The exploration of vegetation patchiness and pattern-driven multistability improves our understanding of ecosystem functioning and resilience in drylands, yet there is still a lack of validation of such early warning indicators and underlying mechanisms for real ecosystems until now. Here, we identify nearly 20 million individual shrub islands across a large-scale continuous environmental gradient in China combining with remote sensing and deep learning. We investigate two indictors of ecosystem functioning and resilience: ecosystem biomass, shrub island patch size. Two indicators follow consistent and non-linear variations along with environmental gradients, as indicated by three stages (gradual change, almost constant and sharp shifts). Such delay of sustained decline in second stage demonstrates the resilience of dryland, determined by multistability driven by vegetation-sand-water interactions at landscape level and self-adaptation of individual shrubs in response to environmental changes. These findings enhance our understanding and managing of discontinuous state shifts in drylands.