Abstract: Abstract: Impervious surfaces have been closely related to various environmental science, especially on the magnitude, position, spatial pattern, and perviousness-imperviousness ratio. The area of impervious surface has also rapidly expanded with the recent acceleration of urbanization. A rapid and accurate spatial distribution of urban impervious surfaces can provide crucial data for the urban ecological environment, rational planning, and regional sustainable development, particularly for the developing sponge cities, ecological and intelligent cities. As a result, remote sensing has received much attention in this field. In this research, an extraction workflow of urban impervious surface was proposed to treat the visible-light images from the Unmanned Aerial Vehicle (UAV) using the Object-Based Image Analysis (OBIA) and Random Forest (RF). First, the image was segmented into the homogenous objects (basic units for classification), according to the optimal segmentation scale determined by the ESP2 plug-in. The classification schemes (S1-S7) were established to sequentially introduce the four additional types of features (41 in total), including vegetation index, texture, geometry, and terrain. The different feature subsets were also constructed, according to the spectral features of objects. In scheme S8, the feature recursive elimination (RFE) was used to determine the optimal features subsets (13 in total). Then, the RF was applied to the S1-S8 for the optimum scheme. Finally, the classifications were carried out using RF, Support Vector Machine (SVM), and K-Nearest Neighbor (KNN), further to evaluate the performance using the feature subset of the best scheme. The results show that the UAV images with the ultra-high resolution were widely expected to serve as the finer ground object recognition. More importantly, the UAV images presented much more morphological and spatial features, compared with the previous satellite and aerial remote sensing images. The object-oriented image analysis provided more information about the objects from various features, compared with the spectral feature alone. All topographic, spectral, and vegetation index features dominated the classification accuracy, especially topographic features (nDSM). Specifically, the classification accuracies of S3-S7 after the introduction of nDSM were substantially improved(22.49-39.67 percentage points). The highest classification accuracy was achieved in the S8 using feature optimization subset, indicating an overall accuracy of 96.18%, and a Kappa coefficient of 0.95. The reason was that the feature optimization for the high-dimensional features resulted in a significant reduction in the number of features, particularly for the higher classification accuracy. Furthermore, the overall accuracy of RF increased by 1.35 and 14.18 percentage points, respectively, compared with the SVM and KNN, indicating better RF performance. Correspondingly, the object-oriented classification combined with the RF presented a higher accuracy, stronger anti-noise ability, and stable performance on the urban impervious surfaces, thereby effectively reducing the fragmentation of classification during extraction. To summarize, it is feasible to extract the urban impervious surface using UAV visible-light images, indicating the high extraction accuracy and the cost saving in the data acquisition. The finding can provide a strong reference to extract information about additional urban features from UAV visible light images, thereby promoting the application of consumer UAVs in urban remote sensing.