不同类型土壤的光谱特征及其有机质含量预测

Spectral Characteristics and Estimation of Organic Matter Contents of Different Soil Types

【目的】构建适合土壤有机质含量估测的高光谱参数及定量反演模型。【方法】系统分析中国中、东部地区5种不同类型土壤风干样本有机质含量与350～2500nm波段范围高光谱反射率之间的关系,利用特征光谱参数和BP神经网络建立土壤有机质的定量估测模型。【结果】光谱一阶导数构成的两波段光谱参数与土壤有机质含量的相关性明显优于原始光谱,尤其采用Norris平滑滤波后导数光谱效果更好。光谱参数构成形式以差值指数最好,其次为比值和归一化指数。与土壤有机质含量相关程度最高的光谱参数是由可见光区554nm和近红外区1398nm两个波段的一阶导数组合而成的差值指数DI(D554,D1398),两者呈显著指数曲线关系,拟合方程为y=184.2×exp[-1297×DI(D554,D1398)],决定系数为0.90。经不同类型土壤的观测资料检验,模型预测决定系数为0.84,均方根误差RMSE为3.64,相对分析误差RPD为2.98,显示估测模型具有较好的预测精度。另外,利用BP神经网络结合偏最小二乘法(PLS)对导数光谱进行分析,提取贡献率达到99.56%的前6个主成分建立了三层BP神经网络模型,模型决定系数为0.98,经不同类型土壤的观测资料检验,模型预测决定系数为0.96,RMSE为2.24,相对偏差RPD为4.83。比较利用DI(D554,D1398)和BP网络进行土壤有机质含量的预测结果,前者精度低于后者,但可以满足土壤有机质监测的需要。【结论】利用差值光谱指数DI(D554,D1398)和BP神经网络模型均可实现对土壤有机质的精确估测。

【Objective】 The objectives of the present study were to determine the key spectral parameters and models for estimating SOM content. 【Method】 The dried sample of five different soil types in China were analyzed for SOM content and hyperspectral reflectance within 350-2 500 nm, quantitative models of SOM using spectral index and BP neural network were established, respectively. 【Result】The results showed that correlation between spectral indices which composed of first derivative and SOM content were obviously stronger than those composed of original reflectance, especially derivative with Norris smoothing filter. The correlation sequence of SOM to different index types was DI〉RI〉ND which composed of spectral reflectance or the first derivative spectra. DI composed of first derivative of 554 nm and 1 398 nm gave a better prediction performance, with equation as y=184.2xexp[-1297×DI（D554, D1398）], coefficient of determination was 0.90. Testing of the monitoring models with independent data from different soil types indicated that R2, RMSE and RPD of validation were 0.84, 3.64 and 2.98, respectively. In addition, the scores computed by PLS were applied as input of BP neural network developed with over 99.56% of cumulative proportion of correlation matrix. R2 of calibration model was 0.98, and R2, RMSE, RPD of validation were 0.96, 2.24 and 4.83, respectively. Compared with BP neural network model, DI（D554, D1398） had a little lower prediction precision, but it could meet need of estimating of SOM content. [Conclusion] It is concluded that both of methods based on DI（D554, D1398） and BP neural network can estimate SOM content accurately.