In this work, we investigate the channel estimation (CE) problem for extremely large-scale multiple-input-multiple-output (XL-MIMO) systems, considering both the spherical wavefront effect and spatial non-stationarity (SnS). Unlike existing non-stationary CE methods that rely on the statistical characteristics of channels in the spatial or temporal domain, our approach seeks to leverage sparsity in both the spatial and wavenumber domains simultaneously to achieve an accurate estimation.To this end, we introduce a two-stage visibility region (VR) detection and CE framework. Specifically, in the first stage, the belief regarding the visibility of antennas is obtained through a structured message passing (MP) scheme, which fully exploits the block sparse structure of the antenna-domain channel. In the second stage, using the obtained VR information and wavenumber-domain sparsity, we accurately estimate the SnS channel employing the belief-based orthogonal matching pursuit (BB-OMP) method. Simulations demonstrate that the proposed algorithms lead to a significant enhancement in VR detection and CE accuracy, especially in low signal-to-noise ratio (SNR) scenarios.