Multi-variable evaluation of agroecosystem models reveals challenges in daily and annual NEE simulation for croplands

Abstract

Net ecosystem exchange (NEE) of CO₂ is the residual of two opposing fluxes, and accuracy at the modelled NEE level can eclipse compensating errors in its components. This study evaluated DNDCv.CAN and EPIC against multi-variable observations at two Manitoba agricultural sites across 17 site-years. The two sites were Carberry (CAB, synthetic fertilizers, irrigated) and South Tobacco Creek (STC, manure-amended, rainfed). NEE was used as a diagnostic indicator of model structure. Component decomposition into gross primary production (GPP) and ecosystem respiration (R eco ), compensating-error analysis, and conditional residual binning were applied. Both models reproduced soil temperature, evapotranspiration, and yield with low error. Daily growing-season NEE from DNDCv.CAN had R² = 0.68 at CAB and R² = 0.33 at STC. EPIC returned R² = 0.49 at CAB and 0.29 at STC. DNDCv.CAN systematically under-predicted both GPP and R eco by similar magnitudes and the same sign at both sites. Roughly 10 to 28 % of growing-season days showed component error cancellation in NEE. R eco showed negative Nash-Sutcliffe efficiency at the daily scale and systematic under-prediction at the annual scale. STC was a net carbon source in 2014, 2015, and 2016, but both models predicted a net sink in all three years. Soybean at STC showed the weakest agreement with observations across all twelve site-crop component evaluations. Structural improvements are needed in respiration partitioning in DNDCv.CAN, in soybean representation under manure-amended soils, and in interannual soil water carryover in EPIC. Daily-scale multi-variable validation with component decomposition provides a diagnostic framework for identifying structural limitations that net-flux metrics obscure.