•A non-destructive pore characterization of coal in different scales during liquid nitrogen cold shocks.•Microscopic pore structure test and analysis by nuclear magnetic resonance technology.•Evolution of mesoscopic pore-fracture network by steromicroscope imaging and quantitative analysis.•Nondestructive characterization of macroscopic damage by ultrasonic wave velocities and dynamic parameters.•Combination of macroscopic, mesoscopic and microscopic pore structure characteristics. Liquid nitrogen cyclic cold shock is a new waterless fracturing method suitable for coal seams in arid areas. The pore structure of coal after cold shock plays a key role in gas drainage, but there is currently a lack of study combining macroscopic-mesoscopic-microscopic scale. Multiscale pore characterization will be helpful to clarify the damage mechanism of liquid nitrogen cold shock to coal. In this paper, the same coal samples were tested for ultrasonic inspection, stereomicroscopic imaging and nuclear magnetic resonance (NMR) after different cold shock times. The macroscopic damage were studied by ultrasonic wave; the changes of mesoscopic pores and fractures (0.1 ~ 10 mm) were studied by stereoscopic microimaging and digital image processing; the changes of microscopic pores (2 nm ~ 0.1 mm) were studied by NMR. Finally, the macroscopic - mesoscopic - microscopic pore characteristic parameters were coupled. The main results are as follows: With the increase of cold shock times, the P/S wave velocities decreased, the dynamic parameters decreased linearly, and the macroscopic damage of coal samples was obvious. The number of mesoscopic pores, surface porosity, and probability entropy increased. As can be seen from the T2 spectra of coal samples, the microscopic pores expanded and increased with cold shock times. The effective porosity increased, among which the increase of micropores contributed the most. The macroscopic wave velocity and dynamic parameters of coal samples were negatively correlated with porosities. The calculation models of T2 cutoff value were established by multiple linear regression analyses, which were in good agreement with the experimental values.