Learning Evaporative Fraction with Memory

Abstract. Evaporative fraction (EF), defined as the ratio of latent heat flux to the sum of sensible and latent heat flux, is a key metric of surface energy partitioning and an indicator of plant water stress. Recognizing the role of vegetation memory effects, we developed an explainable machine learning (ML) model based on a Long Short-Term Memory (LSTM) architecture, which explicitly incorporates memory effects, to investigate the mechanisms underlying EF dynamics. The model was trained using data from 90 eddy-covariance sites across diverse plant functional types (PFTs), compiled from the ICOS, AmeriFlux, and FLUXNET2015 Tier 1 datasets. It accurately captures EF dynamics – particularly during post-rainfall pulses and soil moisture dry-down events – using only routinely available meteorological inputs (e.g., precipitation, radiation, air temperature, vapor pressure deficit) and static site attributes (e.g., PFT, soil properties). The ensemble mean predictions showed strong agreement with observations (R² = 0.82) across sites spanning broad climate and ecosystem gradients. Using explainable ML techniques, we identified precipitation and vapor pressure deficit as the primary drivers of EF in woody savanna, savanna, open shrubland, and grassland ecosystems, while air temperature emerged as the dominant factor in deciduous broadleaf, evergreen needleleaf, and mixed forests. Furthermore, expected gradients revealed variation in memory contributions across PFTs, with evergreen broadleaf forests and savannas exhibiting stronger influences from antecedent conditions compared to grasslands. These memory effects are strongly associated with rooting depth, soil water-holding capacity, and plant water use strategies, which collectively determine the time scales of drought response. Notably, the learned memory patterns could serve as proxies for inferring rooting depth and assessing plant water stress. Our findings underscore the critical role of meteorological memory effects in EF prediction and highlight their relevance for anticipating vegetation water stress under increasing drought frequency and intensity.