Current remote sensing of dryland ecosystems is fundamentally limited by a reliance on vegetation indices ("greenness"), which struggle to disentangle mixed pixel signals and fail to capture the non-photosynthetic structural components critical for resilience. This "greenness bias" obscures the dynamics of sparse vegetation and leads to significant uncertainties in distinguishing true restoration from ephemeral greening. Here, we developed a novel spatio-temporal framework integrating continuous change detection and classification with spectral mixture analysis to decouple sub-pixel photosynthetic vegetation, non-photosynthetic vegetation (NPV), and soil fractions across the temperate drylands of China. By reconstructing 21-year continuous time series of endmember abundance, we mapped previously undetected sparse shrublands, revealing they constitute a dominant "landscape skeleton" covering nearly half of the vegetated area. Our results uncover a profound asymmetry in ecosystem transitions: while degradation typically manifests as an abrupt collapse of vegetation-soil coupling, recovery follows a distinct "structure-first" hysteresis. This recovery pathway is driven by the gradual accumulation of NPV, which stabilizes substrates and facilitates subsequent canopy expansion. Furthermore, trajectory analysis demonstrates that these resilience mechanisms are strictly substrate-dependent, with restoration pathways diverging significantly across sandy, gravelly, and saline substrates. By moving beyond pixel-level greenness to sub-pixel structural characterization, this study provides new insights into the asymmetric processes shaping drylands and offers a robust, spatially explicit tool for monitoring adaptive restoration under climatic and anthropogenic pressures.