Abstract: Objectives: Co-seismic landslide identification relies heavily on the geometric structures and spectral information of remote sensing imagery. However, traditional pan-sharpening methods, primarily designed for generic visual reconstruction, struggle with spectral confusion and fragmented landslide boundaries in complex environments. To address these challenges, this study proposes a task-driven unsupervised pansharpening network, termed OD-PSGAN, which establishes a gradient coupling mechanism spanning from low-level image reconstruction to high-level semantic perception, facilitating deep synergy between spatio-spectral fusion and downstream interpretation tasks. Methods: First, to accommodate the multiscale characteristics of landslide bodies, an asymmetric dual-stream architecture is employed for deep spatio-spectral decoupling, integrated with a multiscale residual attention mechanism to enhance landslide edge representation via non-linear spatial response reweighting. Second, to suppress structural artifacts common in Generative Adversarial Networks (GANs), a gradient branch is designed within the discriminator to explicitly enforce structural consistency between fused and original images in the gradient space. Furthermore, a joint loss function comprising cycle-consistency, structural restoration, and semantic retention is constructed to guide the network in maintaining high spectral fidelity while prioritizing the extraction of discriminative features conducive to landslide identification. Results: End-to-end validation across simulated degradation and full-resolution scenarios demonstrates that OD-PSGAN significantly improves the robustness of land-cover identification in shadowed and complex terrains. Specifically, the identification accuracy (F1-score and mIOU) increased by 5.41% and 2.40%, respectively, compared to the secondbest method. Conclusions: This study confirms the feasibility of the task-driven paradigm in bridging the "semantic gap" between image fusion and semantic perception, providing a theoretical foundation and technical support for task-oriented remote sensing image processing.