Automatic morphologic classification of Martian craters using imbalanced datasets of Tianwen-1’s MoRIC images with deep neural networks

Martian impact craters contain prolific geomorphic information; their types provide vital indicators reflecting the geological evolution and chronology of celestial bodies. Currently, visual inspection is the most commonly used and widely accepted approach for identifying the crater types and their morphological features. However, it is a subjective task and requires professional knowledge. In this work, we present a method for the automatic morphologic classification of Martian craters using imbalanced Mars image datasets. Specifically, we classified and cropped Tianwen-1’s MoRIC images according to the morphological features of craters based on Robbins’ impact crater catalog; a dataset of six different types of Mars impact craters was established using the images. Based on this dataset, we classified the Mars impact craters using three popular neural network models with CNN and Transformer architectures. Meanwhile, to address the imbalanced samples in the network training process, a common problem in planetary remote sensing datasets, we introduce two methods (i.e., label smoothing strategy and weighted loss function) to suppress its influence on classification accuracy. Experimental results show that the Vision Transformer (ViT) model has the highest classification accuracy, reaching 90.3%. The label smoothing strategy performs well in CNN approaches, among which VGGNet11 improves the accuracy by 2.2%. By contrast, the weighted loss function performs well in ViT, and the classification accuracy of ViT improves by 1.3%. These results demonstrate a promising future for applying deep neural networks to identify and morphologically analyze Martian craters automatically.