Interpretable model responsibility for cropland evapotranspiration across environmental gradients

Individual cropland evapotranspiration (ET) models encode different assumptions about energy partitioning, canopy regulation, atmospheric demand, and water limitation, making reliability strongly state dependent. Fusion can exploit complementary model strengths under environmental states relevant to irrigation decisions and crop-stress diagnosis, yet conventional fused estimates usually collapse these differences into a single flux, obscuring conditional reliability. We formulate ET fusion as interpretable responsibility learning. Relative model errors and a weak energy-consistency anchor define soft model responsibilities; an Extreme Gradient Boosting (XGBoost) multi-head regression gate then learns their dependence on environmental predictors and returns simplex-constrained weights for six physically based ET models. A strongly shrunk ElasticNet residual branch captures remaining systematic mismatch. Using 41 eddy-covariance cropland sites and 82,942 daily records, the framework achieved R² = 0.718, root-mean-square error (RMSE) = 22.13 W m⁻², and Kling-Gupta efficiency (KGE) = 0.843 under leave-one-site-out validation. The responsibility-weighted physical branch retained nearly the same skill (R² = 0.709, RMSE = 22.49 W m⁻², and KGE = 0.847), indicating that most predictive capacity came from the dynamic physical mixture. Responsibility landscapes showed overlapping model niches and diffuse competition: 88.1% of retained days jointly had high entropy and low top-two margin. Residual-ratio diagnostics localized shared inadequacy to low-demand, low-energy, and sparse-canopy margins. Matched random-split experiments inflated apparent skill and changed model ranking, underscoring the need to distinguish within-site-pool interpolation from transfer to unseen cropland sites. Responsibility learning therefore links cropland ET fusion to cross-site diagnostics of conditional model credibility, model ambiguity, and residual failure regimes relevant to agricultural water management.