关键词: 盐渍土, 有机质, 光谱响应, 高光谱预测

Abstract:

【Objective】The aim of this study was to investigate the spectral response of soil organic matter and salt, to analyze the effects of different salt content on soil organic matter prediction models, and to establish a rapid and effective hyperspectral prediction model for organic matter content in saline soils.【Method】In this study, according to different salinity contents for indoor hyperspectral measurements, the soil samples were divided into four groups of non-saline (SA), slightly saline (SB), moderately saline (SC), and heavy saline (SD). Then, ANOVA was used to explore the degree of organic matter and salinity spectral response of soils with different salinity degrees respectively. The raw spectra reflectances (raw spectral reflectance, R) were subjected to first order differential reflectance (first order differential reflectance, FD), continuous statistical removal (continuous statistical removal, CR), logarithmic (logarithmic, Log) and multiple scatter correction (multipication scatter correction, MSC) transformations were applied to the raw spectra reflectance; finally, three soil organic matter prediction models, namely multiple linear regression (multiple linear regression, MLR), partial least squares regression (partial least squares regression, PLSR) and support vector regression (support vector machine, SVR), were constructed based on four sets of samples of saline soils combined with the four transformed spectra, respectively.【Result】Soil organic matter and salinity had significant spectral response in the range of 400-900 nm and the change pattern were the same basically, and the sensitive bands of the two overlap. Modeling by dividing different salinity levels could improve the prediction accuracy of soil organic matter, but the prediction accuracy of the model decreased with the increase of salinity content. FD treatment could better highlight the difference of spectral characteristics and improved the correlation between organic matter content and spectral reflectance. Comparing the results of the three models, the highest accuracy of the soil organic matter prediction model was established using FD treatment combined with SVR, and the coefficients of determination R² of the optimal result modeling set and validation set were 0.86 and 0.82, respectively, the root mean square error RMSE was 2.71 and 2.96 g·kg^-1, respectively, and the ratio of prediction to deviation RPD was 2.42.【Conclusion】Soil salinity and organic matter overlapped in the sensitive bands near the visible wavelength (400-900 nm), and the accuracy of the organic matter prediction model could be effectively improved by classifying different salinity levels.

Key words: saline soil, organic matter, spectral response, hyperspectral prediction