We address the problem of uplink channel estimation in TDD massive multiuser multi-input-multi-output (MU-MIMO) systems, when the uplink training duration is limited. Based on the concept of compressive sensing (CS), the uplink channel could be estimated with limited training duration if the channel can be sparsely represented. In this paper, a low-rank matrix approximation (LRMA) based on CS technique is proposed for the massive MU-MIMO channel estimation problem. As such, the channel estimation problem was formulated as a quadratic nuclear norm optimization problem with linear constraint. Consequently, the regularization parameter, which minimizes the error between a data fidelity and convex penalty function, is selected based on cross-validation (CV) curve method. The simulation results demonstrate that the proposed method outperforms the LS method in terms of the estimator performance. In addition, the proposed method reduces the pilot length and the computational cost as well.