Process-level assessment of aridity-regime sensitivity in Penman-Monteith and PT-JPL evapotranspiration models across climate and aridity gradients using Ameriflux observations and deep learning
Abdul Wahed Nab1, Muhammad Jawad1, Ali Behrangi1, Guo-Yue Niu1
1Department of Hydrology and Atmospheric Sciences, University of Arizona, Tucson, AZ, USA
Abstract
Accurate evapotranspiration (ET) estimation remains challenging, particularly in drylands where evaporative demand and surface moisture constraints expose structural weaknesses that simple benchmark comparisons miss. Here, we move beyond a head-to-head model comparison by combining network-scale evaluation (92 AmeriFlux sites), regime-based diagnostics, and hybrid post-processing attribution to o explain divergence between PM and PT-JPL across climate space. Both models were forced and optimized consistently to enable fair comparison, and isolate structural effects. Model skill was evaluated using KGE skill score (KGEss), NSE, R2, RMSE, and model parsimony using Bayesian Information Criterion (BIC). Across the full network, PM achieved substantially higher skill (KGEss: 0.86/0.85, NSE: 0.85/0.82, R2: 0.89/0.88) compared to PT-JPL (KGEss: 0.80/0.73, NSE: 0.47/0.32, R2: 0.57/0.56) with errors roughly halved (RMSE ≈ 14.7/14.8 W m⁻² for PM versus ≈ 27.2/27.8 W m⁻² for PT-JPL) in both calibration (2018–2021) and validation (2022–2023), respectively. Despite its higher parameterization, PM was strongly favored by parsimony diagnostics (ΔBIC of ~10 for PM versus ~522–1382 for PT-JPL). An information-theoretic assessment reinforced this separation: PM’s predictive mutual information centered near ~2 bits, compared to <1 bit for PT-JPL. Aridity-gradient analyses revealed that PM maintained stable performance across energy-limited and water-limited conditions, whereas PT-JPL skill degraded as aridity increased. Deep neural-network post-processing improved both models, but interpretability results suggest the corrected products still reflect their physical baselines. Overall, explicit resistance representation yields a more robust ET baseline in water-limited environments, while PT-JPL remains aridity-sensitive despite calibration and post-processing.