root mean square error are minimum, being 2.13 and 1.404 respectively, and the predictive ability of the soil organic matter (SOM) content are very close to the measured spectrum,so the analysis results can achieve a more practical prediction accuracy for the best fitting model.