In preparation
A leakage-aware per-region framework for NDVI forecasting across 133 global ecoregions
A per-region NDVI forecasting framework combining 89 region-specific CatBoost specialists with a per-cell Bayesian-ridge router, trained on 22 years of satellite-derived climate, soil, fire, and topographic drivers. The methodological contribution is a four-test leakage-audit protocol that catches sister-vegetation features, shared cross-region statistics, leaky CV splits, and pheno-prior contamination — failure modes that inflate published R² scores by 0.05–0.20 in much of the existing literature. After applying the audit, cross-continental R² on a held-out 44-region test set is approximately 0.93.
Planned
- Open benchmark release — public dataset of 133-region NDVI prediction targets + 19-module driver matrix, with the v2 leakage-audited train/test split (89/44). Intended as a community reference for honest evaluation of regional vegetation models.
- Methodology companion paper — standalone treatment of the four-test leakage protocol with worked examples on a smaller open dataset, designed to be cited independently of the Desertflow framework.
Why preprint-first?
Closed-access journals add a 6–18 month delay between methodology lock and public availability. For an evolving framework where the methodology is the contribution, that delay actively harms reproducibility. Posting to EarthArXiv makes the work citable immediately, lets reviewers verify the audit on the live codebase, and stays open after acceptance.
If you're a reviewer, collaborator, or researcher interested in the framework: (click to copy)
← Back home