Global urban tree leaf area under climate change: divergent projected trajectories and the limits of space-for-time inference

Abstract Urban trees moderate city climates, yet how their leaf-area density relates to projected future climates is poorly constrained globally. We fitted a Random Forest model to monthly temperature and precipitation normals and the multi-year mean leaf area index (LAI) derived from a satellite-informed machine-learning urban tree product across 21,590 urban cells, then projected end-of-century change under six Shared Socioeconomic Pathways. Leave-one-continent-out validation gave $$R^2 = 0.846$$ (95% CI $$0.813$$ – $$0.873$$ from spatial-block resampling), but that spatial skill does not transfer to time: driven by observed reanalysis anomalies over exactly the satellite record, the same relationship reproduces neither the magnitude nor the pattern of observed change ( $$\rho = +0.087$$ ). The response variable is itself a machine-learning product built from overlapping climate predictors, so the spatial fit is not independent evidence of that relationship either. Under modified Mann–Kendall tests with false-discovery-rate control, 19.8% of cells greened and 5.3% browned during 2000–2022, roughly half the browning obtained without those corrections. Global mean change is positive under all scenarios, yet 32.6% of cells decline under SSP5-8.5, with declines concentrated in tropical humid cities and gains at high latitudes. Two further results bound them: 10.6% of cells fall outside the fitted climate domain under SSP5-8.5, and four algorithms agree on the sign of change at only 68.1% of cells. We report these as changes implied by transferring a spatial relationship into future climates, not forecasts: the zonal contrast is the most reproducible result across the evaluated algorithms; cell-level magnitudes and city orderings are not.