This paper proposes a low complexity channel estimation algorithm for unmanned aerial vehicle (UAV) three dimension multi-user multiple-input-multiple-output (3D MIMO) systems with the uniform planar array (UPA) at base station (BS) using paired spatial signatures. With the aid of antenna array theory and array signal processing, firstly 3D channel modeling based on the angle between the direction of arrival (DOA) and the x- direction of the array antenna, and the angle between the DOA and the y- direction of the array antenna. And 3D MIMO channels can be projected onto the x- and y-directions, respectively. Then, channel estimation for multiuser uplinks using small amount of training resources, which is divided into two phases. The first phase is the uplink preamble phase, which assigns each user an orthogonal pilot, and obtains the paired spatial signatures and optimal rotation angle of each user through the same pilot sequence. We also propose a user grouping strategy based on three dimension angle-division multiple access (3D-ADMA) to ensure that the user's spatial signatures do not overlap. The second phase is much coherence times after the preamble phase, which assigns the same pilot sequence within a group and assigns orthogonal pilot sequences between groups, and the channel state information of the user's x- and y-directions is recovered by the paired space signatures and optimal rotation angle of each user obtained in the preamble phase, respectively. And dynamically updating the user's paired spatial signatures and optimal rotation angle utilizes the obtained channel components of x- and y-directions. Finally, the channel components of the x- and y-directions are reconstructed by the updated user's space signatures and the optimal rotation angle, and the 3D MIMO channel is generated by the Kronecker product. Compared with the conventional channel estimation method of a 3D MIMO system under UPA using a low rank model, the proposed methods greatly reduce the computational complexity without reducing the estimated performance, it is carried out with limited training resources, and the pilot resource overhead of the system is greatly reduced by the 3D-ADMA packet and the two-stage pilot allocation. Various numerical results are provided to corroborate the proposed studies.