Enhanced estimation of reference evapotranspiration using hybrid deep learning models and remote sensing variables

Abstract

Effective water resources management and irrigation scheduling for agricultural sector highly depend on the precise estimation of reference evapotranspiration, ET_o. This study aims to develop ET_o estimation models using deep learning algorithms with remote sensing variables as the input variables at Pulau Langkawi and Kuantan stations, located in Peninsular Malaysia. Support vector regressor (SVR) was found to satisfactorily estimate the daytime land surface temperature (LST) using a set of significant variables including meteorological and remote sensing variables. It was then used along with downward shortwave radiation and surface reflectance bands to estimate ET_o. Both long short-term memory (LSTM) and gated recurrent unit (GRU) showed their equivalent capability in estimating ET_o and achieved the highest R² of 0.695 and 0.796, respectively. The proposed hybrid deep learning models, combined model of convolutional neural network (CNN) with LSTM and GRU, respectively, achieved higher accuracy compared to individual models. They managed to improve the accuracy of the prediction in most of the cases, with the highest R² = 0.805 and the lowest prediction errors, MAE = 0.265 mm/day, RMSE = 0.343 mm/day and NRMSE = 0.096. It was shown that the incorporation of surface reflectance bands and auxiliary variables (day length, Julian day and solar zenith angle) enhanced the performance of the models. This study provides valuable insights into deep learning algorithms and further confirms the potential of remote sensing variables as an alternative data source for ET_o estimation.