Continuous observations of the central Arctic boundary layer (ABL) remain scarce, and ABL structure is strongly governed by stability, clouds, and wind shear. Using the year-round MOSAiC record, a condition-adaptive, multi-source neural network is developed to retrieve atmospheric boundary layer height (ABLH) from microwave radiometer, infrared hyperspectral radiances, ceilometer backscatter, wind profiles. Overall RMSE of 42.4 m is achieved, with best performance under clear skies and weakest under cloudy conditions. Feature-attribution and ablation analyses indicate wind-field information provides the strongest ABLH constraint. Clear-sky retrievals are improved by infrared observations, but cloud-related interference is introduced under cloudy conditions, whereas cloudy-scene robustness is enhanced by microwave observations. The ABLH annual cycle is found to peak in May and reach minima in August and December, mainly controlled by boundary-layer thermal evolution, cloud frequency, and low-level jets. Daily mean ABLH is closely linked to surface temperature, especially in spring and autumn. Stable and neutral boundary layers dominate: stable cases occur more often when cloud base lies above the ABL, while neutral cases are more frequent when cloud base lies below it. ABL development is promoted by wind shear beneath low-level jets, which increases with jet-core wind speed; wind shear stress, friction velocity, and downward longwave radiation also contribute. Model generalizability is confirmed by extrapolation to other Arctic sites and representative case studies. Overall, complementary active and passive remote-sensing observations are effectively integrated to generate continuous, condition-aware ABLH retrievals in complex Arctic environments, supporting characterization of boundary-layer variability and Arctic atmosphere–surface interactions.