Comparison between neural network and M5 model tree for reconstructing missing evaporation data of Khuzestanation data of khouzestan

Pan Evaporation data used to estimate crop water requirements, but in some cases due to lack of accurate data on measurements or defective equipments failure and lost data as referred in missing data. Since data integration is important for irrigation planning, it is necessary to correct the statistical errors. Many methods were used for finding missing data, In the meantime, neural network and tree models have high degree of accuracy, however these models were not compared and evaluated. The aim of this study is to compare the model tree and neural network for reconstructing missing daily evaporation data for four stations in Khuzestan province.
In this study, the required data from four stations include Aghajari, Bandar Mahshahr, Izeh and Bostan located in Khuzestan province were collected. On the basis of the Koppen climate classification, the climate of these stations is arid. The data related to the years 1997 to 2008 that include daily values of pan evaporation, wind speed, maximum and minimum air temperature, relative humidity, sunshine and the extraterrestrial radiation. The data were divided to two four-year period (2005 to 2008) and 12 years (1997 to 2008) and at any period after intentional removal of 5%, 10% and 20% of the measured data, their values were estimated with the use of tree and neural network models. The results of the model were compared using Statistical indices.
In tree model coefficient of determination for four years period were: 85%, 75% and 85% and for 12 years period were: 90%, 83% and 84% respectively. In neural network model coefficient of determination for 4years period were: 85%, 75% and 85% and for 12years period were: 90%, 82% and 85% respectively. A higher coefficient value for 12 years period showed that models are more accurate to estimate missing data for longer term statistical data. By increasing missing data from 5% to 20%, accuracy of models was diminished. This research also indicated that both models have similar accuracy in the estimation of missing data.
According to the results of this study, when more than 10 years of data are available, both neural network and tree models will have relatively good results. Also, when the number of missing data are less or missed in shorter periods, estimated values will be closer to the actual values. In order to improve and complete results, it is suggested that statistical estimates of missing data for different periods, such as 8 or 15 years are repeated and the best period to determine for best practice models.