Sub-daily reconstruction of eddy-covariance energy fluxes under realistic contiguous outages: model benchmarking and energy-balance diagnostics

Eddy-covariance (EC) measurements of surface energy fluxes are frequently affected by data gaps that occur as contiguous multi-hour outages rather than randomly distributed missing points. However, most gap-filling evaluations still rely on random missingness assumptions, potentially overstating reconstruction skill under operational EC conditions. Here, we benchmark machine- and deep-learning approaches for realistic gap filling of latent heat (LE), sensible heat (H), and ground heat flux (G) using half-hourly observations from six European FLUXNET sites spanning arid to semi-arid, semi-humid, and humid ecosystems. Artificial gaps were imposed as contiguous 6–24 h outages at missing fractions of 5–50%, and performance was evaluated exclusively at masked timestamps. In addition to statistical accuracy, physical realism was assessed independently using surface energy-balance diagnostics based on observed net radiation. Reconstruction skill was controlled more strongly by gap structure than by missing fraction alone, highlighting the dominant influence of contiguous outages. Sensible heat was reconstructed most robustly (RMSE ≈ 41 W m⁻²; R² ≈ 0.93), followed by latent heat (RMSE ≈ 25 W m⁻²), while ground heat flux exhibited greater sensitivity to gap severity. Despite comparable median statistical performance, models differed substantially in physical consistency, with tree-based approaches showing lower and more stable energy-balance residuals (20–30 W m⁻² across sites) than deep-learning models, particularly under severe data loss. These results demonstrate that realistic gap structures and energy-balance diagnostics are essential for evaluating surface energy-flux gap filling under operational EC conditions, especially in water-limited ecosystems where long outages can distort sub-daily energy partitioning.