A very accurate ecological model was predicting space far better than year-to-year change
WHAT THE EVIDENCE SAYS
Researchers analyzed the Remote Sensing Ecological Index across China from 2002–2024 and built one-year-ahead interpretable machine-learning models with spatial and temporal holdouts. XGBoost achieved weighted R² of 0.9712 and RMSE of 0.0262, 13.67% better than a persistence baseline. Yet 97.39% of total variance was between-pixel spatial difference and only 2.61% reflected within-pixel year-to-year change; the factors associated with spatial patterns also differed from those associated with temporal variation.
WHY IT MATTERS / OUR READ
Headline predictive accuracy can hide what is actually being predicted. For change detection and adaptive management, a model that distinguishes persistent spatial structure from interannual dynamics is more informative than one high R² treated as evidence of temporal foresight.
CAVEATS / UNCERTAINTY
RSEI is a composite remote-sensing index rather than a direct biodiversity or ecosystem-function measure. Associations identified by SHAP are not causal, and national-scale predictability does not guarantee local decision performance.
SOURCE
Scientific Reports — primary study
View original evidence ↗
China’s RSEI model reached weighted R² of 0.9712, but variance decomposition showed 97.39% of the signal was between pixels and only 2.61% was within-pixel interannual variability.
Read the Wild Lands assessment →