Machine learning uncovers a multi-year climate memory in permafrost degradation on the Qinghai–Tibet Plateau: the critical roles of precipitation and lagged temperature

The Qinghai–Tibet Plateau (QTP), which hosts the world’s largest area of alpine permafrost, is experiencing accelerated degradation due to climate warming, posing significant threats to regional hydrological cycles and ecosystem stability. While existing research has primarily focused on direct temperature impacts, the influence of precipitation and the multi-year lagged responses of permafrost thermal regimes remain insufficiently quantified. To address these gaps, we employed a machine learning approach that integrates multi-year (0–5 years) lagged climatic features (temperature and precipitation) to model permafrost distribution and active layer thickness (ALT) across the QTP from 1960 to 2020. Our comparative analysis of three machine learning paradigms revealed that CatBoost delivered superior predictive performance (testing set F1-score = 0.979; R² = 0.791). Crucially, this high performance is directly attributable to the model’s capacity to leverage a multi-year “climate memory”, which highlights the importance of incorporating lagged climate features in permafrost change simulation. Interpretability analyses of the CatBoost model further reveal that winter snowfall acts as a key insulator, whereas spring and summer rainfall accelerate thawing by increasing soil thermal conductivity. Spatiotemporal analysis identified a net permafrost retreat of 2.51 × 10⁴ km² per decade. Notably, ALT dynamics exhibited a pronounced regime shift around 1980, transitioning from a thinning trend (−5.2 cm/decade, 1960–1980) to rapid thickening (+4.1 cm/decade, 1980–2000). These results establish that a robust understanding of QTP permafrost dynamics requires moving beyond simple temperature-driven models to incorporate the interacting roles of seasonal precipitation and cumulative climatic legacies.