基于机载高光谱影像的南方水稻土有机质含量估算

Estimation of Organic Matter Content in Southern Paddy Soil Based on Airborne Hyperspectral Images

水稻土中有机质光谱常常受到水分、秸秆等土壤背景的影响,为了减弱或去除非有机质组分对有机质光谱的影响,构建南方水稻土有机质估算模型。利用机载高光谱(GaiaSky-Mini2-VN)作为数据源,对原始反射率进行单一和组合变换(去除包络线、倒数、对数、一阶微分、二阶微分单一变换和倒数一阶微分、对数一阶微分、倒数对数组合变换)提取一维特征光谱;通过对变化后光谱进行比值和归一化处理,提取二维特征光谱;构建基于特征光谱的线性(多元回归、偏最小二乘)和非线性(反向传播神经网络、支持向量机)有机质预测模型,监测南方水稻土有机质含量。结果表明:一维光谱变换能显著增强光谱对有机质响应的敏感度,二维光谱变换能充分挖掘光谱信息,增强有机质与光谱之间的内在联系,提高建模精度。非线性模型(BPNN、SVM)尤其是BPNN对土壤有机质拟合性好,建模精度高。基于原始反射率比值指数建立的BPNN模型建模精度达到0.952,检验精度达到0.889,建模效果最优。该结果适用于南方水稻土有机质监测,对机载高光谱在土壤有机质监测中的特征波段提取和建模方法选择具有重要借鉴意义,对现代农业发展管理提供新的思路。

Organic matter spectra in paddy soils are often influenced by soil background such as water and straw. In order to weaken and remove the influence of non-organic matter components, an estimation model of organic matter in paddy soils in southern China was constructed. In this paper, Airborne Hyperspectral(GaiaSky-Mini2-VN) was used as data source to extract one-dimensional characteristic spectra by single and combined transformation of the original reflectance(continuum removal, inverse, logarithmic, first-order differential, second-order differential single transformation and inverse first-order differential, logarithmic first-order differential, reciprocal logarithmic combined transformation);two-dimensional characteristic spectra were extracted by ratio and normalization processing of the transformed spectra;linear(multiple linear regression, partial least squares) and non-linear(back propagation neural network and support vector machine) organic matter prediction models based on characteristic spectra were constructed to monitor the organic matter of paddy soils in Southern China. The results showed that one-dimensional spectral transformation could significantly enhanced the sensitivity of spectral response to organic matter. Two-dimensional spectral transformation fully mined spectral information, enhanced the intrinsic relationship between organic matter and spectrum, and improved the modeling accuracy. Nonlinear models(BPNN, SVM), especially BPNN, had good fit for soil organic matter and high modeling accuracy. The BPNN model based on the original reflectance ratio indexhad the best modeling effect, with modeling accuracy of 0.952 and testing accuracy of 0.889. Above results were applicable to the monitoring of organic quality of paddy soils in southern China. It could provide important reference in feature band extraction and modeling method selection of soil organic monitoring by airborne hyperspectral spectroscopy and put forward new ideas for the management of modern agricultural development.