In Massive MIMO-OFDM systems, the channel shows strong correlations in both spatial and frequency domain. Aiming at the problem that only spatial or frequency domain correlation is considered in most of the existing compressed feedback algorithms, a joint spatial-frequency compression algorithm is proposed. First, a two dimensional sparsity of channel in spatial-frequency domain is analyzed according to the compressed sensing theory. Then, a joint sparse matrix of channel is derived. Based on the joint sparse matrix, the joint spatial-frequency compression algorithm is presented. Simulation results and analysis show that, the proposed algorithm can significantly reduce the feedback load with acceptable accuracy.