土壤水分在一定程度上决定着地区的生态承载力和土壤理化性质。 准确、 快速的获取土壤水分含量, 对生态环境监测、 土壤退化恢复等具有重要意义。 高光谱遥感在土壤参数反演方面应用广泛, 但对高寒草甸土壤的高光谱特征与参数反演研究有待深入。 为建立适用于高寒脆弱生态系统的高寒草甸土壤含水量高光谱反演模型, 以黄河源区的河南县为例, 利用多元线性逐步回归(MLSR)、 偏最小二乘回归(PLSR)和BP神经网络(BPNN)对土壤样本含水量与土壤光谱及其数学变换的特征波段进行建模, 由决定系数(R2)、 均方根误差(RMSE)和预测残差比(RPD)对模型精度进行验证。 结果表明: (1) 在可见光-近红外波段, 土壤样本的光谱反射率在710、 780和950 nm附近存在强弱不同的水分吸收区间, 且随着含水量的增加, 反射率呈现先迅速降低, 后缓慢增加的趋势。 (2) 通过连续投影算法(SPA)筛选的光谱特征波段作为自变量, 含水量作为因变量, 分别通过MLSR和PLSR建立反演模型, 其中一阶微分(FD)、 对数一阶微分(FDL)变换对应的PLSR模型可实现高寒草甸土壤水分的粗略反演, 且FD变换对应的PLSR模型精度较高。 (3)BPNN反演模型中, 除去包络(CR)建模外, 其余模型R2均大于0.9, RMSE在0.048~0.074之间。 其中FD、 FDL、 LG变换光谱对应的BPNN模型反演精度较高, 验证结果的R2均大于0.8, RPD均大于2.5, 精度最高的为对数(LG)变换对应的BPNN模型, R2、 RMSE和RPD分别达到0.967、 0.038和5.039。 因此, BPNN模型能较精确的实现黄河源区高寒草甸土壤含水量的高光谱反演, 为该地区乃至其他高寒地区生态环境监测与土壤恢复提供技术基础和数据支撑。

Soil moisture determines a region’s ecological carrying capacity and soil physical and chemical properties to a certain extent. It is significant to obtain soil moisture content accurately and quickly for ecological environment monitoring and soil degradation restoration. Hyperspectral remote sensing is widely used in soil parameter inversion, but the research on hyperspectral characteristics and parameter inversion of alpine meadow soil needs further study. Consequently, to develop a hyperspectral inversion model of soil moisture content in alpine meadows applicable to fragile alpine ecosystems, 102 soil samples were collected from Henan County in the Yellow River source area. Multiple linear stepwise regression (MLSR), partial least squares regression (PLSR) and back propagation neural network (BPNN) methods were used to model the soil moisture content with the original spectrum and its mathematically transformed characteristic bands, and the inversion accuracy was verified by the coefficient of determination (R2), root mean square error (RMSE) and the residual ratio of prediction (RPD). The major findings are as follows: (1) In the visible-near infrared band, the spectral reflectance of soil samples has water absorption interval near 710, 780 and 950 nm, and the absorption intensity is different. The reflectance tends to decrease rapidly and increase slowly with increasing soil moisture content. (2) SPA algorithm was used to select the spectrum’s characteristic bands after S-G smoothing, four transformations as independent variables and water content as dependent variables. Then MLSR and PLSR were used to establish the inversion model. The PLSR model corresponding to the first-order differential (FD) and first-order logarithmic differential (FDL) transformations can achieve a rough inversion of soil moisture in alpine meadows, and the PLSR model corresponding to the FD transformation is accurate. (3) In the BPNN inversion models, except for the model corresponding to continue to remove (CR), theR2 of other models is greater than 0.9, and RMSE is between 0.048 and 0.074. In all the models, the BPNN model corresponding to FD, FDL and LG transform is highly accurate, with R2 and RPD greater than 0.8 and 2.5 respectively. The BPNN model corresponding to the LG transform has the highest accuracy, with R2, RMSE and RPD up to 0.967, 0.038 and 5.039, respectively. Therefore, the BPNN model can achieve relatively accurate hyperspectral inversion of soil moisture content of alpine meadow in the source region of the Yellow River, which can provide the technical basis and data support for ecological environment monitoring and soil restoration in this region and even other alpine regions.