Due to load traffic disparity in downlink (DL) and uplink (UL) in heterogeneous cellular networks (Het-Nets), dynamic time-division duplexing (TDD) is proposed to dynamically allocate UL and DL resources. Under the same circumstances, decoupled UL/DL access is introduced to balance between UL and DL transmissions and to further improve the overall system performance. In classical HetNets, coupled UL/DL access (CUDA) mode is adopted, where each user is associated in downlink and uplink with a single cell. In the next generation HetNets, rather than belonging to a specific cell, a mobile user can receive the downlink traffic from one base station (BS) and send uplink traffic through another BS. This situation is referred to as downlink and uplink decoupled access (DUDA). In this paper, we analytically investigate a joint TDD and DUDA statistical model based on a geometric probability approach. Taking all possible TDD subframes combinations between the macro and small cells, four coupled and decoupled cell associations strategies are investigated in details. We derive inter-cell interference and capacity expressions for each of these association policies to assess the improvement brought by DUDA mode to TDD HetNets. In this particular DUDA mode, users are associated in UL with the small cell and in DL with the macro cell. Moreover, we identify the small cell offset under which the decoupled mode performs significantly better and maintains a higher spectral efficiency in both DL and UL. Monte-Carlo simulations results are presented to validate the accuracy of the analytical model.