Comparison of Three Modeling Approaches for Predicting Net Ecosystem Productivity in Slash Pine Plantation in North-Central FL

Carbon fluxes are of great interest because of growing concerns about increasing CO2 concentrations in the air and its impacts on economics and policy. Forests have been shown to be a potentially valuable carbon sink, especially in the southeastern United States, which are predominantly covered by pine plantations. Three years of eddy and micro-meteorological data (2000-2003) from a mid-rotation slash pine plantation in Alachua county, Florida, USA were used as inputs and output verification with three different models: an artificial neural network (ANN), a classification/regression tree (CART), and a physiological model, PnET-Day. The ANN and CART models used the same climate input variables used for PnET-Day: PAR, maximum and minimum temperatures, and precipitation. The ANN yielded an r2 of 0.73, CART yielded an r2 of 0.74, and PnET-Day yielded an r2 of 0.45. Qualitatively, the ANN and CART were able to capture the temporal fluctuations, with the ANN better able to capture extreme values in CO2 fluxes than the CART. No gap filling was applied to any of the missing data. Omitting the dates with missing data, the annual sums for carbon uptake from the tower were 2.55, 1.71, 4.32, and 5.54 Mg C ha-1 yr-1, from 2000 to 2003 respectively. The annual carbon uptake estimates for CART were 1.87, 2.70, 4.57, and 4.98 Mg C ha-1 yr-1 from 2000 to 2003 respectively, 1.76, 2.75, 4.62, and 5.00 Mg C ha-1 yr-1 for the ANN, and 1.46, 4.06, 7.05, and 7.42 Mg C ha-1 yr-1 for PnET-Day. PnET-Day had a strong tendency to overestimate carbon uptake on both a daily basis and annual basis except for 2000, where all three models underestimated total carbon uptake. Overall, the responses from the CART and ANN models were close to each other and were closer to the measured values than PnET-Day. Further refinement and optimization of the models will be conducted on all three models, as well as testing on different time-frames (2004 onwards) and at other field sites to help assess how single intensive site “point” calibration may influence the results of a regional assessment. This study explores a set of promising modeling techniques for predicting carbon fluxes both over large areas and over changing climate conditions.