ABSTRACTLand cover and its change are one of the important factors of global environmental change. The new GaoFen-2 (GF-2) satellite provides abundant spectral features and texture information, and airborne light detection and ranging (lidar) provides accurate three-dimensional coordinates at a finer scale. Fusing these data has the potential to improve land-cover classification. In the article, we selected the Random Forest (RF) as a classifier. The spectral bands of GF-2, normalized difference vegetation index (NDVI), normalized digital surface model derived from lidar data, and their grey-level co-occurrence matrix (GLCM) textures including mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation were generated to create seven scenarios with different combination of RF input variables. We estimated the classification performance on GF-2 satellite data, compared and assessed the individual and combined contributions of GF-2 and lidar data with regard to classification accuracy using fusion data, and identified the optimum combination of input variables that best balances classification performance with the computational challenges for land-cover classification. The results showed that GF-2 multispectral data alone or the fusion of GF-2 data and other sources can provide relatively good classification mapping accuracy in complex urban environments. The classification accuracy of fused data from the GF-2 satellite and lidar data exceeded those of GF-2 or lidar data alone. Moreover, GLCM textures significantly improved the classification performance whether using GF-2 satellite data or lidar height data. A fusion of GF-2 multispectral data, NDVI, lidar data, and their texture features (102 RF variables) achieved the best classification accuracy in seven scenarios. Using the optimum set of variables (55 variables) for RF yielded the most satisfactory classification result, achieving a total accuracy and kappa coefficient of 94.51% and 0.93, respectively. The producer’s accuracy and user’s accuracy exceeded 90% for almost all classes. This study aims to provide a reference for the efficient improvement of land-cover classification and offer support for extending the applications of classification algorithms and data sources.