Activity rhythms of laboratory rodents are usually measured by running wheels, and although wheel running activity-or-rest data enable straightforward rhythmic analyses, it provides limited behavioral information. In subterranean rodents (tuco-tucos), we used bio-loggers (accelerometers) to measure activity rhythms in both lab and field conditions, detecting diverse movements that compose activity. However, understanding these different accelerometer-detected activity components requires more complex analytical tools. Here we used supervised hidden Markov models (HMMs) as a machine learning analysis, to identify behavioral patterns in accelerometer data of tuco-tucos from field enclosures and characterize their behavioral rhythms in this condition. Activity of tuco-tucos was previously video-recorded in the laboratory with simultaneous accelerometer measurements. Video-obtained behavioral data were used in HMM models to refine (train) the classification of accelerometer recordings into different behavioral states. The classification obtained by HMM matched in 93% the one obtained by the video-observed method. Trained models were then used to automatically extract behavior information from accelerometers attached to 20 unobserved tuco-tucos first maintained in field enclosures and then transferred to the laboratory. Activity bouts associated with digging and locomotion were responsible for the diurnal rhythm in field enclosures and the nocturnal rhythm in the laboratory. Bouts of activity spread throughout day and night (cathemeral) were present in both conditions and were associated with feeding, coprophagy, and grooming. Finally, while rest occurs throughout day and night in the laboratory setting, tuco-tucos restrict rest episodes to nighttime under field enclosures, possibly as a behavioral adjustment to challenging environments. HMM models provide more behavioral information from accelerometry data, expanding the scope of activity pattern studies in small mammals under natural conditions.