The geomorphological characterization of Mars is complex. Therefore， to ensure the safe driving of rover， it is essential to understand the surface state around the rover through images captured by on-board digital cameras. The images are first preprocessed using stereo vision to create an aerial view， which is then divided into equal-sized blocks. Next， calibration and prediction datasets are created， containing 315 and 135 datapoints， respectively. Based on these datasets， a neural network model is developed. Finally， the image is classified using the resulting classification model to identify the region of interest. The results of the classification show that its accuracy on the calibration and prediction datasets using ResNet50 is 75.56% and 81.48%， respectively. This method can help researchers characterize the surface types around UGVs and identify the regions of interest that may provide more valuable information from the images. It can also be used for traversability prediction， risk assessment， and automatic path planning.