Abstract: Benggang as a fragmented landform type, can be characterized by a deep-cut slope with various shapes and depressions on the vast weathered crust slopes in southern China. Normally, the gully heads have been continuously collapsed to form a typical landform, such as chair-like erosion. Benggang usually develops rapidly due to large amount of erosion, and thereby to threaten land resources and ecological environment. The identification of Benggang development become necessary to control erosion, and then clarify the behind mechanism. However, conventional methods have low levels of automation for a large-scale work, particularly on local inquiry, in-situ search, and manual interpretation for high-resolution images from satellite remote sensing. This paper proposes a novel Bag of Visual-Topographic Words (BoV-TW) model combining high resolution Digital Orthophoto Map (DOM) and Digital Surface Model (DSM) local features to represent Benggang areas, according to the classification and recognition methods in remote sensing images. The local features of DOM can be set in Harris-Affine and Maximally Stable Extremal Regions (MSER), with Scale-Invariant Feature Transform (SIFT) descriptors. The local features of DSM were extracted by a 3D Douglas-Peucker algorithm, representing by gradient direction-invariant descriptors developed in this study. A Latent Dirichlet Allocation (LDA) was used to balance latent semantic analysis, thereby to construct low-dimensional high-level semantic representations. Support Vector Machine (SVM) was used as a supervised learning training classifier to achieve high-precision and fast automatic Benggang recognition. Three Benggang areas in Tongcheng County, Hubei Province were selected as the experimental objects. The original data were collected by DJI Phantom 3 Pro micro UAV in September 2016. Photoscan processing was used to obtain the DOM and DSM of three Benggang areas. Uniform grid division can be used to achieve DOM and DSM tiles in 0.15 m resolution of 256 × 256 pixels. A high-resolution 3D Benggang model was used to mark the areas with/without Benggang. Two Benggang areas were taken as the training set, and the rest were taken as the test set, in three comparative experiments. The results show that: 1) With changing numbers of LDA topics, the proposed method can maintain a total accuracy of 95%, while the recall rate and precision of Benggang 80%, indicating the maximums were 97.22% and 94.44%, respectively. The total accuracy, recall and precision have increased by 12%, 11% and 32%, respectively, compared with that of only DOM features. The recognition performance was significantly improved after combining with DSM features. 2) With changing sizes of the BoV-TW vocabulary, the total accuracy was 90%, and the maximum was 96.10%, while the recall rate reached 90%, the maximum recall rate of 100%. Precision gradually increased as the increase of vocabulary size, with the maximum of 85.00%. Compared with that of only BoV-TW, three performance indicators increased by 13%, 12% and 30%, respectively, indicating that LDA for latent semantic analysis can greatly improve the recognition detection performance. 3) The proposed method was better than that on the ResNet50 network using DOM and DOM + DSM as the data source. It infers that Benggang as a landform type cannot be well recognized using only DOM, or the simple combination of DOM and DSM. Meanwhile, it needs to combine feature extraction and fusion strategy for efficient application. This finding can deliver a useful tool for the quantitative analysis on the identification, control and erosion mechanism for Benggang.