Estimating millimeter-scale surface roughness of rock outcrops using drone-flyover structure-from-motion (SfM) photogrammetry by applying machine learning model

A quantitative outcrop survey was conducted in three coastal areas in Japan to estimate the relationship between the surface morphology and visual information of well-exposed rocks using photogrammetry of drone flyovers. We generated three-dimensional digital outcrop models in the study areas to produce the hue, saturation, value (HSV) color space images and digital elevation model (DEM) data, together with terrain ruggedness index (TRI) computed from the DEM data. Using the data, we examined whether our machine learning model (MLM) could predict the millimeter-scale surface ruggedness of the given rock outcrop. In the prediction, one of the three studied outcrops was selected for the training data, and various patterns of choices from available georeferenced visual information (i.e., coordinates, H, S, V) and TRI data in the other study areas were used as explanatory variables and response variables, respectively. The MLM with H, S, and V as explanatory variables using the 3σ method for outlier removal showed the smallest Root Mean Square Error of 0.51 × 10⁻³. The results revealed that our MLM provided reasonable quantitative predictions of surface ruggedness. Additionally, our predictions worked well even in the presence of cast shadows on the studied outcrops, suggesting that the shadow effects were likely negligible. Our findings emphasize that the HSV color space data acquired by drone-flyover photogrammetry alone can quantitatively predict the millimeter-scale surface ruggedness of outcrops, facilitating the acquisition of high-resolution surface morphology data without DEMs.