Quantifying Seasonal Vegetation Anomalies in an Arid Ecosystem: A Reproducible, Leakage-Free Machine Learning Framework with Dynamic Climate Predictors and Quantified Uncertainty
Year: 2026
Type: preprint
Abstract: Vegetation stress in arid ecosystems is a critical early-warning indicator for desertification and agricultural risk, yet its modelling remains methodologically challenging: climate forcing is highly stochastic, and standard anomaly pipelines frequently leak future information into historical baselines, inflating apparent predictive skill. This study presents a reproducible, open-source machine-learning framework for modelling seasonal vegetation anomalies in the Cholistan Desert, Pakistan, using MODIS NDVI (MOD13Q1; 250 m, 16-day; 2018–2024). In contrast to conventional workflows built on static climate normals, we derive time-varying seasonal predictors (precipitation, one-season-lagged precipitation as a proxy for antecedent moisture conditions, and temperature) from the NASA POWER archive, alongside SRTM terrain derivatives and MODIS land cover (MCD12Q1). To guarantee an honest evaluation, the seasonal anomaly baseline (pixel- and season-specific mean and standard deviation) is estimated strictly on the 2018–2023 training period, and the entire year 2024 is withheld. Generalized additive models (GAM) and random forests (RF) are compared under 4-fold spatial-block cross-validation (n = 30,646 valid pixel–composite observations assigned to meteorological seasons) and the temporal hold-out (n = 4,179 valid pixel–composite observations from the 2024 hold-out). RF consistently outperformed GAM (mean spatial-CV = 0.261 vs. 0.180; 2024 hold-out = 0.156 vs. 0.100), with seasonal temperature, seasonal precipitation, and lagged precipitation dominating variable importance, supporting the use of dynamic climate predictors. Predictive uncertainty, proxied by GAM–RF disagreement, was markedly elevated where 2024 climate conditions fell outside the training distribution, transparently exposing where predictions amount to extrapolation. The framework provides a leakage-free template for dryland vegetation monitoring under intensifying climate variability and identifies priority spatiotemporal domains for targeted field validation.
Source: Research Square
Author: aamna naveed
Institution: Islamia University of Bahawalpur
Language: English
Subfield: Ecology
Field: Environmental Science
Domain: Physical Sciences
SDGs: Climate action, Life on land
Open Access status: green
OpenAlex ID: W7223885943