Abstract: In order to monitor typical soil salt content (sandy loam soil) in South Xinjiang,China quickly and effectively,and to improve the precision of the soil salt content estimation model by removing the noise of soil hyperspectral absorbance,this paper focused on the inversion relationship between soil spectrum and electrical conductivity (EC) by using multiple spectral pretreatment methods,and then the multiple linear regression (MLR) and the partial least squares regression (PLSR) modelling were applied to establish the salt content model based on the hyperspectral analysis technique.The effective and predictive capacities of different models were validated.This study took the typical arid area in South Xinjiang as the research object,obtained the hyperspectral data and EC by using Near-infrared spectrometer (Zolix Gaia Sorter) and conductivity meter (DDS-307),142 soil samples at 0~20 cm depth were collected and these samples were highly representative for the EC.Seven pretreatment methods were used to pretreat the original spectral data,such as vector normalization (VN),multiplicative scatter correction (MSC),standard normal variate (SNV),moving-average (MA) smoothing,Savitzky-Golay (SG) smoothing,first derivative (1-Der) and second derivative (2-Der),then the characteristic wavelengths were extracted with stepwise multiple regression (SMR) and successive projection algorithm (SPA),which were used as input variables of MLR and PLSR modeling.The results showed that the optimum pretreatment methods were VN,MSC,SNV and 1-Der.According to different pretreatments,in the stepwise multiple linear regression (SMLR) prediction model R_val² was greater than 0.95,RPD was greater than 6.2,and RMSEP was less than 0.44.In the multiple linear regression model based on the successive projections algorithm (SPA-MLR) R_val² was greater than 0.96,RPD was greater than 6.9,and RMSEP was less than 0.36,which were better than those in SMLR. In the partial least squares regression prediction model R_val² was greater than 0.88,RPD was greater than 4.4,RMSEP was less than 0.62 and in the partial least squares regression model based on the successive projections algorithm (SPA-PLSR) R_val² was greater than 0.59,RPD was greater than 2.4,and RMSEP was less than 1.1 which were less than those in SMLR and SPA-MLR. The best prediction of SMLR and SPA-MLR after vector normalization (VN) was as the follows:RMSEP=0.287 6,R_val²=0.979 2,RPD=9.907 8 and RMSEP=0.278 3,R_val²=0.980 5,RPD=11.50 which had fewer characteristic wavelengths.So the VN is a best effective pretreatment method.It is more suitable to establish the prediction model of soil conductivity in jujube tree by MLR with PLSR.It could be an important part in future researches how to choose the right algorithm to analyze the soil salt content with the spectral data.

Key words: soil electrical conductivity, multiple linear regression, partial least squares regression, hyperspectral