• 论文与研究报告 •

### 基于RBF组合模型的山地红壤有机质含量光谱估测

1. 1. 江西农业大学林学院 江西省鄱阳湖流域农业资源与生态重点实验室 南方粮油作物协同创新中心 南昌 330045;
2. 南京农业大学公共管理学院 南京 210095
• 收稿日期:2016-12-20 修回日期:2018-05-07 出版日期:2018-06-25 发布日期:2018-07-02
• 基金资助:
国家自然科学基金项目（41361049）；土壤与农业可持续发展国家重点实验室（中国科学院南京土壤研究所）项目（0812201202）。

### Spectrum Based Estimation of the Content of Soil Organic Matters in Mountain Red Soil Using RBF Combination Model

Xie Wen1, Zhao Xiaomin1, Guo Xi1, Ye Yingcong1, Sun Xiaoxiang1, Kuang Lihua2

1. 1. College of Forestry, Jiangxi Agricultural University Key Laboratory of Poyang Lake Watershed Agricultural Resources and Ecology of Jiangxi Province Southern Regional Collaborative Innovation Center for Grain and Oil Crops in China Nanchang 330045;
2. College of Public Administration, Nanjing Agricultural University Nanjing 210095
• Received:2016-12-20 Revised:2018-05-07 Online:2018-06-25 Published:2018-07-02

Abstract: [Objective] It is of great significance for rapid acquisition of soil fertility parameters using spectral technology to explore the quantitative estimation of the contents of soil organic matters.[Method] Two hundred and forty eight soil samples at 0-30 cm depth were collected from mountain red soil region of northern Jiangxi Province. The spectral reflectance of mountain red soil samples was measured by an ASD FieldSpec3 instrument under laboratory conditions. Meanwhile the content of organic matters of each soil sample was analyzed using potassium dichromate external heating method in laboratory. Correlation analyses between raw mountain red soil spectral reflectance and organic matter content mountain red soil were conducted. Kennard-Stone algorithm was used to divide mountain red soil samples into calibration sets with 186 samples and prediction sets with 62 samples. Based on the full band (400-2 450 nm) of mountain red soil spectra in this study, the partial least squares regression (PLSR), BP neural network (BP) and support vector machine regression (SVMR) were selected to obtain and rebuild the prediction result, and establish fixed weight and variable weights in combination model with the minimum absolute error sum as the goal. Then, the combination model can be built based on radial basis function (RBF) neural network. The optimal combination model with different weighting and with or without reconstructed data was investigated. Accuracies of the content predictions of organic matters of the mountain red soil were evaluated by root mean squared error (RMSE), ratio of partial deviation (RPD) and determination coefficients (R2).[Result] The results showed the best single model is SVMR, its determination coefficient (R2), the root mean square error (RMSE) and the ratio of standard error of performance to standard deviation (RPD) value of validation set were 0.64, 9.76 g·kg-1 and 1.67. Under the condition of prediction data, the combination model of variable weights is better than the fixed weight combination model, while under the condition of prediction data reconstruction, the combination model of fixed weight is slightly better than the variable weight combination model, but the estimation accuracy of the two models is very close. The optimal model of the combination model is the reconstruction of fixed weight combination model, its determination coefficient (R2), the root mean square error (RMSE) and ratio of standard error of performance to standard deviation (RPD) value of validation set were respectively 0.87, 7.91 g·kg-1 and 2.06. Therefore, the accuracy of the RBF combination model is better than that of the single model, indicating that it is feasible to estimate the content of organic matters in mountain red soil by RBF combination model.[Conclusion] The rapid content estimation of organic matters in mountain red soil showed that the single model has the characteristics of simple operation and fast calculation. Therefore, it has a greater value of application. However, the combination model was able to make full use of various sample information, therefore, effectively reducing the impacts of random factors on the use of single model, and hence to improve the precision of estimating the content of organic matters in mountain red soil.