Optimization of JULES model in different sites of the Amazon biome

The Amazon biome is one of the largest carbon reservoirs, a relevant carbon sink in the world. The large extension and diversity of the Amazon biome hampers the assessment of regional-scale carbon budget based solely on local observations. Land surface models can provide carbon flux estimates, but they require proper calibration to represent the dynamics of the different ecosystems, abiotic conditions and vegetation characteristics in the Amazon Basin. One of the most important land surface model is JULES being increasingly used in tropical forests to estimate carbon fluxes. However, there is a lack of parameterization information that can be applied to the Amazon biome. Thus, this study presents an optimization of JULES main sensitivities parameters for different sites of the Amazon biome. For this attempt, we selected four Eddy-covariance flux towers as a reference based on different regions of the Amazon biome: K34 (Manaus, 2.614S/60.12W); K67 (Santarem, 2.85S/54.97W); RJA (Reserva Jaru, 10.08S/61.93W and ATTO (São Sebastião do Uatumã, 2.15S/59.03W). The variables analyzed to reproduce the carbon dynamics were the Net Ecosystem Exchange (NEE), Gross Primary Production (GPP) and eutrophic respiration (RESP) during one year of analysis. JULES most sensitivities parameters adjusted were related to the Upper-temperature threshold for photosynthesis (tupp_io); Scale factor for dark respiration (fd_io); The maximum ratio of internal to external CO₂ (f0_io) and Quantum efficiency (alpha_io).  The optimization was made using the Nelder-Mead method and after a leave-one-out cross-validation method was implemented to evaluate the simulation efficiency in each site. Also, the new parametrization in each site was compared with the default version of JULES and with another model Vegetation Photosynthesis and Respiration Model (VPRM). We selected the Wilmott index of agreement (d) and the Root Mean Square Error (RMSE) to analyze simulation efficiency. The Nelder-Mead optimization method reduced the error in GPP simulations in each Tower in comparison to the two models evaluated however the new parametrization of JULES was not able to improve RESP in these sites. However, the optimization procedure presented better results in NEE in each tower evaluated in the Amazon biome being the ATTO tower that demonstrated the most efficient simulations (d =0.60;  RMSE = 2.03 g C m^-2 day^-1) in comparison to the default version (d= 0.52; 3.09 g C m^-2 day^-1) and VPRM (d = 0.58; 2.29 g C m^-2 day^-1). In general, results demonstrated that the new parametrization of JULES reduced the error of simulation compared to the last version of JULES for tropical forests and better represented the seasonality compared to the VPRM model.