Abstract: Geological bodies with high genus are complex structures that feature a variety of cavities, such as the interbedded strata and fracture zones at dam foundation. In handling these structures, the common geological modeling methods based on surface reconstruction have poor performance in automation and low accuracy, while methods using volume element representation can realize automatic modeling, but at a cost of redundant voxel data. Besides, previous algorithms can hardly suit the unevenly distributed borehole data or the multivariate shapes of high genus strata in hydropower engineering. This paper develops an automatic geological voxel modeling method based on the K-means-modified extremely randomized trees (Kmeans-ERT). First, to classify the ambiguous and complex high genus strata, ERT is selected as the base prediction model because of its robustness. Then, the K-means algorithm is adopted to modify ERT by adding a clustering analysis progress at each node to calculate dynamically the distribution of random split values. Moreover, a boundary recognition algorithm is constructed to optimize the model by hiding interior voxels. Engineering application shows our new model can automatically reconstruct high genus strata. Compared to SVC, KNN, random forest, deep forest, and BP neural network, the model improves the average accuracy by 17.4%, 19.1%, 4.7%, 6.5% and 17.1% respectively, and it sees a 69.3% decrease in memory cost. This verifies our new method has accuracy and efficiency superior to manual geological modeling or other automatic algorithms.

Key words: hydropower engineering, 3D geological modeling, implicit modeling, high genus geological structure, K-means, extremely randomized trees