Estimation of actual evapotranspiration: A novel hybrid method based on remote sensing and artificial intelligence

Actual evapotranspiration (AET) is one of the decisive factors controlling the water balance at the catchment level, particularly in arid and semi-arid regions, but measured data for which are generally unavailable. In this study, performance of a base artificial intelligence (AI) model, adaptive neuro-fuzzy inference system (ANFIS), and its hybrids with two bio-inspired optimization algorithms, namely shuffled frog leaping algorithm (SFLA) and grey wolf optimization (GWO), in estimating monthly AET was evaluated over 2001–2010 across Neishaboor watershed in Iran. The inputs of these models were categorized into three groups including meteorological, remotely sensed, and hybrid-based predictors, and defined in the form of 8 different scenarios. Net radiation (Rn), land surface temperature (LST), normalized difference vegetation index (NDVI), soil adjusted vegetation index (SAVI), and soil wetness deficit index (SWDI) were the remotely sensed predictors, computed using MODIS satellite images on the monthly scale for the study area. The results showed that the SWDI predictor has played a significant role in improving the accuracy of AET estimation, with the highest error reduction (12.5, 17 and 26.5% for ANFIS, ANFIS-SFLA, and ANFIS-GWO, respectively) obtained under scenarios including SWDI compared to corresponding scenarios excluding this predictor. In testing set, the three aforementioned models exhibited their best performance under Scenario 8 (RMSE = 11.93, NSE = 0.69, RRMSE = 0.37), Scenario 4 (RMSE = 11.06, NSE = 0.74, RRMSE = 0.37) and Scenario 4 (RMSE = 10.9, NSE = 0.76, RRMSE = 0.36), respectively. Coupling the SFLA and GWO optimization algorithms to the base model improved the accuracy of AET estimation, with the maximum error reduction for the two algorithms being about 12% (Scenarios 2 and 4) and 14% (Scenario 4), respectively. Examining the performance of the best scenarios of the three models in three intervals including the first, middle, and last third of measured AET values showed that all models were the most accurate in the first third interval. The results also indicated that all models have had higher accuracies in the first and middle third intervals of under-estimation set and the last interval of over-estimation set.