Machine learning estimates of eddy covariance carbon flux in a scrub in the Mexican highland

Arid and semi-arid ecosystems contain relatively high species diversity and are subject to intense use, in particular extensive cattle grazing, which has favoured the expansion and encroachment of perennial thorny shrubs into the grasslands, thus decreasing the value of the rangeland. However, these environments have been shown to positively impact global carbon dynamics. Machine learning and remote sensing had enhanced our knowledge about carbon dynamics, but they need to be further developed and adapted to particular analysis. We measured the net ecosystem exchange of C (NEE) with the Eddy Covariance (EC) method and estimated GPP in a thorny scrub at Bernal in Mexico. We tested the agreement between EC estimates and remotely sensed GPP estimates from MODIS, and also with two alternative modelling methods: ordinary least squares multiple regression (OLS) or ensembles of machine learning algorithms (EML). The variables used as predictors were Moderate Resolution Spectroradiometer (MODIS) spectral bands, vegetation indices and products, as well as gridded environmental variables. The Bernal site was a carbon sink despite it was overgrazed, the average NEE during fifteen months of 2017 and 2018 was -0.78 g C m^-2 d^-1 and the flux was negative or neutral during the measured months. The probability of agreement (θs) represented the agreement between observed and estimated values of GPP across the range of measurement. According to the mean value of θs, agreement was higher for the EML (0.6) followed by OLS (0.5) and then MODIS (0.24). This graphic metric was more informative than r² (0.98, 0.67, 0.58 respectively) to evaluate the model performance. This was particularly true for MODIS because the maximum θs of 4.3 was for measurements of 0.8 g C m^-2 d^-1 and then decreased steadily below 1 θs for measurements above 6.5 g C m^-2 d^-1 for this scrub vegetation. In the case of EML and OLS the θs was stable across the range of measurement. We used an EML for the Ameriflux site US-SRM, which is similar in vegetation and climate, to predict GPP at Bernal, but θs was low (0.16) indicating the local specificity of this model. Although cacti were an important component of the vegetation, the night time flux was characterized by positive NEE, suggesting that the photosynthetic dark-cycle flux of cacti was lower than ecosystem respiration. The discrepancy between MODIS and EC GPP estimates stresses the need to understand the limitations of both methods.