AE-TransUNet+: An Enhanced Hybrid Transformer Network for Detection of Lunar South Small Craters in LRO NAC Images.

Impact craters are the most significant topographic features on the lunar surface. They are also essential factors affecting the construction of lunar bases and activities on the lunar surface in the future. Nevertheless, it is difficult to extract the impact craters by image processing methods due to the influence of illumination in the lunar polar region. Furthermore, the digital elevation model (DEM) is unable to extract small impact craters. Therefore, we propose an extraction framework for small craters, namely AE-TransUNet+, for the future lunar South Pole region of interest. The algorithm is applied to grayscale images (0.5 m/pixel) captured by the Lunar Reconnaissance Orbital (LRO) narrow-angle camera (NAC). To improve the channel extract and spatial relationship between features, the convolutional block attention modules (CBAMs) and depthwise-separable convolution (DSC) are added to the core TransUNet model, and the redesigned jump connection enhancement module focuses more on small impact crater features. Experiments on the NAC dataset show that AE-TransUNet+ has a higher F1-Score (0.743) and IOU (0.847) compared to other crater detection models. Besides, based on transfer learning, the visualization results of Mars and Mercury show that the proposed method is more efficient in the processing of remote sensing data for different deep space explorations providing better mobility. [ABSTRACT FROM AUTHOR]