Nonlinear Interactions Between Solar‐Induced Fluorescence, Atmospheric Demand, and Soil Moisture Constrain Ecosystem Evapotranspiration Estimation

Abstract

Evapotranspiration (ET) links the terrestrial water and carbon cycles, yet its accurate estimation remains limited by simplified representations of plant‐atmosphere interactions. Solar‐induced chlorophyll fluorescence (SIF) has emerged as a promising physiological constraint on ET through its link to stomatal conductance, but current SIF‐based semi‐mechanistic models assume fixed, linear relationships that limit their applicability under variable environmental conditions. Here, we evaluate a widely used SIF‐driven semi‐mechanistic ET model against eddy covariance observations from 31 flux tower sites spanning eight plant functional types, and employ an interpretable machine learning (ML) framework with SHAP (SHapley Additive exPlanations) analysis to identify where these formulations become limiting. The performance gap widens markedly under environmental stress, with the semi‐mechanistic model showing reduced skill under low soil moisture ( = 0.53 vs. 0.70 for ML) and high vapor pressure deficit (VPD) ( = 0.45 vs. 0.65), indicating that fixed conductance formulations fail where nonlinear physiological regulation intensifies. SHAP diagnostics reveal that SIF and VPD are the dominant predictors of ET but their influence is conditional, with SIF‐ET coupling modulated by soil water content and VPD exhibiting threshold‐dependent behavior. PM‐ML analysis confirms that this limitation lies in the linear conductance formulation based on SIF, and global‐scale validation confirms these patterns across vegetated land. These results demonstrate that nonlinear coupling between photosynthetic activity, atmospheric demand, and soil water availability represents a first‐order control on ecosystem ET that current semi‐mechanistic formulations cannot adequately capture, and that interaction‐aware parameterizations of stomatal conductance are needed to reduce biases across ecosystems.