Vegetation–Environment Resilience Diagnostic Intelligence — proxy-label construction, data processing and the three-stage framework.

VERDI couples vegetation state and environmental conditions to assess three operational, proxy-based resilience indicators for urban vegetation, and to translate the resulting factor readouts into diagnostic narratives. The implementation covers municipal tree censuses, Sentinel-2 and Landsat imagery, ERA5-Land reanalysis, SMAP soil moisture and static GIS layers for three metropolitan areas (New York City, Paris, Melbourne) on a 100 m grid.

Code accompanying the paper "From observation to evaluation: Coupling vegetation state and environmental conditions for real-time urban vegetation resilience assessment" (Urban Forestry & Urban Greening).

Contents

The label-construction and data-processing code implements Section 3.1 and Appendix A of the paper; the Stage 1–3 model code implements Section 3.2 and Appendix E. Stage 3 returns the diagnostic prompts. Trained weights, the fine-tuned language model and processed city datasets are not included; see Data.

The three labels are operational indicators of selected resilience dimensions; they do not measure recovery, persistence, functional stability or post-disturbance regeneration.

Architecture

flowchart LR
    A["Observed and externally<br/>derived inputs"]:::data --> B["Stage 1<br/>Spatial World Model<br/>coupled latent, K=6"]:::model
    B --> C["Stage 2<br/>Masked Sensor Transformer<br/>fusion under missing modalities"]:::model
    C --> D["Stage 3<br/>attribution, typology,<br/>diagnostic prompts"]:::stage3
    D --> E["Proposed decision support<br/>retrospective scenarios"]:::use
    L["Proxy labels<br/>R*A, R*B, R*C"]:::label -.-> B
    L -.-> C
    classDef data fill:#10221C,stroke:#5AAAC8,color:#E9F4EE
    classDef model fill:#10221C,stroke:#8CC85A,color:#E9F4EE
    classDef label fill:#10221C,stroke:#9FB8AE,color:#E9F4EE
    classDef stage3 fill:#10221C,stroke:#8CC85A,color:#E9F4EE
    classDef use fill:#10221C,stroke:#9FB8AE,color:#E9F4EE

Loading

Repository layout

Path	Purpose
verdi/config.py	All parameters and their defaults
verdi/labels/	R*_A, R*_B, R*_C and the label-fusion strategies
verdi/data/	Census harmonisation, grid construction, exclusion rules, z-scores, spatial partitioning, output schema
verdi/model/	Stages 1–3 and their losses
verdi/eval/	Prediction, representation and calibration metrics
scripts/	Runnable entry points for the labels and the spatial split
docs/	Label definitions, data sources, processing rules, model reference
tests/	Synthetic-data checks of the formulas, the split and the model shapes

Proxy labels

Label	Meaning	Built from	Availability
R*_A	Relative cross-environment condition	Census health and within-species NDVI and DBH percentiles	NYC, Melbourne (Paris has no census health field)
R*_B	Short-term NDVI retention around qualifying heat events	ERA5-Land daily maximum temperature, Sentinel-2 pre/post pairs	Cells with a qualifying event and a cloud-free pair
R*_C	Relative deviation from a species baseline	Sentinel-2 NDVI	All cells in all three cities; the primary reported target

Formulas, thresholds and configurable settings are documented in docs/labels.md.

Quick start

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Build the three labels on a synthetic city and write the output table
python scripts/build_labels.py --city nyc --cells 400

# Create and summarise the spatial split for a city
python scripts/make_partition.py --city melbourne

# Agreement between the labels that follows from their construction alone
python scripts/label_overlap.py --cells 4000 --seeds 5

# Run the checks
python -m pytest tests/

The scripts find the package from the repository root. build_labels.py simulates the city, so its output exercises the plumbing and the label formulas; replace the synthetic inputs with the provider data listed in docs/data_sources.md to build the real tables.

Data

Raw provider data is not distributed here. The censuses (TreesCount! 2015, Les Arbres de Paris, Melbourne Urban Forest Visual), imagery and derived products are obtained from their original providers under their own licences: Sentinel-2 and Landsat are open (Copernicus, USGS public domain), ERA5-Land and SMAP are free with attribution, SoilGrids and ESA WorldCover are CC-BY, OSM data is ODbL, and GHSL is open. Per-variable sources, resolutions and derivations are in docs/data_sources.md; the processing rules, exclusion criteria and partition scheme are in docs/processing.md.

License

MIT — see LICENSE. Provider data retains the licence of its source.