Human-induced vegetation shifts drive global declines in urban vegetation transpiration

Description This repository contains the source and processed tabular datasets supporting the manuscript “Human-induced vegetation shifts drive global declines in urban vegetation transpiration”. The study develops an observation-constrained hybrid convolutional neural network–Penman–Monteith (CNN–PM) framework to estimate urban vegetation transpiration (VTᵤ) across 3,560 global cities during the historical period (1986–2024) and project its evolution from 2025 to 2100 under SSP1-2.6, SSP2-4.5 and SSP5-8.5. The current archive contains 27 CSV files accompanied by 11 dataset-specific README files. The datasets support the principal results, figures and supplementary analyses reported in the manuscript. Dataset contents The repository includes the following thematic components: Eddy-covariance observations and VTᵤ partitioning Daily VTᵤ estimates from 20 urban eddy-covariance sites derived using three observation-based evapotranspiration partitioning methods—the underlying water-use efficiency (uWUE), Pérez-Priego (PP) and transpiration estimation algorithm (TEA) methods—together with their ensemble fusion estimates. Flux-footprint and land-cover characterization Site-level flux-footprint source areas, footprint-scale land-cover composition and supporting information for matching eddy-covariance observations with remotely sensed surface characteristics. CNN–PM model training and evaluation Training and validation datasets, model feature importance, Kling–Gupta efficiency distributions, genetic-algorithm optimization trajectories, extreme-condition performance tests and repeated held-out-site generalization results. Anthropogenic heat flux Historical and future anthropogenic heat flux estimates for 3,560 cities, together with population, nighttime-light and impervious-surface information used for spatial allocation. Global urban VTᵤ patterns City-level long-term mean VTᵤ, temporal trends and anomaly distributions for the historical and future periods, including regional and SSP-specific summaries. Vegetation structure and plant functional types Trends in fractional vegetation cover, impervious surface fraction, woody vegetation fraction and grassland fraction; VTᵤ changes across nine plant functional types; and vegetation-transition information associated with urbanization. Trend attribution City-, region- and scenario-level attribution results for 13 landscape, vegetation and climate variables. These include raw XGBoost–SHAP outputs and aggregated contributions from vegetation structural change and climate–atmospheric forcing. Any derived or transformed SHAP tables are identified in the accompanying README and should be distinguished from the raw model outputs. Spatial and temporal coverage Spatial coverage: 3,560 cities worldwide and 20 urban eddy-covariance sites Historical period: 1986–2024 Future period: 2025–2100 Future scenarios: SSP1-2.6, SSP2-4.5 and SSP5-8.5 Primary temporal resolutions: daily, monthly, annual and period-level summaries Primary units: VTᵤ in mm day⁻¹ or mm year⁻¹; VTᵤ trends in mm year⁻²; anthropogenic heat flux in W m⁻² File format All data tables are provided as comma-separated values (CSV) files with descriptive column headers and units. Each thematic dataset is accompanied by a README file documenting its variables, units, spatial and temporal coverage, and relationship to the manuscript figures and tables.