While large-scale rock strata affected by underground coal mining have been widely studied through numerical modeling, there are still some aspects that can be better understood. Specifically, researchers do not fully utilize borehole logs and test results of rock specimens to reveal rock mass property variation along horizontal directions and strata lateral thickness variation at bed level. In this paper, we address this knowledge gap by proposing a data-intensive numerical modeling method (DINMM) that can make full use of these data with consideration of the modeling limitations of BlockRanger, a grid generation tool. Both the proposed method and the conventional numerical modeling method (CNMM) are applied to the Ying-Pan-Hao coal mine via the FLAC3D (Fast Lagrangian Analysis of a Continua in 3 Dimensions) package, and their predictions are then calibrated and compared to discuss the validity. Results show that, compared to the CNMM-based predictions, the root mean square error of 70 monitoring points is decreased at least by 27.4% in the DINMM-based prediction, and the relative error of maximum subsidence is reduced by 5.1% with a reduction rate of 66.5% on average, even though the CNMM-based model was originally better calibrated. We also find that a DINMM-based model is more in line with field observations and theoretical understanding in terms of displacement, stress, and failure propagation. The notion of data-intensive modeling seems to be quite promising and the DINMM should be useful for a better understanding of strata movement and subsidence prediction. Highlights • We propose a data-intensive modeling method (DINMM) to build numerical models based on multiple borehole logs, rather than on a single or generalized borehole log. • We realized rock mass property variation along horizontal directions and strata lateral thickness variation at bed level when modeling large-scale rock strata. • We find that the root mean square error of 70 monitoring points is decreased at least by 27.4% in the DINMM-based prediction, even though the control model was originally better calibrated.