【摘要】 干旱地区土壤水分是影响气候动态变化、植被生态恢复和土地荒漠化治理的重要指示因子。本研究采用Landsat8 OLI/TIRS多光谱遥感影像，在9个传统光谱指数基础上引入热红外波段(b₁₀)进行改进，通过显著性检验和多重共线性检验后的优选光谱指数作为本研究的建模因子，并结合地形数据采用多元线性回归（multivariable linear regression, MLR）和随机森林（random forest, RF）算法构建荒漠土壤水分综合反演模型，选取最优模型分析土壤水分空间分布特征及驱动因素，结果表明：（1）改进后，光谱指数EBSI,ECI,ECal, ENDVI和EPDI相关系数提升了0.02～0.11;（2）光谱指数经改进后，线性和非线性模型预测集R²分别提升了0.12和0.05,相对分析误差提升了0.35和0.49,其中，RF-II模型的相对分析误差高达3.12,能精准地对土壤水分进行预测；（3）非线性模型的精度明显优于线性模型，MLR线性模型预测集的R²仅为0.59和0.71,而RF非线性模型预测集的R²达到0.86和0.91;（4）土壤水分分布受到自然、人为2种驱动因素影响，东北部沙漠呈现[0,5)%和[5,12)%,南部农田交错分布，北部及中部荒漠-绿洲过渡带受植被覆盖程度和地表盐结皮抑制土壤水分蒸散困难，多呈现[15,20)%和[20,40)%。更多还原

【Abstract】 Soil moisture is an important indicator affecting dynamic climate changes, vegetation ecological recovery, and land desertification control in arid regions. Using Landsat8 OLI/TIRS multispectral remote sensing images, this study determined the optimal spectral indices by introducing thermal infrared(b10) band to improve nine traditional spectral indices and through significance tests and multiple covariance tests. Then, with the improved spectral indices as the modeling factors and based on the terrain data, this study constructed multispectral comprehensive inversion models of desert soil moisture using the multivariate linear regression(MLR) and random forest(RF) algorithms. Finally, the spatial distribution characteristics of soil moisture and their driving factors were analyzed using the optimal model. The results are as follows:(1) The correlation coefficients of the improved spectral indices EBSI, ECI, ECal, ENDVI, and EPDI increased by 0.02～0.11;(2) For the prediction datasets of linear and non-linear models, their R~2 increased by 0.12 and 0.05, respectively and their RPD values increased by 0.35 and 0.49, respectively after the spectral indices were improved. Moreover, the RPD value of model RF-II was up to 3.12, and thus this model can accurately predict soil moisture.(3) The accuracy of the non-linear models was significantly higher than that of the linear models. The R~2 of the prediction datasets of MLR-based linear models was only 0.59 and 0.71, while that of the RF-based non-linear models reached 0.86 and 0.91.(4) The distribution of soil moisture was influenced by both natural and artificial factors. The soil moisture content is [0, 5)% and [5, 12)% in the northeastern desert and shows cross-distribution in the southern farmland. Soil moisture is difficult to evaporate in the northern and central desert-oasis transition zones due to inhibiting factors such as the vegetation coverage and surface salt crust, with the content of [15, 20)% and [20, 40)% mostly.更多还原