Local Overfitting in Machine Learning Models Drives Spatially Uneven Performance in Atmospheric Observations

Abstract

Atmospheric and environmental monitoring relies on observation networks with highly uneven spatial density, yet machine-learning models trained on such networks prioritize data-dense regions, producing systematic uneven accuracy precisely where reliable predictions are most needed. Existing fairness methods from machine learning address such gaps but require explicit group definitions and domain expertise. We show that uneven accuracy is strongly associated with local overfitting: optimal model complexity and training schedule diverge across time-, space-, and spatiotemporal cross-validation splits, so that no single hyperparameter setting is optimal for all. Using a range of environmental datasets, AI models tend to memorize dense regions while others generalize better but lose local structure. These dynamics motivate our Overfit-to-Generalization framework, which integrates into existing pipelines. Across nine model configurations, OG improves R 2 by 10.8–29.7% over baselines, largest in data-sparse regions. A complementary diagnostic predicts model accuracy at unsampled locations from local density and spatial proximity.