Abstract:
Recognition and classification of Mars analog terrain aim to simulate and study the Mars environment by analyzing Mars analog terrain images, which holds significant research value for exploring scientific questions such as formation, evolution, and potential habitability of Mars. In response to the challenge of balancing classification performance and model lightweighting in current Mars terrain classification algorithms, a lightweight, rapid recognition and classification method for Mars analog terrain is proposed (LWNet). This algorithm constructs a dual-branch teacher-student network, employs knowledge distillation to reduce the number of parameters and computational load of the model, and integrates attention mechanism to enhance the capability of terrain classification and recognition, achieving high accuracy and lightweight classification models. To verify the classification performance of the proposed method, a dataset of Mars analog terrain on Earth was established, including four typical Mars landforms: cliff, desert, channel, and yardang, with each type of terrain consisting of 800 images. The dataset was employed to conduct rapid recognition and classification experiments with LWNet. The results indicate the overall accuracy reaches 97.81%, which only decreases by 1.25% compared with Swin-Transformer, while its Parameters and FLOPs are only 1.3% and 4.8% of Swin-Transformer, respectively. Experimental results verify the effectiveness and superiority of the LWNet.