Modeling the Net Primary Productivity: A Study Case in the Brazilian Territory

The net primary productivity is one of the main indicators of an ecosystem’s health. The objectives of the present study were to assess the performance of machine learning techniques in net primary productivity modeling and to assess regional trends for the Brazilian territory. Net primary production was modeled using evapotranspiration estimates, the normalized difference vegetation index, hypsometry and meteorological data. The models adopted for estimating net primary productivity were stepwise regression, Bayesian regularized neural network and Cubist regression. A linear trend model was applied pixel by pixel in order to verify a significant change in net primary productivity across the Brazilian territory. The Cubist model performed best among the evaluated models, with root-mean-squared error of 135.6 g C m−2 year−1 and R2 equal to 0.78. While assessing the net primary productivity time series, an increased trend was observed for the Brazilian Savannah biome, which may be attributed to the replacement of some Savannah formations and degraded grasslands to agriculture. The developed model has shown a great potential for filling the gap of spatial net primary productivity data in large scales.