Lunar crater identification based on SA-BRU-Net

Identification of crater impacts is crucial during lunar exploration. Previous methods primarily rely on visual interpretation, which is time-consuming and lacks objective standards. Moreover, these methods often fail to accurately identify small craters. To address these issues, we propose a novel convolutional neural network, SA-BRU-Net (Spatial Attentive Bottleneck Residual U-Net). Building upon U-Net, SA-BRU-Net introduces spatial attention mechanisms and employs bottleneck residual blocks instead of traditional residual blocks. These improvements reduce network parameters while enhancing the identification accuracy of lunar craters. Experimental results demonstrate that our model achieves a recall rate of 81.38% and a precision of 79.03% on DEM images, outperforming other network models.