Development of random forest model as decision support tool in water resources management of Ogun headwater catchments

Abstract

AbstractWater resources cannot be effectively managed unless potential evapotranspiration is determined with high accuracy at headwater catchments. The study presents the most suitable feature combinations for building a reliable potential evapotranspiration (PET) model in the headwater catchments of Ogun River Basin, Southwest Nigeria. Using rainfall (R), wind speed (U2), sunshine hour (S), relative humidity (Rh), minimum temperature (Tmin) and maximum temperature (Tmax) as input features, a Random Forest (RF) model was developed to predict PET. Although the model yielded satisfactory results, it was subjected to the minimal depth and percentage increase in mean square error (%IncMSE). This was done to reduce the input features and to increase model accuracy. Thereafter various combinations of important input features were examined in order to establish the best combinations required to yield optimum results. The study revealed that although Tmax (%IncMSE of 652.09, p value < 0.05) and Rh (%IncMSE of 254.36, p value < 0.05) were the most important predictors of PET, a more reliable RF model was achieved when S and U2 were combined with them. Consequently, this study presents RF with a combination of four parameters (Tmax, Rh, S and U2) as an excellent computational technique for the prediction of PET in headwater catchments.