Estimating forest carbon fluxes using four different data-driven techniques based on long-term eddy covariance measurements: Model comparison and evaluation.

With the recent availability of large amounts of data from the global flux towers across different terrestrial ecosystems based on the eddy covariance technique, the use of data-driven techniques has been viable. In this study, two advanced techniques, namely adaptive neuro-fuzzy inference system (ANFIS) and extreme learning machine (ELM), were developed and investigated for their viability in estimating daily carbon fluxes at the ecosystem level. All the data used in this study were based upon the long-term chronosequence observations derived from the flux towers in eight forest ecosystems. Both ANFIS and ELM methods were further compared with the most widely used artificial neural network (ANN) and support vector machine (SVM) methods. Moreover, we also focused on probing into the effects of internal parameters on their corresponding approaches. Our estimates showed that most variation in each carbon flux could be effectively explained by the developed models at almost all the sites. Moreover, the forecasting accuracy of each method was strongly dependent upon their respective internal algorithms. The best training function for ANN model can be acquired through the trial and error procedure. The SVM model with the radial basis kernel function performed considerably better than the SVM models with the polynomial and sigmoid kernel functions. The hybrid ELM models achieved similar predictive accuracy for the three fluxes and were consistently superior to the original ELM models with different transfer functions. In most instances, both the subtractive clustering and fuzzy c-means algorithms for the ANFIS models outperformed the most popular grid partitioning algorithm. It was demonstrated that the newly proposed ELM and ANFIS models were able to produce comparable estimates to the ANN and SVM models for forecasting terrestrial carbon fluxes.