Prediction of diffuse photosynthetically active radiation using different soft computing techniques

Knowledge of diffuse photosynthetically active radiation (PARd) is important for many applications dealing with climate change, environmental engineering, and terrestrial productivity. It is necessary to estimate the PARd using different techniques due to the absence of direct observations of this radiometric flux in most parts of the world. In this study, Adaptive Neuro-Fuzzy Inference Systems (ANFIS) with grid partition (ANFIS-GP), ANFIS with subtractive clustering (ANFIS-SC)) and M5 model tree (M5Tree) are optimized and applied for predicting hourly PARd in different ecosystems. Hourly climatic variables at six stations from the AmeriFlux network are used for training, validating and testing above models. It is observed that the mean bias errors (MBE) values range −11 ~ 18%, −16 ~ 9% and −10 ~ 9% for ANFIS-GP, ANFIS-SC and M5Tree models, respectively; root mean square errors (RMSE) range 25-44%, 22-41% and 29-51%, respectively for different stations in testing period. The results show that the ANFIS-SC model can generally bring more accurate estimations than other models, and the statistical indices (MBE and RMSE) at CA_Gro stations (mixed forests) are lower than other stations for each model. The models underestimate some high PARd values (>900 umolm−2s−1) for some stations, which may due to the differences of data ranges and distributions. This study will lay the foundation for accurately mapping regional and global distributions of PARd and its associated ecological applications.