9:15 AM - 9:30 AM
[SCG51-02] Applying a machine learning model to quantify the surface ruggedness using high-resolution 3D outcrop model
Keywords:Machine Learning, Outcrop, Surface Ruggeddness, HSV, 3D digital outcrop model
A recent 3D digitization technology has been advanced developing 3D digital outcrop models. Yet, few studies report the effort in extracting the physical properties of rocks from 3D digital outcrop models. The 3D digital outcrop models constructed by the UAV aerial photography provide visual information, including the color and surface morphology of the outcrops. Surface morphology is traditionally used to study surface ruggedness in the field of geomorphology. On a large spatial scale, Yamada et al. (2013) found that localized heterogeneity in the shape, including ruggedness and thickness of underlying faults, controls fault slip using 3D reflection seismic data. On a small scale, numerous studies in metallurgy have pointed out that the type of metal, heat treatment, and degree of processing determine the ruggedness of the fracture surface. However, few works have studied the medium-spatial scale ruggedness such as that observed on geological outcrops, hampering to understand the linkage between the surface ruggedness and rock properties of the outcrop.
We aim to investigate the relationship between the surface morphology and the physical properties of the rocks at outcrops. First, we created high-resolution DEM images with different spatial resolutions from 3D high-resolution digital outcrop models, and calculated Terrain Ruggedness Index (TRI), Roughness, and Topographic Position Index (TPI). The 3D digital outcrop models were generated by the drone flyover in the Nogita Coast of Itoshima Peninsula and the Unosaki Coast of Oga Peninsula, Japan. Second, we produced orthoimages using the 3D digital outcrop models and extracted them to the data in HSV color space. We examined the correlation between H, S, and V color spaces and a ruggedness index GTRI, newly defined in this study. Third, we examined the use of a machine learning model to predict the ruggedness of the orthoimage in the Unosaki coast, with the Nogita data, HSV color space data, and GTRI data as the supervisory data, explanatory variables, and the objective function, respectively.
We found that Roughness and TRI were more affected by the general surface slope of the outcrop surface than the decrease in spatial resolution, while TPI was not affected by the change in resolution because it is a method for evaluating the surface slope. The most effective index representing the ruggedness was TRI. The values of the new ruggedness parameter GTRI showed good correlations with those in the HSV color space. Among the good correlations, GTRI showed the best correlation with V, suggesting a correlation between the outcrop ruggedness and the V component in the orthoimage. We further found that our machine learning model generally reconstructed well the GTRI image of the Unosaki coast, with the successful prediction of the presence of relatively large-scale ruggedness such as cracks. The results suggest that our method by integrating the TRI parameter and HSV color space acquired from UAV photography can be a powerful method to estimate rock properties.
We aim to investigate the relationship between the surface morphology and the physical properties of the rocks at outcrops. First, we created high-resolution DEM images with different spatial resolutions from 3D high-resolution digital outcrop models, and calculated Terrain Ruggedness Index (TRI), Roughness, and Topographic Position Index (TPI). The 3D digital outcrop models were generated by the drone flyover in the Nogita Coast of Itoshima Peninsula and the Unosaki Coast of Oga Peninsula, Japan. Second, we produced orthoimages using the 3D digital outcrop models and extracted them to the data in HSV color space. We examined the correlation between H, S, and V color spaces and a ruggedness index GTRI, newly defined in this study. Third, we examined the use of a machine learning model to predict the ruggedness of the orthoimage in the Unosaki coast, with the Nogita data, HSV color space data, and GTRI data as the supervisory data, explanatory variables, and the objective function, respectively.
We found that Roughness and TRI were more affected by the general surface slope of the outcrop surface than the decrease in spatial resolution, while TPI was not affected by the change in resolution because it is a method for evaluating the surface slope. The most effective index representing the ruggedness was TRI. The values of the new ruggedness parameter GTRI showed good correlations with those in the HSV color space. Among the good correlations, GTRI showed the best correlation with V, suggesting a correlation between the outcrop ruggedness and the V component in the orthoimage. We further found that our machine learning model generally reconstructed well the GTRI image of the Unosaki coast, with the successful prediction of the presence of relatively large-scale ruggedness such as cracks. The results suggest that our method by integrating the TRI parameter and HSV color space acquired from UAV photography can be a powerful method to estimate rock properties.