Effectiveness of Predicting Event-Level Interception Losses Across Diverse Vegetated Sites Using Statistical Models

Abstract

Understanding evaporation from wet canopies across ecosystems is challenged by its spatiotemporal variability and associated observational challenges. Precipitation and throughfall were measured at 22 sites of the National Ecological Observatory Network using consistent methodologies across diverse climates and ecosystems, providing a novel opportunity to examine the performance of statistical models in predicting interception loss from meteorological and canopy-structure data. We used those data to quantify event-level interception losses and found wide variation; median interception loss of small storms (< 10 mm) was 37.7% (39% inter-quartile range), and of large storms (> 50 mm) was 19.8% (20% interquartile range). We found storm gross-precipitation depth was the most important variable for predicting the amount of interception loss (predicting ~70% of the variation), followed by mean canopy height and air temperature. Storm gross-precipitation depth was also an important predictor of interception loss as a percent of storm depth, but much less variation was explained (R² = 0.11, RMSE = 24%). Prediction of percent interception loss improved (R² = 0.32 and RMSE = 19%) by including additional meteorological and vegetation structure characteristics in a random forest model. In addition to demonstrating the greater importance of storm traits over vegetation traits in predicting event interception losses, this analysis showed that relationships between storm traits and interception losses differed among sites; these inconsistent relationships across sites limited the ability for any statistical model to perform well in predicting event-level interception losses across sites, which may justify the use of alternative approaches (e.g. process-based models).