Regional heterogeneity of hyperspectral characteristics of salt-affected soil and salinity inversion

Abstract: The objectives of this study were to analyze regional heterogeneity of hyperspectral characteristics of salt-affected soils from Xinjiang Uygur Autonomous Region, Zhejiang and Jilin provinces and to establish hyperspectral inversion model of salinity for cross-regional salt-affected soils with high precision. One hundred and fifty-nine soil samples at 0-20 cm depth were taken from Xinjiang Uygur Autonomous Region (58 soil samples), Zhejiang (68 soil samples) and Jilin (33 soil samples) provinces, respectively. Electrical conductivity(1:5 soil to water, EC1:5) and spectral reflectance (SR) of all the 159 soil samples were determined. Regression models between EC1:5 and SR were fitted using principal component regression (PCR), multiple linear regression (MLR), and partial least squares regression (PLSR) based on local and global models, respectively. The prediction accuracies of these models were assessed by comparing determination coefficients (R2), relative percent deviation (RPD) and root mean squared error (RMSE) between predicted and measured EC1:5. Results showed that there were obvious differences not only in spectral reflectances but also in spectral curve shapes among the salt-affected soils from different regions. After a first derivative data processing, however, these differences were decreased. Values of R2, RMSE and RPD between the predicted and measured EC1:5 for the global model were 0.06, 0.93 and 1.03 for PCR equation, 0.10, 0.91 and 1.04 for MLR equation, 0.71, 0.51 and 1.85 for PCR equation, and 0.71, 0.51, 1.86 for PLSR equation, respectively. Values of R2, RMSE and RPD between the predicted and measured EC1:5 for the local model were 0.45, 0.73 and 1.30 for LR equation, 0.50, 0.69 and 1.38 for MLR equation, 0.76, 0.46 and 2.05 for PCR equation, and 0.78, 0.44, 2.15 for PLSR equation, respectively. The values of R2 and RPD between the predicted and measured EC1:5 were higher for local models than those of global models, but the values of RMSE of local models between the predicted and measured EC1:5 were lower than that of global models. This indicated that the local models were more accurate than the global models in predicting EC1:5 from soil spectral reflectances. In order to improve the prediction accuracy of global model, 25 data processing methods were carried out for soil spectral reflectances. It was shown that the sensitive bands of EC1:5 varied with study regions. Among all of the 25 data processing methods, the prediction accuracy of global model based on Savitzky Golay Second Derivative (SGSD) method decreased drastically compared with that based on the spectral reflectance method. Prediction accuracies of inversion models decreased slightly based on area normalization (AN), mean normalization (MN), unite vector normalization(UVN), maximum normalization (MAN), range normalization(RN), linear baseline correction(LBC), Savitzky Golay Smoothing (SCS) and multiplicative scatter correction (MSC). Six data processing methods including three-point savitzky golay first derivative (SGFD3), five-point cavitzky golay first derivative (SGFD5), seven-point savitzky golay first derivative (SGFD7), LBC+SGFD3, SGS+ SGFD3 and SGS+LBC+SGFD3 improved inversion accuracies of global models. The values of PRD were greater than 2.0 for inversion equations based on these 6 data processing methods. The inversion accuracy based on SGS+SGFD3 data processing method was best with the R2, RMSE and RPD of 0.80, 0.43, and 2.23, respectively. The study can provide valuble information for aerospace hyperspectral remote sensing of cross-regional soil salinization.