Aiming to take full advantage of the backward scattering characteristics for different land cover types in different temporal and polarization SAR data, the recursive nonnegative matrix underapproximation (RNMU) was used for the fusion of multisource SAR data, and the fused SAR image was used to achieve a highprecision land cover classification. According to the characteristics of different SAR image modes, the input SAR images were preprocessed firstly, and then the matrix decomposition of the SAR images and the iterative solution of the optimal matrix were implemented based on RNMU. To verify the effect of application of integrated SAR image on land cover classification, taking Da’an City in Jilin Province as an example, RNMU was used for the fusion of multitemporal VV/VH dual-polarization Sentinle-1 SAR image and HH/HV dualpolarization GF-3 data. The main types of land cover in the study area were classified with the fused SAR data based on RNMU. The results illustrated that SAR data fused based on RNMU algorithm had sound performance in the land cover classification with 93.11% overall accuracy and 0.86 Kappa coefficient, which outperformed the Gram-Schmid (G-S) fusion method with 6.83 percentage points and 0.12 higher in overall accuracy and Kappa coefficient respectively. The attempt of multi-source SAR fusion provided an effective means for SAR image fusion and provided more high-precision data resources for land cover classification.