Abstract:It is significant to take best agricultural measures and improve salinization to rapidly and accurately determinate the composition and content of soil salt. The hyperspectral integrated inversion model based on transformation of hyperspectral, characteristic bands, characteristic spectral indices screening and support vector machine (SVM) was established to improve the accuracy of watersoluble salt ions content by taking the saline soil of Yongji irrigation area of Hetao Irrigation District. The results showed that the correlation between the original spectral reflectance by pretreatment and watersoluble salt ions content was relatively low and the maximum correlation coefficient was 0.18, and the sequence of them from big to small was as follows: Ca2+, SO2-4, Mg2+, the content of salt, Na++K+ and Cl-. The optional transformation forms of salt content, Na++K+, Cl-, SO2-4, Ca2+ and Mg2+ were (1/R)″, (1/R)″, (lnR)′, (lnR)″, R′ and (lnR)″, respectively. The numbers of sensitive bands (P<0.01) were 41, 7, 9, 65, 76 and 28, respectively. Stepwise regression method was used to filtrate the characteristic bands from sensitive bands, and the average of determination coefficient (R2) and the average of root mean square error (RMSE) of each ion in the regression model based on the characteristic band were 0.35 and 0.87g/kg, of which R2 was the largest and the smallest were SO2-4 (0.52) and Ca2+ (0.20), respectively. Combined with the stepwise regression method, the characteristic bands were substituted into the spectral index to determine that there were three characteristic spectral indices for Mg2+, there were two characteristic spectral indices for salt content, and there were one characteristic spectral index for Na++K+, SO2-4 and Ca2+, respectively. The R2 of model for watersoluble salt ions content based on the characteristic bands and characteristic spectral indices was increased by 5867%, and the RMSE was decreased by 2460%, of which the maximum R2 was SO2-4 (0.74), RMSE was 0.47g/kg. 〖JP〗The model of SVM based on the characteristic bands and characteristic spectral indices combined had a significant improvement in the prediction than that merely based on the characteristic bands, for example, the average relative analysis error (RPD) was increased by 110.27%, the R2 was increased by 37.54% and the RMSE was decreased by 4012% in the training set, the R2 was increased by 5604% and the RMSE was decreased by 3939% in the verification set. The results showed that the RPD of SO2-4 reached 3000, which showed a good prediction ability. The model of salt content and Mg2+ had good quantitative prediction ability which can be used for assessment or correlation prediction, respectively. The SVM models of Na++K+ and Ca2+ had the ability to distinguish between high and low values.
