Environmental degradation from rapid urbanization significantly threatens ecological resilience (ER). Nevertheless, accurately evaluating ER remains a persistent challenge. Prior studies’ limited attention to resilience’s cross-scale complexity has hindered evidence-based management. This study, based on long-term time series remote sensing and multi-source data, developed a cross-scale spatiotemporal ER analysis framework integrating landscape ecology and panarchy perspectives. A local “resistance–adaptation–recovery” substrate resilience evaluation was combined with telecoupling-based global network resilience to quantify multi-scale ER from 2000 to 2020. Key drivers across time scales were identified using a hybrid XGBoost–SHAP and genetic algorithm (GA)–optimized dynamic Bayesian network (DBN), and spatial optimization scenarios were simulated with patch-generating land use simulation (PLUS) model. ER decreased slightly from 0.4856 in 2000 to 0.4503 in 2020, with dynamic fluctuations across periods. A clear spatial pattern emerged, with higher ER in the east and lower in the west. Forest land contributed strongly to ER, while construction and cropland reduced it. Spatial composition factors—especially the proportions of forest and construction land—were dominant drivers, outweighing structural factors such as landscape pattern. DBN backward inference revealed nonlinear threshold effects among socio–natural–spatial drivers. Scenario-based simulations confirmed that regulating spatial composition via our optimization pathway can enhance ER. This is particularly effective when expanding forestland in mountainous regions while restraining the growth of built-up areas. This study proposes an integrated framework of “resilience assessment—driver analysis—spatial optimization,” which not only advances the theoretical basis for nested ER assessment but also offers a transferable approach for optimizing spatial patterns and sustainable land management, thereby enhancing ecological resilience in rapidly urbanizing regions.