Abstract:

As one of the important environmental problems in the global arid region, soil salinization not only restricts agricultural production, but also causes serious land degradation. This problem has attracted wide attention from many scholars at home and abroad. In the recent stage, the monitoring and evaluation of soil salinization is mainly based on the use of different remote sensing data combined with different model inversion methods, and few research was seen on the integration of ground measured spectra and satellite remotely sensed data to obtain soil salinization information. Therefore, this study explores the effect of different satellite broadband simulations on soil conductivity and reflects soil salinization. In this study, the Ebinur Lake Wetland Nature Reserve in Xinjiang Uygur Autonomous Region, China was taken as an example.40 soil samples were collected in the field, and the soil conductivity (EC) and visible-near-infrared (VIS-NIR) spectral data were measured. The broadband data of the Landsat 8 OLI, Sentinel 2, and Sentinel 3 satellites were simulated by spectral response techniques. The three-dimensional spectral index was constructed by simulating wide-band data and its six spectral forms after five pretreatments: first-order differential (FD),second-order differential (SD),continuum removal (CR),absorbance conversion (ABS),and multivariate scattering correction (MSC).Three satellite soil EC estimation models were established by combining the gradient boosting regression tree (GBRT) algorithm, and the accuracy of the model estimation was compared after adding the three-dimensional spectral index. Then the soil estimation potential of the three satellites was analyzed. The results show that: Landsat 8 OLI, Sentinel 2, and Sentinel 3 have similar spectral trends under different soil EC conditions. With the increase of soil EC, the reflectance of each band of the three satellites is correspondingly increased, and both have high reflectance in red and near infrared. The correlation between TDSI and soil EC was basically above 0.4, and Sentinel 3 had the highest correlation based on SD pretreatment, which was-0.72. TDSI retains the information of red, green, blue, near-infrared and short-wave infrared bands sensitive to soil EC, which indicates that TDSI increases the effective information of soil EC, while taking into account the remote sensing mechanism, which has certain scientific significance. GBRT algorithm has outstanding performance for soil EC, estimation. Landsat 8 OLI, Sentinel 2, Sentinel 3 based on the estimation model of broadband simulation data produced the best prediction accuracy R² as 0.831, 0.847 and 0.903 respectively. After adding the TDSI, the estimation accuracy of different spectrum forms of the three satellites has been significantly improved. The corresponding prediction accuracy R² of the three satellites was increased to 0.83 5,0.857 and 0.935,respectively.The comprehensive analysis found that Sentinel 3 had the best estimation effect for soil EC(R²=0.935,RMSE=2.986mS·cm^-1, AIC=57.500),followed by Sentinel 2(R²=0.857,RMSE=4.596 mS·cm^-1, AIC=45.247),and finally Landsat 8 OLI (R²=0.835,RMSE=4.348 mS·cm^-1, AIC=32.765).In this study, the spectral response technique combined with GBRT algorithm has been verified that Landsat 8 OLI, Sentinel 2 and Sentinel 3 satellites have a good estimation effect on soil EC in arid area. TDSI can deeply mine the synergy information between bands and improve the prediction accuracy of the model. The combination of the two methods could provide favorable guidance for quantitative monitoring and prevention of soil salinization in arid areas.

Key words: spectral response, soil conductivity, spectral pretreatment, three-dimensional spectral index, gradient boosting regression tree