Lunar impact crater extraction using an improved U-Net model assisted by pyramid random cropping strategy

Lunar impact craters are one of the important topographic features on the lunar surface, and they are critical for future scientific exploration of the Moon. The existing impact crater identification methods are affected by various factors, including prior rules and the complexity of impact craters, which often have large errors and low identification accuracies, particularly for small or overlapping impact craters. This research provides a method to automatically extract lunar impact craters with high accuracy using an improved U-Net model, which is aided by the pyramid random cropping strategy (PRCS). First, the lunar digital elevation model (DEM) images are cropped using PRCS. The improved U-Net predicts crater edges, followed by template matching filtration and DEM coordinate transformation to output crater locations (lat/lon) and sizes. The model is based on the U-Net architecture, which employs depthwise separable convolution and also adds the self-attention mechanism. Experiments show that with metrics of 77.17% recall, 72.49% precision, and 76.19% F2-score, our model shows optimal performance for recall, F2-score, and extreme-scale craters detection (20 km). The method we proposed can be used to extract impact craters more correctly and automatically, as well as detect overlapping impact craters and missed smaller or larger impact craters.