Because the training serials are unipolar in multi-input multi-output free space optical communication (MIMO-FSO) systems, which leads to the signal energy loss easily, the result of traditional singular value decomposition (SVD) channel estimation algorithm is inaccurate. A kind of modified adaptive SVD estimation algorithm is put forward which can compensate for the estimation errors of SVD algorithm. So the modified adaptive SVD estimation algorithm can be better fit for MIMO-FSO systems. The simulation results indicate that compared with the SVD algorithm, the mean square error (MSE) performance of the modified algorithm is improved by two and three orders of magnitude when the signal-to-noise ratio (SNR) is 15 dB and 30 dB, respectively. Compared with the average value modified SVD algorithm, the modified algorithm has 1 dB improvement of the MSE performance under the same condition. The modified algorithm possesses strong transplanting property and can be adopted in other channel estimation methods.