基于高光谱特征指数的土壤有机质含量建模

以江苏中部的水稻土和潮土为研究对象,采集178个表层土壤(0~20 cm)样品,并测定了土壤有机质含量(Soil Organic Matter,SOM)。运用ASD FieldSpec 3光谱仪测量了土壤的高光谱曲线,首先对原始光谱进行倒数对数和去包络线变换,分析了不同SOM含量梯度和土壤类型的高光谱特征。其次,基于原始光谱、倒数对数变换和去包络线变换等三种光谱数据,分别计算弓曲差、差值指数、比值指数和归一化指数等光谱特征指数,并分析其与SOM含量的相关性。最后,筛选光谱特征指数建立SOM的回归预测模型,并比较模型精度。结果表明:(1)SOM含量与原始光谱呈极显著负相关,与倒数对数光谱呈极显著正相关,且在400~900 nm波段相关性最强,相关系数绝对值在0.6以上。去除包络线处理后,土壤光谱曲线特征差异明显,在420 nm、480 nm、660 nm和900 nm附近出现了明显吸收谷。(2)原始光谱、倒数对数变换和去包络线变换光谱在600 nm处的弓曲差与SOM含量极显著相关(P<0.01),相关系数分别为–0.66,0.61和–0.33。(3)利用3种光谱数据的差值指数、比值指数和归一化指数分别结合弓曲差,建立的SOM预测模型效果较好,建模的R2和RMSE分别介于0.56~0.64和4.98~5.50 g·kg^–1,验证的R2和RMSE介于0.67~0.73和3.21~3.51 g·kg^-1。为快速有效测定苏中平原SOM含量提供技术支持。

Modeling of soil organic matter content using hyperspectral characteristic index

Objective] The analysis of soil properties using routine chemical analysis method is rather costly and time-consuming, and so hard to meet the requirement for handling large volumes of soil samples fast and efficiently to monitor soil properties. Continuous soil spectral curves obtained with the aid of the hyperspectral technology encompass abundant spectral information, and reflect comprehensively various soil attribute information. Therefore, modeling can be done to predict some soil properties efficiently and accurately based on the hyperspectral technology. This paper was oriented to build a model for predicting soil organic matter (SOM) content based on hyperspectral feature indices with an expectation to provide a new method for rapid and effective determination of SOM content. Method] In this study, a total of 178 soil samples were collected from the surface soil layers (0~20 cm) of farmlands of paddy soil and fluvo-aquic soil in the central plain of Jiangsu province for analysis of SOM content. Hyperspectral curves of the soil samples were obtained with the aid of the ASD FieldSpec 3 spectrometer. Firstly, the original spectra were processed with the algorithms of reciprocal log transformation (Log(1/R)) and continuum-remova l (CR) for analysis of hyperspectral characteristics of the soil samples different in SOM content and in soil type. Secondly, based on the data of the original, Log (1/R) and CR spectra, spectral feature indices, including deviation of arch (DOA), difference index (DI), ratio index (RI), and normalized difference index (NDI), were calculated, and relationships of SOM content with the four indices were analyzed. Finally, linear regression models for SOM content were established based on the selected spectral feature indices. Accuracies of the models were evaluated and compared. Result] Results show:(1) SOM content was significantly and negatively related to original spectra, but significantly and positively to reciprocal log spectra. The relationship was the most significant at the waveband of 400~900 nm, with the absolute correlation coefficient value reached above 0.6. After the spectral curves being CR transformed, their differences in characteristic became extraordinarily significant different, and significant absorption valleys appeared near 420 nm, 480 nm, 660 nm, and 900 nm; (2) The DOAs of the original, Log (1/R) and CR spectra showed extra-significant relationships with SOM content (P<0.01) at wavelength of 600 nm, with correlation coefficient being -0.66, 0.61 and-0.33, respectively; and (3) Based on the combinations of DOA with DI, RI, and NDI of the three different spectra, the model established for predicting SOM content performed quite effectively, with relative percent deviation ranging from 1.78 to 1.94, R² from 0.56 to 0.64, and RMSE from 4.98 g·kg^-1to 5.50 g·kg^-1. And the validation sets had R² ranging from 0.67 and 0.73, and RMSE from 3.21 g·kg^-1 to 3.51 g·kg^-1.Conclusion] The spectral feature indices, including DOA, DI, RI, and NDI, can be used effectively for modeling for SOM content, and the model may explain about 67%~73% of the SOM variability. The model established with RI and DOA of the Log (1/R) spectrum is an optimal one.