Mapping Fog Frequency: A Machine Learning Approach

Fog is an important yet often-overlooked non-rainfall water source for many coastal, mountainous, and arid-to-semi-arid regions worldwide. From the limited ecological and agricultural studies conducted to date, we know that fog is particularly beneficial for plants during dry spells. Specifically, fog events support plant hydration and growth via foliar water uptake, enhance water- and light-use efficiency by reducing evapotranspiration and increasing light scattering, and contribute water and nitrogen inputs to soils. Currently, a major limitation in further assessing the effects of fog on vegetation, as well as changes in fog patterns under ongoing climate warming, is the scarcity of fog data with broad spatiotemporal coverage. To address this gap, we built three ensemble decision-tree machine learning models—Random Forest, LightGBM, and XGBoost—to predict fine-scale monthly fog frequency using ERA5-Land data and physiographic and temporal parameters. Hourly fog observations from 136 ASOS weather stations in California were used as ground truth, and spatial and temporal holdout strategies were applied to ensure generalization. Overall, the models effectively rank and classify monthly fog frequency across seasons, with the strongest performance during July and August, when dry spells are most prevalent in California. Our methodology demonstrates how large-scale climatic data can be paired with physiographic and temporal information to map fog frequency, with models that are agnostic to fog-formation mechanisms and transferable for use in other regions. Beyond frequency mapping, this work provides insights into the drivers of fog formation through Shapley Additive exPlanations (SHAP) analysis. Dewpoint depression was found to be the most influential predictor, making it a good candidate for informing projections of future shifts in fog patterns. While climate-change studies have focused on important climatic variables such as temperature and precipitation, little do we understand about how fog patterns have changed in the recent past, or will shift in the future. Our approach can serve as the basis for assessing both.