基于环境变量的渭干河-库车河绿洲土壤盐分空间分布

Spatial distribution of soil salinity in Ugan-Kuqa River delta oasis based on environmental variables

土壤属性的数字制图对精准农业生产和环境保护治理至关重要。为了在大尺度上尽可能精确的监测土壤盐分空间变异性,该文使用普通克里格(ordinary kriging,OK)、地理加权回归(geographically weighted regression,GWR)和随机森林(random forest,RF)方法,结合地形、土壤理化性质和遥感影像数据等16个环境辅助变量,绘制渭干河-库车河绿洲表层土壤盐分分布图。基于决定系数(R^2)、均方根误差(RMSE)和平均绝对误差(MAE)验证模型精度。结果表明:不同方法预测的盐分分布趋势没有显著差异,大体上从研究区的西北向东南部方向增加;结合辅助变量的不同预测方法中,RF方法预测精度最高,R^2为0.74,RMSE和MAE分别为9.07和7.90 mS/cm,说明该模型可以有效地对区域尺度的土壤盐分进行定量估算;RF方法对电导率(electric conductivity,EC)低于2 mS/cm时预测精度最高,RMSE为3.96 mS/cm,很好的削弱了植被覆盖对电导率EC的影响。

Digital soil mapping(DSM) is the creation and population of spatial soil information systems by numerical models inferring the spatial and temporal variations of soil types and soil properties from soil observations and knowledge and from related environmental variables. DSM is critical to precise agricultural production and environmental protection. Accurately mapping soil salinity through remote sensing techniques has been an active research area in the past few decades particularly for agricultural lands. A total of 73 cropland topsoil samples(0-10 cm) were collected from Ugan-Kuqa River Delta Oasis, southern parts of Xinjiang Uyghur Autonomous Region of China for the measurement of soil electrical conductivity(EC) based on 1:5 soil-water extraction solution. Three spatial prediction models, i.e., ordinary kriging(OK), geographically weighted regression(GWR) and random forest(RF) methods were employed for digital mapping of soil salinity. Multi-source remote sensing data were resampled in the spatial resolution of 30 m and calculated various derived environmental variables, such as terrain attributes, soil physiochemical properties, and spectral indices. We selected 16 most sensitive variables to calibrate the estimation models based on the correlation analysis. Finally, the validation results of different models were compared under different intervals of EC and vegetation coverage. The mean absolute prediction error(MAE), root mean square error(RMSE) and coefficient of determination(R^2) were used to evaluate and compare the performance of the above methods. The spatial distribution patterns of EC obtained by different methods were quite similar, in general the distribution of salt increased from northwest to southeast of the study area, salt soil and severe salinity soil were concentrated in the southeast of the region. Among the different prediction methods combined with the variables, the OK method lacked a detailed description of the spatial variation of the EC content, and the internal map fragmentation of the GWR method made the details of the drawing effect more abundant. For the RF method the RMSE and MAE of both datasets were lower than OK and GWR method, R^2, RMSE and MAE were 0.74, 9.07 and 7.90 mS/cm, could effectively estimate the soil salinity at the regional scale. From the segmentation statistics of EC, the error of the RF method in the low and high values was small. The RF method had the highest prediction accuracy of 3.96 mS/cm for the EC of 0-2 mS/cm, which weakens the influence of vegetation cover on EC. Both the OK and the GWR methods had the largest prediction error between 0.1 and 0.2 of NDVI, but the RF method had little change in RMSE under different vegetation coverage. The best predicting model in these methods was selected based on corresponding performance and accuracy measures. The effect of GWR and RF modeling by nonlinear regression was obviously better than that of OK method. The local variation information of EC content was described in more detail. This study could provide a basis for the next step in the promotion of salinization monitoring in arid or semi-arid areas, selecting more effective environmental synergy variables, and improving the accuracy of soil mapping digital mapping.