Human-induced vegetation shifts drive global declines in urban vegetation transpiration
Zenodo Description Spatial and temporal coverage Key findings supported by these data
- Source
- Zenodo
- Published
- 2026
- Region
- Global
- Format
- csv
- Access
- Open access
- License
- cc-by-4.0
- Themes
- Plants & Vegetation, Sustainable Cities
About this dataset
Zenodo Description
Title
Source data for: Human-induced vegetation shifts drive global declines in urban vegetation transpiration
Description
This repository contains the complete source datasets underlying the manuscript "Human-induced vegetation shifts drive global declines in urban vegetation transpiration." The study develops a hybrid physics–deep learning framework (CNN-PM) to reconstruct and project urban vegetation transpiration (VTu) across 3,560 global cities from 1986 to 2100, and attributes the observed and projected VTu decline to urbanization-driven vegetation structural transformation rather than climatic forcing.
Dataset overview
The repository comprises 10 thematic dataset packages totaling over 30 individual CSV files, organized by their role in the analysis pipeline:
1. Eddy covariance observations and transpiration partitioning
Daily vegetation transpiration (VTu) estimates derived from three independent high-frequency turbulence partitioning methods — Flux Variance Similarity (FVS), Conditional Eddy Covariance (CEC), and Modified Relaxed Eddy Accumulation (MREA) — plus a multi-method fusion product, for 20 globally distributed urban eddy covariance flux tower sites. Each method provides a central estimate and 5th–95th percentile uncertainty bounds in mm/day.
2. Flux footprint and land-use characterization
Footprint model summaries (equivalent radii at 60%, 70%, 80%, and 90% source-area contours), baseline land-use composition (impervious, vegetation, water, bare/other), and source-area-resolved land-use fractions for the top-eight sites, along with visual encoding lookup tables for figure reproduction.
3. Model training and hyperparameter optimization
Generation-by-generation convergence trajectories of the Genetic Algorithm (GA) optimization for the two sub-CNN models (CNN_c for climatic forcing, CNN_a for anthropogenic forcing) across 300 cross-validation iterations and 70 generations, recording validation loss, RMSE, KGE, learning rate, weight decay, and batch size with 5th–95th percentile envelopes.
4. Model validation and performance evaluation
Feature importance rankings, extreme-scenario stress test results (RMSE and bias for 10 scenarios × 6 model configurations), observed-versus-predicted scatter data for training (15 sites), validation (3 sites), and independent generalization (2 sites) datasets, and KGE distribution samples. Summary statistics: training KGE = 0.94, validation KGE = 0.93, generalization KGE = 0.90.
5. Anthropogenic heat flux (AHF) estimates
City-level AHF for 3,560 cities under historical and future conditions (W/m²), derived from nighttime light remote sensing calibrated with energy consumption statistics (historical) and impervious surface projections (future). Global mean AHF increases from 20.8 to 31.1 W/m², a 50% rise.
How the data was collected
Zenodo Description
Title
Source data for: Human-induced vegetation shifts drive global declines in urban vegetation transpiration
Description
This repository contains the complete source datasets underlying the manuscript "Human-induced vegetation shifts drive global declines in urban vegetation transpiration." The study develops a hybrid physics–deep learning framework (CNN-PM) to reconstruct and project urban vegetation transpiration (VTu) across 3,560 global cities from 1986 to 2100, and attributes the observed and projected VTu decline to urbanization-driven vegetation structural transformation rather than climatic forcing.
Dataset overview
The repository comprises 10 thematic dataset packages totaling over 30 individual CSV files, organized by their role in the analysis pipeline:
1. Eddy covariance observations and transpiration partitioning
Daily vegetation transpiration (VTu) estimates derived from three independent high-frequency turbulence partitioning methods — Flux Variance Similarity (FVS), Conditional Eddy Covariance (CEC), and Modified Relaxed Eddy Accumulation (MREA) — plus a multi-method fusion product, for 20 globally distributed urban eddy covariance flux tower sites. Each method provides a central estimate and 5th–95th percentile uncertainty bounds in mm/day.
2. Flux footprint and land-use characterization
Footprint model summaries (equivalent radii at 60%, 70%, 80%, and 90% source-area contours), baseline land-use composition (impervious, vegetation, water…
3. Model…