关键词: 土壤盐渍化, 可见/近红外光谱, 黄河口区, 随机森林, 支持向量机

Abstract: 【Objective】It is crucial to quantitatively and accurately detect saline soil for the prevention of soil salinization and agricultural sustainable development. The objective of this study is to find the characteristic spectra of soil salt and its main ions, and build a quantitative analysis model of soil salinization which is suitable for the estuary area of Yellow River to improve the accuracy and stability of the quantitative analysis. 【Method】 Kenli County in Shandong Province was selected as the experimental area. Firstly, 96 representative soil samples were collected on October 5th-9th, 2014. After soil sample were dried, the contents of soil salt and ions were analyzed. Then the visible and near infrared reflectance hyperspectra of the soil samples were measured in the laboratory by ASD Fieldspec 3 spectrometer, smoothed and transformed to the first deviation. Secondly, the response spectra of salt and its main ions (Cl^-, Na⁺, Ca²⁺）were analyzed, first soil salinity and main ions content and the first derivative spectra of reflectance by band correlation analysis, according to the principle of correlation coefficient and significant, their sensitive bands were selected as the characteristic bands, feature bands with maximum correlation coefficient were chosen as the significant feature bands. Then the characteristic spectra which can represent soil salt and main ions (Cl^-, Na⁺, Ca²⁺) were analyzed synthetically using correlation analysis and identified. Finally, the methods of multiple linear regression (MLR), support vector machine (SVM) and random forest (RF) were used respectively to build quantitative analysis models of soil salinity and ions contents. 【Result】 The overall shape and trend of the spectra curves of soil salinity and major ions (Cl^-, Na⁺, Ca²⁺) content in study area were very similar. Soil salinity and major ions (Cl^-, Na⁺, Ca²⁺) spectra response regions were determined to be 1 320-1 495, 1 790-1 920, 2 120-2 290 nm. On account of the correlation between the first derivative of the reflectance and the soil salinity and its main ions content, the sensitive spectral regions were 1 490-1 520, 1 890-1 930 nm, final integrated spectra analysis and correlation analysis the characteristic bands were  1 493, 1 801, 1 911 and 2 289 nm and the significant characteristic bands were 1 493 and 1 911 nm. The models’ accuracy based on the first deviation of reflectance on the significant characteristic bands matched the models’ accuracy based on the four characteristic bands, which indicates that the significant characteristic spectra were effective for quantitative analysis of soil salt and its main ions. Compared the three modeling methods, the prediction ability of the RF was the best, followed by the SVM, the MLR models’ precision was the lowest. The models using the above-mentioned three methods could be used for quantitative analysis of salt, Cl^- and Na⁺, and had good stability and high precision, however the prediction accuracy of Ca²⁺ contents was still to be improved. In comprehensive comparison and analysis, among the built models, the RF models based on the first deviation of reflectance on the significant characteristic bands (1 493 and 1 911 nm) had higher accuracy and stability for the quantitative analysis of soil salt, Cl^- and Na⁺, and could be applied to the quantitative estimation of Ca²⁺.【Conclusion】The significant characteristic spectra (the first deviation of reflectance on 1 493 and 1 911 nm) of soil salt and its main ions were selected synthetically using correlation analysis based on the spectral response, then the quantitative estimation models of salt and its main ions were built using the RF regression method, which is suitable for the effective extraction of soil salinization information in the estuary area of Yellow River.

Key words: soil salinization, visible/near infrared spectroscopy, estuary area of Yellow River, random forest, support vector machine