Fig. 2: General rockfall diameter distribution and details of Fig. 1.

The histogram shows the general distribution of estimated rockfall diameters. Estimated rockfall diameters are based on the output of the neural network and are potentially biased for smaller diameters (section indicated in white), as the CNN’s detection accuracy is reduced when detecting small features. Insets a–h are details from Fig. 1 and represent graben (a), pyroclastic vents (b), craters (c, f, h), regular slopes (d), and recent impacts (e, g). Some craters show a heterogeneous distribution of rockfall abundance, such as c, f, h. The peak count of spatial rockfall density is shown in the insets per 2° by 2° quadrangle. In e, g, we identified a series of rockfalls that are directly induced by small-scale impact events (impact-ejected); for details refer to Supplementary Fig. 2. ^aApollo 17 landing site, ^bMorse et al.³⁰ proposed an impact trajectory for f based on the distribution of the ejecta blanket (axis NW–SE), which agrees with the rockfall distribution (white arrow). Global WAC mosaic in the background⁴⁰.